[
  {
    "paper_id": 9745,
    "title": "EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models",
    "abstract": "Diffusion models have demonstrated remarkable capabilities in image synthesis and related generative tasks.\nNevertheless, their practicality for low-latency real-world applications is constrained by substantial computational costs and latency issues.\nQuantization is a dominant way to compress and accelerate diffusion models, where post-training quantization (PTQ) and quantization-aware training (QAT) are two main approaches, each bearing its own properties.\nWhile PTQ exhibits efficiency in terms of both time and data usage, it may lead to diminished performance in low bit-width settings. On the other hand, QAT can help alleviate performance degradation but comes with substantial demands on computational and data resources.\nTo\ncapitalize on\nthe advantages while avoiding their respective drawbacks,\nwe introduce a data-free, quantization-aware and parameter-efficient fine-tuning framework for low-bit diffusion models, dubbed EfficientDM, to achieve QAT-level performance with PTQ-like efficiency.\nSpecifically, we propose a quantization-aware variant of the low-rank adapter (QALoRA) that can be merged with model weights and jointly quantized to low bit-width.\nThe fine-tuning process distills the denoising capabilities of the full-precision model into its quantized counterpart, eliminating the requirement for training data.\nTo further enhance performance, we introduce scale-aware LoRA optimization to address ineffective learning of QALoRA due to variations in weight quantization scales across different layers.\nWe also employ temporal learned step-size quantization to handle notable variations in activation distributions across denoising steps.\nExtensive experimental results demonstrate that our method significantly outperforms previous\nPTQ-based\ndiffusion models while maintaining similar\ntime and data efficiency.\nSpecifically, there is only a marginal0.050.050.050.05sFID increase when quantizing both weights and activations of LDM-4 to 4-bit on ImageNet256×256256256256\\times 256256 × 256.\nCompared to QAT-based methods, our EfficientDM also boasts a16.2×16.2\\times16.2 ×faster\nquantization\nspeed with comparable generation quality, rendering it a compelling choice for practical applications.\nCode is available athttps://github.com/ThisisBillhe/EfficientDM.",
    "introduction": "Diffusion models (DM)(Ho et al.,2022b; Dhariwal & Nichol,2021; Rombach et al.,2022a; Ho et al.,2022a)have demonstrated remarkable capabilities in image generation and related tasks.\nNonetheless, the iterative denoising process and the substantial computational overhead of the denoising model limit the efficiency of DM-based image generation.\nTo expedite the image generation process, numerous methods(Bao et al.,2022; Song et al.,2021; Liu et al.,2022; Lu et al.,2022)have been explored to reduce the number of denoising iterations, effectively reducing the previously required thousands of iterations to mere dozens.\nHowever, the significant volume of parameters within the denoising model still demands a substantial computational burden for each denoising step, resulting in considerable latency,\nhindering\nthe practical application of DM in real-world settings with latency and computational resource constraints.\n\n\nModel quantization, which compresses weights and activations from 32-bit floating-point values into lower-bit fixed-point formats, alleviating both memory and computational burdens.\nThis effect can be increasingly pronounced as the bit-width decreases.\nFor instance, leveraging Nvidia’s CUTLASS(Kerr et al.,2017)implementation, an 8-bit model’s inference speed can be2.03×2.03\\times2.03 ×faster than that of a full-precision (FP) model, and the acceleration ratio reaches3.34×3.34\\times3.34 ×for a 4-bit model.\nTherefore, it possesses substantial potential for significantly compressing and accelerating diffusion models, making them highly suitable for deployment on resource-constrained devices such as mobile phones.\n\n\nNevertheless, the challenges associated with low-bit quantization for diffusion models have not received adequate attention.\nTypically, model quantization can be executed through two predominant approaches: post-training quantization (PTQ) and training-aware quantization (QAT).\nPTQ calibrates the quantization parameters with a small calibration dataset, which is time- and data-efficient.\nHowever, they introduce substantial quantization errors at low bit-width.\nAs illustrated in Figure1, when quantizing both weights and activations to 4-bit with the PTQ method(He et al.,2023a), diffusion models fail to maintain their denoising capabilities.\nIn contrast, QAT methods(Krishnamoorthi,2018; Esser et al.,2019)can recover performance losses at lower bit-width by fine-tuning the whole model.\nHowever, this approach requires significantly more time and computing resources compared to PTQ method(He et al.,2023a), as evidenced by a2.6×2.6\\times2.6 ×increase in GPU memory consumption (31.4GB vs. 11.7GB) and a18.9×18.9\\times18.9 ×longer execution time (54.5 GPU hours vs. 2.88 GPU hours) when fine-tuning LDM-4(Rombach et al.,2022a)on ImageNet256×256256256256\\times 256256 × 256.\nMoreover, in some cases, it may be challenging or even impossible to obtain the original training dataset due to privacy or copyright concerns.\n\n\nFigure 1:An overview of the efficiency-vs-quality tradeoff across various quantization approachs. Data is collected on LDM-8(Rombach et al.,2022a)with 4-bit weights and activations on LSUN-Churches. The GPU memory consumption is visualized by circle size.\n\n\nIn this paper, we introduce a data-free and parameter-efficient fine-tuning framework for low-bit diffusion models, denoted as EfficientDM, which demonstrates the capability to achieveQAT-level performancewhile upholdingPTQ-level efficiencyduring fine-tuning in terms of data and time.\nThe foundation of our approach lies in a quantization-aware form of the low-rank adapter, as depicted in Figure2b.\nThis variant enables the joint quantization of LoRA weights with model weights, thereby obviating additional storage and calculations.\nCompared to previous QAT method(So et al.,2023), our fine-tuning process is executed in a data-free manner, accomplished by minimizing the mean squared error (MSE) between the estimated noise of full-precision denoising model and its quantized counterpart, as illustrated in Figure2c, thus eliminating the need for the original training dataset.\nDue to quantization, the relationship between full-precision LoRA weights and quantized updated model weights becomes step-like with step size,i.e., the quantization scale parameter.\nEffective updates are mostly observed in layers with small scales, while other layers with larger scales do not benefit. To address this, we introduce scale-aware LoRA optimization that adaptively adjusts the gradient scales of LoRA weights in different layers to ensure an effective optimization.\nFurthermore, we extend the learned step size quantization (LSQ) method(Esser et al.,2019)into the denoising temporal domain for activations, effectively mitigating quantization errors due to varying activation distributions across time steps.\n\n\nFigure 2:An overview of the proposed EfficientDM fine-tuning framework. Here,𝐬wsubscript𝐬𝑤\\mathbf{s}_{w}bold_s start_POSTSUBSCRIPT italic_w end_POSTSUBSCRIPTandsxsubscript𝑠𝑥s_{x}italic_s start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPTrepresent the learnable quantization scales for weights and activations, respectively. Compared to QLoRA layer, both updated weights and activations in our QALoRA are quantized to enable efficient bitwise operations during inference.\nFine-tuning is performed by minimizing the mean squared error between the estimated noises of FP and quantized models.\n\n\nIn summary, our contributions are as follows:•We introduce EfficientDM, an efficient fine-tuning framework for low-bit diffusion models which can achieve QAT performance with the efficiency of PTQ. The framework is rooted in the quantization-aware form of low-rank adapters (QALoRA) and distills the denoising capabilities of full-precision models into their quantized counterparts.•We propose scale-aware LoRA optimization to alleviate the ineffective learning of QALoRA resulting from substantial variations in weight quantization scales across different layers. We also introduce TALSQ, an extension of activation learned step size quantization within the temporal domain to tackle the variation in activation distributions across denoising steps.•Extensive experiments on CIFAR-10, LSUN and ImageNet demonstrate that our\nEfficientDM\nreaches a new state-of-the-art performance for low-bit quantization of diffusion models.\n\n",
    "related_work": "Model quantization.Quantization is a widely employed technique for compressing and accelerating neural networks. Depending on whether fine-tuning of the model is necessitated, this method can be categorized into two approaches: post-training quantization (PTQ)(Nagel et al.,2020; Li et al.,2021; Wei et al.,2022; Lin et al.,2022)and quantization-aware training (QAT)(Esser et al.,2019; Gong et al.,2019; Nagel et al.,2022; Louizos et al.,2019; Jacob et al.,2018; Zhuang et al.,2018; He et al.,2023b). PTQ does not involve fine-tuning the model’s weights and only requires a small dataset to calibrate the quantization parameters. This approach is fast and data-efficient\nbut may result in suboptimal performance especially when employing low bit-widths. Recent advancements in reconstruction-based PTQ methods(Li et al.,2021; Wei et al.,2022)utilize second-order error analysis and gradient descent algorithms to optimize quantization parameters. Such methods demonstrate robust performance, which can perform well even at 4-bit on image classification tasks. On the other hand, QAT entails fine-tuning the model weights to achieve low-bit quantization while minimizing performance degradation. Nevertheless, the substantial data prerequisites and high computational overhead render QAT much slower than PTQ. For instance, when applied to ResNet-18 model for the ImageNet classification task, it is reported to be240×240\\times240 ×slower than the PTQ method(Li et al.,2021).\nThis hindrance poses challenges to its widespread adoption, particularly for models with massive parameters.\n\n\nGiven the demand to compress diffusion models and accelerate their sampling, quantization is considered an effective approach. Currently, most quantization work on diffusion models has focused on PTQ(Shang et al.,2023; Li et al.,2023; He et al.,2023a). Since diffusion models can generate samples from random Gaussian noise, these methods are conducted in a data-free manner, with the calibration sets collected from the denoising process. For example, PTQ4DM(Shang et al.,2023)and Q-Diffusion(Li et al.,2023)apply reconstruction-based PTQ methods to diffusion models. PTQD(He et al.,2023a)further decomposes quantization noise and fuses it with diffusion noise. Nevertheless, these methods experience significant performance degradation at 4-bit or lower bit-widths.\nRecent work TDQ(So et al.,2023)fine-tunes diffusion models in a QAT manner and employs additional MLP modules to estimate quantization scales for each step.\nHowever, it requires the original dataset and training the entire quantized model with these extra MLP modules, incurring significant costs. In contrast, our approach introduces only a few trainable quantization scales per layer and can attain QAT-level performance with PTQ-level efficiency.\n\n\nParameter-efficient fine-tuning.Parameter-Efficient Fine-Tuning (PEFT) has\nemerged as a potent alternative to full fine-tuning, which tunes a small subset of parameters while keeping the vast majority frozen, to ease storage burden of large pretrained models.\nNoteworthy approaches within this realm include Prefix Tuning(Li & Liang,2021; Liu et al.,2021), Prompt Tuning(Lester et al.,2021; Wang et al.,2022), and Low-rank adapters (LoRA)(Hu et al.,2022; Zhang et al.,2022), which have found extensive utility within large language models (LLMs).\n\n\nAmong these PEFT methods, LoRA(Hu et al.,2022)keeps the pre-trained weights fixed and learns a set of low-rank decomposition matrices as updates to the weights. This method substantially reduces the number of trainable parameters, and\ndoes not impose additional computational overhead due to re-parameterization.\nQLoRA(Dettmers et al.,2023)combines this technique with quantized LLMs, further curtailing memory requirements while mitigating performance degradation brought by quantization. However, its full-precision LoRA weights cannot be integrated with quantized model weights, leading to extra storage and computational costs for deployment. In contrast, we propose to merge LoRA weights with full-precision model weights and jointly quantize them to the target bit-width, enabling their seamless integration without incurring any additional overhead for deployment.\nAdditionally, we identify and address the issue of ineffective learning of LoRA weights stemming from the variations in weight quantization scales, further enhancing the performance of low-bit diffusion models.\n\n",
    "origin_citations": [
      "Up or down? adaptive rounding for post-training quantization.",
      "BRECQ: pushing the limit of post-training quantization by block reconstruction.",
      "Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization.",
      "Fq-vit: Post-training quantization for fully quantized vision transformer.",
      "Learned step size quantization.",
      "Differentiable soft quantization: Bridging full-precision and low-bit neural networks.",
      "Overcoming oscillations in quantization-aware training.",
      "Relaxed quantization for discretized neural networks.",
      "Quantization and training of neural networks for efficient integer-arithmetic-only inference.",
      "Towards effective low-bitwidth convolutional neural networks.",
      "Bivit: Extremely compressed binary vision transformer.",
      "Post-training quantization on diffusion models.",
      "Q-diffusion: Quantizing diffusion models.",
      "Ptqd: Accurate post-training quantization for diffusion models.",
      "Temporal dynamic quantization for diffusion models.",
      "Prefix-tuning: Optimizing continuous prompts for generation.",
      "P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks.",
      "The power of scale for parameter-efficient prompt tuning.",
      "Multitask prompt tuning enables parameter-efficient transfer learning.",
      "Lora: Low-rank adaptation of large language models.",
      "Adaptive budget allocation for parameter-efficient fine-tuning.",
      "Qlora: Efficient finetuning of quantized llms."
    ],
    "citations": [
      191955,
      118706,
      517176,
      337681,
      264652,
      239014,
      501857,
      443793,
      9139,
      528096,
      32264,
      242559,
      490179,
      261490,
      307359,
      1591,
      1590,
      8536,
      78036,
      1589,
      1090069,
      2253
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 90,
      "Publication_Potential": 95,
      "Final_Score": 91.5,
      "raw_text": "Content_Quality: 90\nPublication_Potential: 95\nFinal_Score: 91"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 95,
      "Final_Score": 90.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 95\nFinal_Score: 90"
    },
    "overall_score": 90.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "UmMa3UNDAz",
        "forum": "UmMa3UNDAz",
        "number": 3007,
        "title": "EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models",
        "normalized_title": "efficientdm efficient quantization aware fine tuning of low bit diffusion models",
        "abstract": "Diffusion models have demonstrated remarkable capabilities in image synthesis and related generative tasks. Nevertheless, their practicality for low-latency real-world applications is constrained by substantial computational costs and latency issues. Quantization is a dominant way to compress and accelerate diffusion models, where post-training quantization (PTQ) and quantization-aware training (QAT) are two main approaches, each bearing its own properties. While PTQ exhibits efficiency in terms of both time and data usage, it may lead to diminished performance in low bit-width settings. On the other hand, QAT can help alleviate performance degradation but comes with substantial demands on computational and data resources. To capitalize on the advantages while avoiding their respective drawbacks, we introduce a data-free, quantization-aware and parameter-efficient fine-tuning framework for low-bit diffusion models, dubbed EfficientDM, to achieve QAT-level performance with PTQ-like efficiency. Specifically, we propose a quantization-aware variant of the low-rank adapter (QALoRA) that can be merged with model weights and jointly quantized to low bit-width. The fine-tuning process distills the denoising capabilities of the full-precision model into its quantized counterpart, eliminating the requirement for training data. To further enhance performance, we introduce scale-aware optimization to address ineffective learning of QALoRA due to variations in weight quantization scales across different layers. We also employ temporal learned step-size quantization to handle notable variations in activation distributions across denoising steps. Extensive experimental results demonstrate that our method significantly outperforms previous PTQ-based diffusion models while maintaining similar time and data efficiency. Specifically, there is only a marginal $0.05$ sFID increase when quantizing both weights and activations of LDM-4 to 4-bit on ImageNet $256\\times256$. Compared to QAT-based methods, our EfficientDM also boasts a $16.2\\times$ faster quantization speed with comparable generation quality, rendering it a compelling choice for practical applications.",
        "authors": [
          "Yefei He",
          "Jing Liu",
          "Weijia Wu",
          "Hong Zhou",
          "Bohan Zhuang"
        ],
        "decision": null,
        "pdate": 1705410862231,
        "cdate": 1695260879012
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "UmMa3UNDAz",
          "forum": "UmMa3UNDAz",
          "number": 3007,
          "title": "EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models",
          "normalized_title": "efficientdm efficient quantization aware fine tuning of low bit diffusion models",
          "abstract": "Diffusion models have demonstrated remarkable capabilities in image synthesis and related generative tasks. Nevertheless, their practicality for low-latency real-world applications is constrained by substantial computational costs and latency issues. Quantization is a dominant way to compress and accelerate diffusion models, where post-training quantization (PTQ) and quantization-aware training (QAT) are two main approaches, each bearing its own properties. While PTQ exhibits efficiency in terms of both time and data usage, it may lead to diminished performance in low bit-width settings. On the other hand, QAT can help alleviate performance degradation but comes with substantial demands on computational and data resources. To capitalize on the advantages while avoiding their respective drawbacks, we introduce a data-free, quantization-aware and parameter-efficient fine-tuning framework for low-bit diffusion models, dubbed EfficientDM, to achieve QAT-level performance with PTQ-like efficiency. Specifically, we propose a quantization-aware variant of the low-rank adapter (QALoRA) that can be merged with model weights and jointly quantized to low bit-width. The fine-tuning process distills the denoising capabilities of the full-precision model into its quantized counterpart, eliminating the requirement for training data. To further enhance performance, we introduce scale-aware optimization to address ineffective learning of QALoRA due to variations in weight quantization scales across different layers. We also employ temporal learned step-size quantization to handle notable variations in activation distributions across denoising steps. Extensive experimental results demonstrate that our method significantly outperforms previous PTQ-based diffusion models while maintaining similar time and data efficiency. Specifically, there is only a marginal $0.05$ sFID increase when quantizing both weights and activations of LDM-4 to 4-bit on ImageNet $256\\times256$. Compared to QAT-based methods, our EfficientDM also boasts a $16.2\\times$ faster quantization speed with comparable generation quality, rendering it a compelling choice for practical applications.",
          "authors": [
            "Yefei He",
            "Jing Liu",
            "Weijia Wu",
            "Hong Zhou",
            "Bohan Zhuang"
          ],
          "decision": null,
          "pdate": 1705410862231,
          "cdate": 1695260879012
        }
      ]
    }
  },
  {
    "paper_id": 10235,
    "title": "Pi-DUAL: Using privileged information to distinguish clean from noisy labels",
    "abstract": "Label noise is a pervasive problem in deep learning that often compromises the generalization performance of trained models. Recently, leveraging privileged information (PI) – information available only during training but not at test time – has emerged as an effective approach to mitigate this issue. Yet, existing PI-based methods have failed to consistently outperform their no-PI counterparts in terms of preventing overfitting to label noise. To address this deficiency, we introduce Pi-DUAL, an architecture designed to harness PI to distinguish clean from wrong labels. Pi-DUAL decomposes the output logits into a prediction term, based on conventional input features, and a noise-fitting term influenced solely by PI. A gating mechanism steered by PI adaptively shifts focus between these terms, allowing the model to implicitly separate the learning paths of clean and wrong labels. Empirically, Pi-DUAL achieves significant performance improvements on key PI benchmarks (e.g.,+6.8%percent6.8+6.8\\%+ 6.8 %on ImageNet-PI), establishing a new state-of-the-art test set accuracy. Additionally, Pi-DUAL is a potent method for identifying noisy samples post-training, outperforming other strong methods at this task. Overall, Pi-DUAL is a simple, scalable and practical approach for mitigating the effects of label noise in a variety of real-world scenarios with PI.",
    "introduction": "Many deep learning models are trained on large noisy datasets,\nas obtaining cleanly labeled datasets at scale\ncan be expensive and time consuming(Snow et al.,2008; Sheng et al.,2008). However, the presence of label noise in the training set tends to damage generalization performance as it forces the model to learn spurious associations between the input features and the noisy labels(Zhang et al.,2017; Arpit et al.,2017). To mitigate the negative effects of label noise,\nrecent methods have primarily tried to prevent overfitting to the noisy labels, often utilising the observation that neural networks tend to first learn the clean labels before memorizing the wrong ones(Maennel et al.,2020; Baldock et al.,2021).\nFor instance, these methods\ninclude filtering out incorrect labels, correcting them, or enforcing regularization on the training dynamics(Han et al.,2018; Liu et al.,2020; Li et al.,2020a). Other works, instead, try to\ncapture the noise structure in an input-dependent fashion(Patrini et al.,2017; Liu et al.,2022; Collier et al.,2022,2023).\n\n\nThe above methods are however designed for a standard supervised learning setting, where models are tasked to learn\nan association\nbetween input features𝒙∈ℝd𝒙superscriptℝ𝑑{\\bm{x}}\\in\\mathbb{R}^{d}bold_italic_x ∈ blackboard_R start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPTand targetsy∈{1,…,K}𝑦1…𝐾y\\in\\{1,\\dots,K\\}italic_y ∈ { 1 , … , italic_K }(assumingK𝐾Kitalic_Kclasses) from a training set of pairs{(𝒙i,y~i)}i∈[n]subscriptsubscript𝒙𝑖subscript~𝑦𝑖𝑖delimited-[]𝑛\\{({\\bm{x}}_{i},\\tilde{y}_{i})\\}_{i\\in[n]}{ ( bold_italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT , over~ start_ARG italic_y end_ARG start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) } start_POSTSUBSCRIPT italic_i ∈ [ italic_n ] end_POSTSUBSCRIPTof features and\n(possibly)\nnoisy labelsy~∈{1,…⁢K}~𝑦1…𝐾\\tilde{y}\\in\\{1,\\dots K\\}over~ start_ARG italic_y end_ARG ∈ { 1 , … italic_K }.\nAs a result, they need to model the noise in the targets as a function of𝒙𝒙{\\bm{x}}bold_italic_x.\nYet, in many practical situations, the mistakes introduced during the annotation process may not solely depend on the input𝒙𝒙{\\bm{x}}bold_italic_x, but rather be mostly explained by annotation-specific side information, such as the experience of the annotator or the attention they paid while annotating.\nFor this reason, a recent line of work(Vapnik & Vashist,2009; Collier et al.,2022; Ortiz-Jimenez et al.,2023)has proposed to useprivileged information(PI) to mitigate the affects of label noise.\nPI is defined as additional features available at training time but not at test time. It can include annotation features such as the annotator ID, the amount of time to provide the label, or their experience.\n\n\nRemarkably, having access to PI at training time, even when it is not available at test time, has been shown to be an effective tool for dealing with instance-dependent label noise. Most notably,Ortiz-Jimenez et al. (2023)showed that by exploiting PI it is possible to activate positive learning shortcuts to memorize, and therefore explain away, noisy training samples, thereby improving generalization. Nevertheless, and perhaps surprisingly, current PI-based methods do not systematically outperform no-PI baselines in the presence of label noise, making them a less competitive alternative in certain cases(Ortiz-Jimenez et al.,2023).\n\n\nFigure 1:Illustration of the architecture of Pi-DUAL.(Left) During training, Pi-DUAL fits the noisy target labely~~𝑦\\tilde{y}over~ start_ARG italic_y end_ARGcombining the output of a prediction network (which takes the regular features𝒙𝒙{\\bm{x}}bold_italic_xas input) and a noise network (which takes the PI𝒂𝒂{\\bm{a}}bold_italic_aas input). The outputs of these sub-networks are weighted based on the output of a gating network (which also has𝒂𝒂{\\bm{a}}bold_italic_aas input) and then passed through asoftmaxsoftmax\\operatorname{softmax}roman_softmaxoperator to obtain the predictions. (Right) During inference, when only𝒙𝒙{\\bm{x}}bold_italic_xis available, Pi-DUAL does not need access to PI and simply uses the prediction network to predict the clean targety𝑦yitalic_y.\n\n\nIn this work, we aim to improve the performance of PI strategies\nby proposing a new PI-guided noisy label architecture:Pi-DUAL, aPrivilegedInformation network toDistinguishUntrustworthyAnnotations andLabels. Specifically, during training, we propose to decompose the output logits into a weighted combination of a prediction term, that depends only on the regular features𝒙𝒙{\\bm{x}}bold_italic_x, and a noise-fitting term, that depends only on the PI features𝒂∈ℝp𝒂superscriptℝ𝑝{\\bm{a}}\\in\\mathbb{R}^{p}bold_italic_a ∈ blackboard_R start_POSTSUPERSCRIPT italic_p end_POSTSUPERSCRIPT. Pi-DUAL toggles between these terms using a gating mechanism, also solely a function of𝒂𝒂{\\bm{a}}bold_italic_a, that decides if a sample should be learned primarily by the prediction network, or explained away by the noise network (see Fig.1). This dual sub-network design adaptively routes the clean and wrong labels through the prediction and noise networks so that they are fit based on𝒙𝒙{\\bm{x}}bold_italic_xor𝒂𝒂{\\bm{a}}bold_italic_a, respectively.\nThis protects the prediction network from overfitting to the label noise. Pi-DUAL is simple to implement, effective, and can be trained end-to-end with minimal modifications to existing training pipelines. Unlike some previous methods Pi-DUAL also scales to training on very large datasets. Finally, in public benchmarks for learning with label noise, Pi-DUAL achieves state-of-the-art results on datasets with rich PI features (+4.5%percent4.5+4.5\\%+ 4.5 %on CIFAR-10H,+1.3%percent1.3+1.3\\%+ 1.3 %on ImageNet-PI (low-noise) and+6.8%percent6.8+6.8\\%+ 6.8 %on ImageNet-PI (high-noise)); and performs on par with previous methods on benchmarks with weak PI or no PI at all, despite not being specifically designed to work in these regimes.\n\n\nOverall, the main contributions of our work are:•We present Pi-DUAL, a novel PI method to combat label noise based on a dual path architecture that implicitly separates the noisy fitting path from the clean prediction path during training.•We show that Pi-DUAL achieves strong performance on noisy label benchmarks without sacrificing scalability, outperforming previous state-of-the-art methods when given access to high-quality PI features.•We provide a simple yet effective method to detect wrong labels in the training set using the prediction network of Pi-DUAL, achieving strong detection performance.\n\n\nIn summary, our work advances the state-of-the-art in noisy label learning by effectively leveraging privileged information through the novel Pi-DUAL architecture. Pi-DUAL can be easily integrated into any learning pipeline, requires minimal hyperparameters, and can be trained end-to-end in a single stage. Overall, Pi-DUAL is a scalable and practical approach for mitigating the effects of label noise in a variety of real-world scenarios with PI.\n\n",
    "related_work": "Table 1:Comparison of different representative methods to learn with label noisevsPi-DUAL on several design axes: ability to leverage PI, ability to explicitly model the noise signal, parameter scalability, and whether training requires multiple models and training stages. Scalability indicates whether the number of parameters for the method remains constant regardless of the number of samples or the number of classes in the training set.MethodsLeverage PIExplicitnoise modelingScalability w.r.t.num. of samplesScalability w.r.t.num. of classesTrainingcomplexityForward-T(Patrini et al.,2017)✗✓✓✗1 model, 2 stagesCo-Teaching(Han et al.,2018)✗✗✓✓2 models, 1 stageDivide-Mix(Li et al.,2020a)✗✗✓✓2 models, 1 stageELR(Liu et al.,2020)✗✗✗✗1 model, 1 stageSOP(Liu et al.,2022)✗✓✗✗1 model, 1 stageHET-XL(Collier et al.,2023)✗✓✓✓1 model, 1 stageDistill. PI(Lopez-Paz et al.,2015)✓✗✓✓2 model, 2 stageAFM(Collier et al.,2022)✓✗✓✓1 model, 1 stageTRAM++(Ortiz-Jimenez et al.,2023)✓✗✓✓1 model, 1 stagePi-DUAL (Ours)✓✓✓✓1 model, 1 stage\n\n\nNoisy label methods mostly fall into two broad categories:\nthose that explicitly model the noise signal, and those that rely on implicit network dynamics to correct or ignore the wrong labels(Song et al.,2022). Noise modeling techniques aim to learn the function that governs the noisy annotation process explicitly during training, inverting it during inference to obtain the clean labels. Some methods model the annotation function using a transition matrix(Patrini et al.,2017); others model uncertainty via a heteroscedastic noise term(Collier et al.,2021,2023); and recently, some works explicitly parameterize the label error signal as a vector for each sample in the training set(Tanaka et al.,2018; Yi & Wu,2019; Liu et al.,2022). Implicit-dynamics based approaches, on the other hand, operate under the assumption that wrong labels are harder to learn than the correct labels(Zhang et al.,2017; Maennel et al.,2020). Using this intuition, different methods have come up with different heuristics to correct(Jiang et al.,2018; Han et al.,2018; Yu et al.,2019)or downweight(Liu et al.,2020; Menon et al.,2020; Bai et al.,2021)the influence of wrong labels during training. This has sometimes led to very complex methods that require multiple stages of training(Patrini et al.,2017; Bai et al.,2021; Albert et al.,2023; Wang et al.,2023), higher computational cost(Han et al.,2018; Jiang et al.,2018; Han et al.,2018; Yu et al.,2019), and many additional parameters that do not scale well to large datasets(Yi & Wu,2019; Liu et al.,2020,2022).\n\n\nThe introduction of privileged information (PI) offers an alternative dimension to tackle the noisy label problem(Hernández-Lobato et al.,2014; Lopez-Paz et al.,2015; Collier et al.,2022). In this regard,Ortiz-Jimenez et al. (2023)showed that most PI methods work as implicit-dynamics approaches. They rely on the use of PI to enable learning shortcuts, to avoid memorizing the incorrect labels using the regular features.\nMoreover, these approaches are attractive for their scalability, as they usually avoid the introduction of extra training stages or parameters. However, current PI methods can sometimes lag behind in performance with respect to no-PI baselines. The main reason is that these methods still try to learn the noise predictive distributionp⁢(y~|𝒙)𝑝conditional~𝑦𝒙p(\\tilde{y}|{\\bm{x}})italic_p ( over~ start_ARG italic_y end_ARG | bold_italic_x )by marginalizing𝒂𝒂{\\bm{a}}bold_italic_ainp⁢(y~|𝒙,𝒂)𝑝conditional~𝑦𝒙𝒂p(\\tilde{y}|{\\bm{x}},{\\bm{a}})italic_p ( over~ start_ARG italic_y end_ARG | bold_italic_x , bold_italic_a ), when they should actually aim to learn the clean distributionp⁢(y|𝒙)𝑝conditional𝑦𝒙p(y|{\\bm{x}})italic_p ( italic_y | bold_italic_x )directly. However, prior PI methods do not have an explicit mechanism to identify or correct\nthe wrong labels.\n\n\nOur proposed method, Pi-DUAL, tries to circumvent these issues by explicitly modeling theclean distribution, exploiting the ability of PI to distinguish clean and wrong labels. Our design allows Pi-DUAL to scale effectively across large datasets and diverse class distributions, while maintaining high performance and low training complexity as seen in Tab.1. We further note that our design is reminiscent of mixtures of experts (MoE) that were shown to be a competitive architecture for language modeling(Shazeer et al.,2017)and computer vision(Riquelme et al.,2021). By analogy, we can see Pi-DUAL as an MoE containing a single MoE layer with two heterogeneous experts—the prediction and noise networks—located at the logits of the model and with a dense gating.\n\n",
    "origin_citations": [
      "Making deep neural networks robust to label noise: A loss correction approach.",
      "Co-teaching: Robust training of deep neural networks with extremely noisy labels.",
      "Dividemix: Learning with noisy labels as semi-supervised learning.",
      "Early-learning regularization prevents memorization of noisy labels.",
      "Robust training under label noise by over-parameterization.",
      "Massively scaling heteroscedastic classifiers.",
      "Unifying distillation and privileged information.",
      "Transfer and marginalize: Explaining away label noise with privileged information.",
      "When does privileged information explain away label noise?",
      "Learning from noisy labels with deep neural networks: A survey.",
      "Correlated input-dependent label noise in large-scale image classification.",
      "Joint optimization framework for learning with noisy labels.",
      "Probabilistic end-to-end noise correction for learning with noisy labels.",
      "Understanding deep learning requires rethinking generalization.",
      "What do neural networks learn when trained with random labels?",
      "Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels.",
      "How does disagreement help generalization against label corruption?",
      "Can gradient clipping mitigate label noise?",
      "Understanding and improving early stopping for learning with noisy labels.",
      "Is your noise correction noisy? pls: Robustness to label noise with two stage detection.",
      "ProMix: combating label noise via maximizing clean sample utility.",
      "Mind the nuisance: Gaussian process classification using privileged noise.",
      "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.",
      "Scaling vision with sparse mixture of experts."
    ],
    "citations": [
      389500,
      3340,
      4253,
      8735,
      907564,
      212199,
      455196,
      174908,
      78536,
      4454,
      397298,
      4184,
      4187,
      3630,
      169382,
      136899,
      4140,
      969370,
      459876,
      421190,
      114150,
      680,
      194116
    ],
    "extract_rate": 0.9583333333333334,
    "article_scores": {
      "Content_Quality": 88,
      "Publication_Potential": 85,
      "Final_Score": 87.1,
      "raw_text": "Content_Quality: 88\nPublication_Potential: 85\nFinal_Score: 87"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 90,
      "Final_Score": 90.5,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 95\nSynthesis_Analysis: 90\nFinal_Score: 91"
    },
    "overall_score": 89.13999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "9oAXix8da9",
        "forum": "9oAXix8da9",
        "number": 5772,
        "title": "Pi-DUAL: Using privileged information to distinguish clean from noisy labels",
        "normalized_title": "pi dual using privileged information to distinguish clean from noisy labels",
        "abstract": "Label noise is a pervasive problem in deep learning that often compromises the generalization performance of trained models. Recently, leveraging privileged information (PI) -- information available only during training but not at test time -- has emerged as an effective approach to mitigate this issue. Yet, existing PI-based methods have failed to consistently outperform their no-PI counterparts in terms of preventing overfitting to label noise. To address this deficiency, we introduce Pi-DUAL, an architecture designed to harness PI to distinguish clean from wrong labels. Pi-DUAL decomposes the output logits into a prediction term, based on conventional input features, and a noise-fitting term influenced solely by PI. A gating mechanism steered by PI adaptively shifts focus between these terms, allowing the model to implicitly separate the learning paths of clean and wrong labels. Empirically, Pi-DUAL achieves significant performance improvements on key PI benchmarks (e.g., +6.8% on ImageNet-PI), establishing a new state-of-the-art test set accuracy. Additionally, Pi-DUAL is a potent method for identifying noisy samples post-training, outperforming other strong methods at this task. Overall, Pi-DUAL is a simple, scalable and practical approach for mitigating the effects of label noise in a variety of real-world scenarios with PI.",
        "authors": [
          "Ke Wang",
          "Guillermo Ortiz-Jimenez",
          "Rodolphe Jenatton",
          "Mark Collier",
          "Efi Kokiopoulou",
          "Pascal Frossard"
        ],
        "decision": null,
        "pdate": 1714610398356,
        "cdate": 1706802783444
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "9oAXix8da9",
          "forum": "9oAXix8da9",
          "number": 5772,
          "title": "Pi-DUAL: Using privileged information to distinguish clean from noisy labels",
          "normalized_title": "pi dual using privileged information to distinguish clean from noisy labels",
          "abstract": "Label noise is a pervasive problem in deep learning that often compromises the generalization performance of trained models. Recently, leveraging privileged information (PI) -- information available only during training but not at test time -- has emerged as an effective approach to mitigate this issue. Yet, existing PI-based methods have failed to consistently outperform their no-PI counterparts in terms of preventing overfitting to label noise. To address this deficiency, we introduce Pi-DUAL, an architecture designed to harness PI to distinguish clean from wrong labels. Pi-DUAL decomposes the output logits into a prediction term, based on conventional input features, and a noise-fitting term influenced solely by PI. A gating mechanism steered by PI adaptively shifts focus between these terms, allowing the model to implicitly separate the learning paths of clean and wrong labels. Empirically, Pi-DUAL achieves significant performance improvements on key PI benchmarks (e.g., +6.8% on ImageNet-PI), establishing a new state-of-the-art test set accuracy. Additionally, Pi-DUAL is a potent method for identifying noisy samples post-training, outperforming other strong methods at this task. Overall, Pi-DUAL is a simple, scalable and practical approach for mitigating the effects of label noise in a variety of real-world scenarios with PI.",
          "authors": [
            "Ke Wang",
            "Guillermo Ortiz-Jimenez",
            "Rodolphe Jenatton",
            "Mark Collier",
            "Efi Kokiopoulou",
            "Pascal Frossard"
          ],
          "decision": null,
          "pdate": 1714610398356,
          "cdate": 1706802783444
        }
      ]
    }
  },
  {
    "paper_id": 43487,
    "title": "Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities",
    "abstract": "Contrastive learning methods, such asCLIP, leverage naturally paired data—for example, images and their corresponding text captions—to learn general representations that transfer efficiently to downstream tasks.\nWhile such approaches are generally applied to two modalities, domains such as robotics, healthcare, and video need to support many types of data at once.\nWe show that the pairwise application ofCLIPfails to capture joint information between modalities, thereby limiting the quality of the learned representations.\nTo address this issue, we presentSymile, a simple contrastive learning approach that captures higher-order information between any number of modalities.Symileprovides a flexible, architecture-agnostic objective for learning modality-specific representations.\nTo developSymile’s objective, we derive a lower bound on total correlation, and show thatSymilerepresentations for any set of modalities form a sufficient statistic for predicting the remaining modalities.Symileoutperforms pairwiseCLIP, even with modalities missing in the data, on cross-modal classification and retrieval across several experiments including on an original multilingual dataset of 33M image, text and audio samples and a clinical dataset of chest X-rays, electrocardiograms, and laboratory measurements.\nAll datasets and code used in this work are publicly available athttps://github.com/rajesh-lab/symile.",
    "introduction": "Contrastive learning leverages naturally paired data to learn general representations that transfer efficiently to downstream tasks[53,35,3].\nA common contrastive approach is to maximize the mutual information between the paired modalities, ensuring that the learned representations retain sensitivity to all correlations between them.\nWhile SimCLR[12]popularized the use of the mutual information estimator InfoNCE[38]for data augmentations, CLIP[40]applied the approach to distinct modalities—for example, images and their corresponding text captions—where representations are learned using any encoder for each modality.\n\n\nWhile contrastive approaches are generally applied to two modalities, there is a rapidly expanding range of domains that require the integration of many types of data at once.\nFor example, in robotics, agents combine information from visual, proprioceptive, and tactile sensors[18,28];\nhealthcare providers analyze various types of patient data including imaging, biosignals, and genomics[10,29]; and video encompasses RGB frames, audio waveforms, and text transcripts[55].\nOne strategy for handling multimodal data has been to design specialized architectures capable of processing all data types at once, which limits their general applicability and increases operational complexity[2,47].\nAnother common approach is to apply two-modality contrastive objectives, such asCLIP, to pairs of available modalities[15,44].\n\n\nIn this paper, we show that, despite its popularity, the pairwise application ofCLIPfails to capture higher-order conditional information between modalities, thereby limiting the quality of the representations it learns.\nFor instance, given three modalities𝐚𝐚\\mathbf{a}bold_a,𝐛𝐛\\mathbf{b}bold_b, and𝐜𝐜\\mathbf{c}bold_c, pairwiseCLIPcaptures dependencies between𝐚𝐚\\mathbf{a}bold_aand𝐛𝐛\\mathbf{b}bold_b,𝐛𝐛\\mathbf{b}bold_band𝐜𝐜\\mathbf{c}bold_c, and𝐚𝐚\\mathbf{a}bold_aand𝐜𝐜\\mathbf{c}bold_c, yet cannot capture any conditional dependencies, such as between𝐚𝐚\\mathbf{a}bold_aand𝐛𝐛\\mathbf{b}bold_bgiven𝐜𝐜\\mathbf{c}bold_c.\nWe show inSection2.2that even in a simple one-dimensional controlled setting where the target𝐛𝐛\\mathbf{b}bold_bis perfectly predictable from𝐚𝐚\\mathbf{a}bold_aand𝐜𝐜\\mathbf{c}bold_c,CLIPperforms no better than random chance.\nEffective contrastive learning for more than two modalities requires a model-agnostic approach capable of learning modality-specific representations—like CLIP—yet also captures higher-order information betweenanynumber of modalities—unlike CLIP.\n\n\nMethodological contributions.This paper presentsSymile, a simple contrastive learning approach that captures higher-order information between any number of modalities.Symileprovides a flexible, architecture-agnostic objective for learning modality-specific representations.\nTo developSymile’s objective, we derive a total correlation estimator, employing a generalization of inner products to more than two vectors that allows for the simultaneous contrasting of all modalities and enables zero-shot applications such as classification and retrieval.\nWe then show that the representations produced bySymilefor any set of modalities form a sufficient statistic for predicting the remaining modalities not considered in the set.\nBecause it targets total correlation,Symilecapturesstrictly moreinformation thanCLIP, guaranteeing performance that matches or surpassesCLIP, except in cases where it known thatonlypairwise statistics are relevant.\nGiven that such prior knowledge is rarely available,Symileshould be favored overCLIP.\n\n\nEmpirical contributions.We demonstrate thatSymileoutperforms pairwiseCLIPon cross-modal classification and retrieval across several experiments including on a multilingual dataset of images, text and audio of over 33M examples and a clinical dataset of chest X-rays, electrocardiograms, and laboratory measurements.\nWe show thatSymileretains its advantage over pairwiseCLIPeven with modalities missing in the data. We publicly release both the multilingual and the clinical datasets, which are specifically designed to test a model’s ability to capture higher-order information between three distinct high-dimensional data types.\n\n",
    "related_work": "Contrastive learning beyond two modalities.As discussed, previous work has extended contrastive learning to multiple modalities by applyingCLIPto pairs of available modalities.Tian et al. [49]distinguish between two such pairwise approaches: core view and full graph.\nThe core view strategy fixes one modality and then averages the loss terms between that primary modality and each of the other modalities[1,11,44].\nImageBind[15]exemplifies this approach, usingCLIPto align image embeddings with embeddings from five other modalities: text, audio, depth, thermal, and motion sensor data.\nOne advantage of this strategy is that it avoids the need for datasets with all modalities (though each dataset must still align with a primary modality).\nAs discussed inSections3.2and5.2,Symilerepresentations can be learned even with modalities missing in the data.The full graph strategy—which we have referred to as pairwiseCLIPin this paper—is to consider all(M2)binomial𝑀2{M\\choose 2}( binomial start_ARG italic_M end_ARG start_ARG 2 end_ARG )contrastive losses[43,9,14,34,33].\nFor example,Guzhov et al. [19]extendCLIPto include audio with text-to-image, text-to-audio, and image-to-audio losses.\nWhile this pairwise strategy captures strictly more information than the one used by ImageBind, neither pairwise approach is able to capture the higher-order information thatSymiledoes.PairwiseCLIPhas also been applied to architecture-specific fusion models that simultaneously process modalities to capture cross-modal interactions[2,52,21].\nFor example,Shvetsova et al. [47]train a Transformer to accept any number of modalities, using a weighted sum of contrastive losses across all input combinations.\nSuch fusion approaches face a combinatorial explosion not only in the number of weighting coefficients to tune, but also in the number of forward passes required per batch.\nIn contrast,Symileis architecture-agnostic and can learn modality-specific representations.\n\n\nTargeting higher-order information with contrastive learning.The use of contrastive methods to target higher-order information has been explored primarily within the context of multiple augmentations of the same data.\nFor instance,Bai et al. [5]derive a total correlation estimator by recursively decomposing total correlation into a summation of mutual information terms, to which variational estimators are applied (in contrast, Symile optimizes only a single term when targeting total correlation).\nThey then use their estimator to maximize the total correlation between four text augmentations.Shidani et al. [46]develop a pairwise contrastive approach for image representation learning by generalizing a lower bound on mutual information to one-vs-rest mutual information across multiple augmentations.Liang et al. [31]maximize the information in two modalities for a specific downstream task by targeting higher-order information.The relationship between these studies and our work is analogous to that between SimCLR[12]and CLIP.\nSimCLR popularized the use of the InfoNCE mutual information estimator for contrastive learning on two data augmentations.\nBuilding on this framework, CLIP applied the approach to distinct modalities, where representations are learned separately for each modality using any encoder.\nSimilarly, while existing work leverages total correlation or mutual information estimators for multi-augmentation contrastive learning, to our knowledge only pairwise applications of CLIP have applied such estimators to more than two distinct modalities.\nOur work parallels the contributions of InfoNCE and CLIP for cases involving more than two modalities: like InfoNCE, we develop a simple estimator that recovers all possible information between any number of modalities, and like CLIP, we show how this estimator can be used to learn modality-specific representations using any encoder.\n\n",
    "origin_citations": [
      "Contrastive multiview coding.",
      "Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text.",
      "Valor: Vision-audio-language omni-perception pretraining model and dataset.",
      "Accommodating audio modality in clip for multimodal processing.",
      "Imagebind: One embedding space to bind them all.",
      "Avlnet: Learning audio-visual language representations from instructional videos.",
      "Multimodal clustering networks for self-supervised learning from unlabeled videos.",
      "Routing with self-attention for multimodal capsule networks.",
      "Imu2clip: Multimodal contrastive learning for imu motion sensors from egocentric videos and text.",
      "Hybrid contrastive learning of tri-modal representation for multimodal sentiment analysis.",
      "Audioclip: Extending clip to image, text and audio.",
      "Self-supervised multimodal versatile networks.",
      "Clip-vip: Adapting pre-trained image-text model to video-language representation alignment.",
      "Clover: Towards a unified video-language alignment and fusion model.",
      "Everything at once - multi-modal fusion transformer for video retrieval.",
      "Estimating total correlation with mutual information estimators.",
      "Poly-view contrastive learning.",
      "Factorized contrastive learning: Going beyond multi-view redundancy.",
      "A simple framework for contrastive learning of visual representations."
    ],
    "citations": [
      2525,
      5342,
      54599,
      76304,
      2236,
      169883,
      163804,
      153199,
      334325,
      172332,
      393006,
      167095,
      228445,
      422281,
      5313,
      198000,
      289759,
      305836,
      7000
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 95,
      "Final_Score": 90.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 95\nFinal_Score: 91"
    },
    "overall_score": 88.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "8PWvdaRQAu",
        "forum": "8PWvdaRQAu",
        "number": 14964,
        "title": "Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities",
        "normalized_title": "contrasting with symile simple model agnostic representation learning for unlimited modalities",
        "abstract": "Contrastive learning methods, such as CLIP, leverage naturally paired data—for example, images and their corresponding text captions—to learn general representations that transfer efficiently to downstream tasks. While such approaches are generally applied to two modalities, domains such as robotics, healthcare, and video need to support many types of data at once. We show that the pairwise application of CLIP fails to capture joint information between modalities, thereby limiting the quality of the learned representations. To address this issue, we present Symile, a simple contrastive learning approach that captures higher-order information between any number of modalities. Symile provides a flexible, architecture-agnostic objective for learning modality-specific representations. To develop Symile's objective, we derive a lower bound on total correlation, and show that Symile representations for any set of modalities form a sufficient statistic for predicting the remaining modalities. Symile outperforms pairwise CLIP, even with modalities missing in the data, on cross-modal classification and retrieval across several experiments including on an original multilingual dataset of 33M image, text and audio samples and a clinical dataset of chest X-rays, electrocardiograms, and laboratory measurements. All datasets and code used in this work are publicly available at https://github.com/rajesh-lab/symile.",
        "authors": [
          "Adriel Saporta",
          "Aahlad Manas Puli",
          "Mark Goldstein",
          "Rajesh Ranganath"
        ],
        "decision": null,
        "pdate": 1727288084010,
        "cdate": 1715755644647
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "8PWvdaRQAu",
          "forum": "8PWvdaRQAu",
          "number": 14964,
          "title": "Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities",
          "normalized_title": "contrasting with symile simple model agnostic representation learning for unlimited modalities",
          "abstract": "Contrastive learning methods, such as CLIP, leverage naturally paired data—for example, images and their corresponding text captions—to learn general representations that transfer efficiently to downstream tasks. While such approaches are generally applied to two modalities, domains such as robotics, healthcare, and video need to support many types of data at once. We show that the pairwise application of CLIP fails to capture joint information between modalities, thereby limiting the quality of the learned representations. To address this issue, we present Symile, a simple contrastive learning approach that captures higher-order information between any number of modalities. Symile provides a flexible, architecture-agnostic objective for learning modality-specific representations. To develop Symile's objective, we derive a lower bound on total correlation, and show that Symile representations for any set of modalities form a sufficient statistic for predicting the remaining modalities. Symile outperforms pairwise CLIP, even with modalities missing in the data, on cross-modal classification and retrieval across several experiments including on an original multilingual dataset of 33M image, text and audio samples and a clinical dataset of chest X-rays, electrocardiograms, and laboratory measurements. All datasets and code used in this work are publicly available at https://github.com/rajesh-lab/symile.",
          "authors": [
            "Adriel Saporta",
            "Aahlad Manas Puli",
            "Mark Goldstein",
            "Rajesh Ranganath"
          ],
          "decision": null,
          "pdate": 1727288084010,
          "cdate": 1715755644647
        }
      ]
    }
  },
  {
    "paper_id": 11651,
    "title": "AnyText: Multilingual Visual Text Generation and Editing",
    "abstract": "Diffusion model based Text-to-Image has achieved impressive achievements recently. Although current technology for synthesizing images is highly advanced and capable of generating images with high fidelity, it is still possible to give the show away when focusing on the text area in the generated image, as synthesized text often contains blurred, unreadable, or incorrect characters, making visual text generation one of the most challenging issues in this field. To address this issue, we introduceAnyText, a diffusion-based multilingual visual text generation and editing model, that focuses on rendering accurate and coherent text in the image. AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy. AnyText can write characters in multiple languages, to the best of our knowledge, this is the first work to address multilingual visual text generation. It is worth mentioning that AnyText can be plugged into existing diffusion models from the community for rendering or editing text accurately. After conducting extensive evaluation experiments, our method has outperformed all other approaches by a significant margin.\nAdditionally, we contribute the first large-scale multilingual text images dataset,AnyWord-3M, containing 3 million image-text pairs with OCR annotations in multiple languages. Based on AnyWord-3M dataset, we propose AnyText-benchmark for the evaluation of visual text generation accuracy and quality. Our project will be open-sourced soon\nonhttps://github.com/tyxsspa/AnyTextto improve and promote the development of text generation technology.",
    "introduction": "Figure 1:Selected samples of AnyText. For text generation, AnyText can render the specified text from the prompt onto the designated position, and generate visually appealing images. As for text editing, AnyText can modify the text content at the specified position within the input image while maintaining consistency with the surrounding text style. Translations are provided in parentheses for non-English words in prompt, blue boxes indicate positions for text editing. See more inA.6.\n\n\nDiffusion-based generative modelsSaharia et al. (2022); Ramesh et al. (2022); Rombach et al. (2022); Zhang & Agrawala (2023)demonstrated exceptional outcomes with unparalleled fidelity, adaptability, and versatility. Open-sourced image generation models (for example, Stable DiffusionRombach et al. (2022), DeepFloyd-IFDeepFloyd-Lab (2023)), along with commercial services (MidjourneyInc. (2022), DALL-E 2Ramesh et al. (2022), etc.) have made substantial impacts in various sectors including photography, digital arts, gaming, advertising, and film production. Despite substantial progress in image quality of generative diffusion models, most current open-sourced models and commercial services struggle to produce well-formed, legible, and readable visual text. Consequently, this reduces their overall utility and hampers potential applications.\n\n\nThe subpar performance of present open-source diffusion-based models can be attributed to a number of factors. Firstly, there is a deficiency in large-scale image-text paired data which includes comprehensive annotations for textual content. Existing datasets for large-scale image diffusion model training, like LAION-5BSchuhmann et al. (2022), lack manual annotations or OCR results for text content. Secondly, as pointed out inLiu et al. (2023), the text encoder used in open-source diffusion models, such as the CLIP text encoder, employs a vocabulary-based tokenizer that has no direct access to the characters, resulting in a diminished sensitivity to individual characters. Lastly, most diffusion models’ loss functions are designed to enhance the overall image generation quality and lack of dedicated supervision for the text region.\n\n\nTo address the aforementioned difficulties, we presentAnyTextframework andAnyWord-3Mdataset. AnyText consists of a text-control diffusion pipeline with two components: auxiliary latent module encodes auxiliary information such as text glyph, position, and masked image into latent space to assist text generation and editing; text embedding module employs an OCR model to encode stroke information as embeddings, which is then fused with image caption embeddings from the tokenizer to render texts blended with the background seamlessly; and finally, a text perceptual loss in image space is introduced to further enhance writing accuracy.\n\n\nRegarding the functionality, there are five differentiating factors that set apart us from other competitors as outlined in Table1: a)Multi-line: AnyText can generate text on multiple lines at user-specified positions. b)Deformed regions: it enables writing in horizontally, vertically, and even curved or irregular regions. c)Multi-lingual: our method can generate text in various languages such as Chinese, English, Japanese, Korean, etc. d)Text editing: which provides the capability for modifying the text content within the provided image in consistent font style. e)Plug-and-play: AnyText can be seamlessly integrated with stable diffusion models and empowering them with the ability to generate text. We present some selected examples in Fig.1and AppendixA.6.\n\n\nTable 1:Comparison of AnyText with other competitors based on functionality.FunctionalityMulti-lineDeformed regionsMulti-lingualText editingPlug-and-playGlyphDraw×\\times××\\times××\\times××\\times××\\times×TextDiffuser✓×\\times××\\times×✓×\\times×GlyphControl✓×\\times××\\times××\\times×✓AnyText✓✓✓✓✓\n\n",
    "related_work": "Text-to-Image Synthesis. In recent years, significant strides has been made in text-to-image synthesis using denoising diffusion probabilistic modelsHo et al. (2020); Ramesh et al. (2021); Song et al. (2021); Dhariwal & Nichol (2021); Nichol & Dhariwal (2021); Saharia et al. (2022); Ramesh et al. (2022); Rombach et al. (2022); Chang et al. (2023). These models have advanced beyond simple image generation and have led to developments in interactive image editingMeng et al. (2022); Gal et al. (2023); Brooks et al. (2022)and techniques incorporating additional conditions, such as masks and depth mapsRombach et al. (2022). Research is also exploring the area of multi-condition controllable synthesisZhang & Agrawala (2023); Mou et al. (2023); Huang et al. (2023). Compositing subjects into scenes presents more specific challenges, and approaches like ELITEWei et al. (2023), UMM-DiffusionMa et al. (2023b), and InstantBoothShi et al. (2023)utilize the features from CLIP image encoder to encode the visual concept into textual word embeddings. Similarly, DreamIdentityChen et al. (2023c)developed a specially designed image encoder to achieve better performance for the word embedding enhancing scheme.\n\n\nText Generation. Progress in image synthesis has been substantial, but integrating legible text into images remains challengingRombach et al. (2022); Saharia et al. (2022). Recent research has focused on three key aspects of text generation:\n\n\nControl Condition.Introducing glyph condition in latent space has been a predominant approach in many recent methods. GlyphDrawMa et al. (2023a)originally used an explicit glyph image as condition, with characters rendered at the center. GlyphControlYang et al. (2023)further extends it by aligning the text based on its location, which also incorporates font size and text box position in an implicit manner. TextDiffuserChen et al. (2023b)utilizes a character-level segmentation mask as control condition, and in addition, it introduces a masked image to simultaneously learn text generation and text in-painting branches. In our work, we adopt a similar way as GlyphControl to render glyph image but incorporate position and masked image as additional conditions. This design enables AnyText to generate text in curved or irregular regions, and handle text generation and editing simultaneously.\n\n\nText Encoder.The text encoder plays a crucial role in generating accurate visual text. Recent methods such as ImagenSaharia et al. (2022), eDiff-IBalaji et al. (2022), and Deepfloyd IFDeepFloyd-Lab (2023)achieve impressive results by leveraging large-scale language models (e.g., T5-XXL). However, most image generation models still rely on character-blind text encoders, and even character-aware text encoders struggle with non-Latin text generation like Chinese, Japanese, and KoreanLiu et al. (2023). To address Chinese rendering, GlyphDraw fine-tunes the text encoder on Chinese images and utilizes the CLIP image encoder for glyph embeddingsMa et al. (2023a). DiffUTE replaces the text encoder with a pre-trained image encoder to extract glyphs in image editing scenariosChen et al. (2023a). In AnyText, we propose a novel approach to transform the text encoder by integrating semantic and glyph information. This aims to achieve seamless integration of generated text with the background and enable multi-language text generation.\n\n\nPerceptual Supervision.OCR-VQGANRodriguez et al. (2023)employs a pre-trained OCR detection model to extract features from images, and supervise text generation by constraining the differences between multiple intermediate layers. In contrast, TextDiffuserChen et al. (2023b)utilizes a character-level segmentation model to supervise the accuracy of each generated character in the latent space. This approach requires a separately trained segmentation model, and the character classes are also limited. In AnyText, we utilize an OCR recognition model that excels in stroke and spelling to supervise the text generation within the designated text region only. This approach provides a more direct and effective form of supervision for ensuring accurate and high-quality text generation.\n\n",
    "origin_citations": [
      "Denoising diffusion probabilistic models.",
      "Zero-shot text-to-image generation.",
      "Score-based generative modeling through stochastic differential equations.",
      "Diffusion models beat gans on image synthesis.",
      "Improved denoising diffusion probabilistic models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Hierarchical text-conditional image generation with CLIP latents.",
      "High-resolution image synthesis with latent diffusion models.",
      "Muse: Text-to-image generation via masked generative transformers.",
      "Sdedit: Guided image synthesis and editing with stochastic differential equations.",
      "An image is worth one word: Personalizing text-to-image generation using textual inversion.",
      "Instructpix2pix: Learning to follow image editing instructions.",
      "Adding conditional control to text-to-image diffusion models.",
      "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.",
      "Composer: Creative and controllable image synthesis with composable conditions.",
      "ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation.",
      "Unified multi-modal latent diffusion for joint subject and text conditional image generation.",
      "Instantbooth: Personalized text-to-image generation without test-time finetuning.",
      "Dreamidentity: Improved editability for efficient face-identity preserved image generation.",
      "Glyphdraw: Learning to draw chinese characters in image synthesis models coherently.",
      "Glyphcontrol: Glyph conditional control for visual text generation.",
      "Textdiffuser: Diffusion models as text painters.",
      "ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers.",
      "Deepfloyd if.",
      "Character-aware models improve visual text rendering.",
      "Diffute: Universal text editing diffusion model.",
      "OCR-VQGAN: taming text-within-image generation."
    ],
    "citations": [
      4827,
      7339,
      431228,
      165,
      5920,
      5926,
      2558,
      2240,
      303968,
      96204,
      236013,
      2235,
      489610,
      228018,
      227373,
      225547,
      496101,
      19656,
      187973,
      1087045,
      256999,
      261369,
      34739,
      139671,
      261390,
      336122
    ],
    "extract_rate": 0.9629629629629629,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 87,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 90,
      "Final_Score": 89.9,
      "raw_text": "Content_Coherence: 87\nCitation_Quality: 92\nSynthesis_Analysis: 90\nFinal_Score: 90"
    },
    "overall_score": 88.54,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "ezBH9WE9s2",
        "forum": "ezBH9WE9s2",
        "number": 3269,
        "title": "AnyText: Multilingual Visual Text Generation and Editing",
        "normalized_title": "anytext multilingual visual text generation and editing",
        "abstract": "Diffusion model based Text-to-Image has achieved impressive achievements recently. Although current technology for synthesizing images is highly advanced and capable of generating images with high fidelity, it is still possible to give the show away when focusing on the text area in the generated image, as synthesized text often contains blurred, unreadable, or incorrect characters, making visual text generation one of the most challenging issues in this field. To address this issue, we introduce AnyText, a diffusion-based multilingual visual text generation and editing model, that focuses on rendering accurate and coherent text in the image. AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy. AnyText can write characters in multiple languages, to the best of our knowledge, this is the first work to address multilingual visual text generation. It is worth mentioning that AnyText can be plugged into existing diffusion models from the community for rendering or editing text accurately. After conducting extensive evaluation experiments, our method has outperformed all other approaches by a significant margin. Additionally, we contribute the first large-scale multilingual text images dataset, AnyWord-3M, containing 3 million image-text pairs with OCR annotations in multiple languages. Based on AnyWord-3M dataset, we propose AnyText-benchmark for the evaluation of visual text generation accuracy and quality. Our project will be open-sourced soon to improve and promote the development of text generation technology.",
        "authors": [
          "Yuxiang Tuo",
          "Wangmeng Xiang",
          "Jun-Yan He",
          "Yifeng Geng",
          "Xuansong Xie"
        ],
        "decision": null,
        "pdate": 1705410869328,
        "cdate": 1695283051195
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "ezBH9WE9s2",
          "forum": "ezBH9WE9s2",
          "number": 3269,
          "title": "AnyText: Multilingual Visual Text Generation and Editing",
          "normalized_title": "anytext multilingual visual text generation and editing",
          "abstract": "Diffusion model based Text-to-Image has achieved impressive achievements recently. Although current technology for synthesizing images is highly advanced and capable of generating images with high fidelity, it is still possible to give the show away when focusing on the text area in the generated image, as synthesized text often contains blurred, unreadable, or incorrect characters, making visual text generation one of the most challenging issues in this field. To address this issue, we introduce AnyText, a diffusion-based multilingual visual text generation and editing model, that focuses on rendering accurate and coherent text in the image. AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy. AnyText can write characters in multiple languages, to the best of our knowledge, this is the first work to address multilingual visual text generation. It is worth mentioning that AnyText can be plugged into existing diffusion models from the community for rendering or editing text accurately. After conducting extensive evaluation experiments, our method has outperformed all other approaches by a significant margin. Additionally, we contribute the first large-scale multilingual text images dataset, AnyWord-3M, containing 3 million image-text pairs with OCR annotations in multiple languages. Based on AnyWord-3M dataset, we propose AnyText-benchmark for the evaluation of visual text generation accuracy and quality. Our project will be open-sourced soon to improve and promote the development of text generation technology.",
          "authors": [
            "Yuxiang Tuo",
            "Wangmeng Xiang",
            "Jun-Yan He",
            "Yifeng Geng",
            "Xuansong Xie"
          ],
          "decision": null,
          "pdate": 1705410869328,
          "cdate": 1695283051195
        }
      ]
    }
  },
  {
    "paper_id": 25244,
    "title": "Looking Backward: Streaming Video-to-Video Translation with Feature Banks",
    "abstract": "This paper introduces StreamV2V, a diffusion model that achieves real-time streaming video-to-video (V2V) translation with user prompts.\nUnlike prior V2V methods using batches to process a limited number of frames, we opt to process frames in streaming fashion, to support an unlimited number of frames.\nAt the heart of StreamV2V lies abackward-lookingapproach that relates the present to the past.\nThis is realized by maintaining a feature bank that archives information from past frames.\nFor incoming frames, StreamV2V extends self-attention to include banked keys and values, and directly fuses similar past features into the output.\nThe feature bank is continually updated by merging stored and new features, making it compact yet informative.\nStreamV2V stands out for its adaptability and efficiency, seamlessly integrating with image diffusion models without fine-tuning.\nStreamV2V can run 20 FPS on one A100 GPU, being 15×\\times×, 46×\\times×, 108×\\times×, and 158×\\times×faster than FlowVid, CoDeF, Rerender, and TokenFlow, respectively.\nQuantitative metrics and user studies confirm StreamV2V’s exceptional ability to maintain temporal consistency. Demo, code, and models are available on theproject page.",
    "introduction": "Text-driven video-to-video (V2V) translation, which aims to alter the input video following given prompts, has wide applications, such as creating short videos, and more broadly in the film industry.\nMost diffusion model based methods(Wu et al.,2023b; Yang et al.,2023; Ouyang et al.,2023; Wang et al.,2023a; Khachatryan et al.,2023; Qi et al.,2023; Zhang et al.,2023; Wang et al.,2023b; Chen et al.,2023; Zhao et al.,2023; Geyer et al.,2023; Liang et al.,2023; Wu et al.,2023a; Singer et al.,2024)use batches to process recorded videos, as shown in Figure2(a).\nHowever, batch processing necessitates loading all frames into GPU memory, thereby limiting the video length they can handle, typically up to 4 seconds.\nFurthermore, these methods do not accommodate real-time translation and typically require several minutes to process a single 4-second clip.\n\n\nFigure 2:(a) Existing V2V methods process frames in batches, restricting them to a limited number of frames. (b) Our StreamV2V framework processes frames in streaming fashion, can operate on streaming videos in real-time. (c) Batch processing requiresO⁢(N)𝑂𝑁O(N)italic_O ( italic_N )memory for the video lengthN𝑁Nitalic_N, whereas our StreamV2V only needsO⁢(1)𝑂1O(1)italic_O ( 1 )memory regardless of video length.\n\n\nThis paper targets streaming V2V applications, such as webcam video translation and AI-assisted drawing, where users want to modify the streaming video iteratively.\nThis necessitates the model to handle input videos of varying lengths and perform real-time translation.\nTo tackle this challenge, we introduce StreamV2V, an approach that processes frames in streaming fashion, as shown in Figure2(b).\nLeveraging advancements in one-/few-step diffusion models(Song et al.,2023; Sauer et al.,2023; Luo et al.,2023b), StreamDiffusion(Kodaira et al.,2023)has designed a pipeline for real-time interactive image generation.\nHowever, directly applying StreamDiffusion for V2V tasks leads to noticeable pixel flickering across frames. This is because StreamDiffusion treats each frame independently, disregarding the continuity of videos.\nIn contrast, humans implicitly memorize visual elements across frames and see the current frame as it links with past observations.\nTo generate consistent videos, it is critical to integrate a mechanism that can effectively bridge the current frame to its predecessors.\n\n\nRecent studies have shown that diffusion features(Tang et al.,2024; Luo et al.,2023a)captured during U-Net’s forward process contain rich correspondences between images.\nInspired by this, our StreamV2V maintains a feature bank, which stores the intermediate features of past frames.\nFor incoming frames, we extend self-attention by incorporating the corresponding stored keys and values.\nThis can be interpreted as a weighted sum of similar regions across frames via attention, effectively aligning the current frame with previous frames.\nAdditionally, to ensure the consistency of fine-grained details, we directly fuse the block’s output with similar features from past frames.\n\n\nThe challenge then becomes:How can we implement this feature bank?A straightforward approach might store a constant number of frames, such as employing a sliding window technique.\nHowever, this method is sub-optimal, as it inadvertently discards valuable data when a frame is omitted, and generates redundancy when the stored frames are similar.\nTo address this, we propose to continuously update the bank by merging redundant features within incoming and stored features.\nThis allows us to preserve the most representative features while keeping a consistent bank size over time.\nThrough our experiment, we find the feature bank can be condensed to the size needed to store just one frame.\n\n\nStreamV2V requires no training or fine-tuning, making it compatible as an add-on with any image diffusion models.\nIt excels in efficiency, capable of processing high-resolution video (512×\\times×512) in real-time at 20 frames per second (FPS) on a single A100 GPU.\nThis speed surpasses current V2V methods—FlowVid(Liang et al.,2023), CoDeF(Ouyang et al.,2023), Rerender(Yang et al.,2023), and TokenFlow(Geyer et al.,2023)—by factors of 15×\\times×, 46×\\times×, 108×\\times×, and 158×\\times×, respectively.\nWe evaluate our method with quantitative metrics, such as CLIP score(Radford et al.,2021)and warp error(Lai et al.,2018), and a user study.\nOur findings indicate that users significantly favor our StreamV2V over StreamDiffusion(Kodaira et al.,2023)(with over 70% win rates) and CoDeF(Ouyang et al.,2023)(with over 80% win rates). While our method may not yet match the performance of state-of-the-art (SOTA) methods like TokenFlow and FlowVid, its rapid real-time execution opens up new venues for streaming V2V applications.\n\n\nOur contributions are three-fold: (1) To the best of our knowledge, our approach is the first approach to tackle real-time video-to-video translation for streaming videos. (2) Our method, StreamV2V, employs a simple yet effective looking-backward principle by maintaining a feature bank to improve consistency. (3) We develop a dynamic feature bank updating strategy that merges redundant features, ensuring the feature bank remains both compact and descriptive.\n\n",
    "related_work": "2.1Video-to-video translationSignificant progress has been made in the domain of text-guided image-to-image (I2I) translation(Brooks et al.,2023; Hertz et al.,2022; Tumanyan et al.,2023; Mou et al.,2023), greatly supported by large pre-trained text-to-image diffusion models(Ramesh et al.,2022; Rombach et al.,2022; Saharia et al.,2022).\nSimilarly, video-to-video (V2V) translation, which aims to generate consistent videos, has attracted substantial interest as well.\nTo generate coherent multiple frames, most existing works(Esser et al.,2023; Wu et al.,2023b; Wang et al.,2023a; Guo et al.,2023; Chen et al.,2023; Khachatryan et al.,2023; Ceylan et al.,2023; Qi et al.,2023; Geyer et al.,2023; Wu et al.,2023a; Liang et al.,2023; Ku et al.,2024; Singer et al.,2024)process batches of frames simultaneously with cross-frame attention mechanisms.\nHowever, as the memory usage increases with an increased number of frames, these methods are typically constrained to about 4 seconds length.\nAdditionally, they tend to rely on expensive DDIM inversion(Song et al.,2020; Qi et al.,2023; Geyer et al.,2023)or optical flow computation(Yang et al.,2023; Liang et al.,2023), leading to long processing time.\nIn contrast, our StreamV2V handles videos in real-time and in streaming fashion, allowing for processing videos of any length.\nCompared to concurrent work Live2Diff(Xing et al.,2024)that requires additional video fine-tuning to train uni-directional temporal attention, our method is training-free and can serve as an add-on for any image diffusion model.\n\n\n2.2Accelerating diffusion modelsWhile achieving great generation quality, diffusion models are commonly limited by their slow speed due to the need for multiple denoising steps.\nRecent advancements have introduced reusing cached features in denoising steps(Ma et al.,2023; Wimbauer et al.,2023)and one-/few-step diffusion models through distillation(Song et al.,2023; Meng et al.,2023; Luo et al.,2023b; Sauer et al.,2023; Yin et al.,2023; Lin & Yang,2024).\nStreamDiffusion(Kodaira et al.,2023)proposes a pipeline to leverage these developments for real-time image generation.\nHowever, its application to video without adjustments brings unsatisfactory results.\nLeveraging StreamDiffusion’s groundwork, we enhance frame consistency by implementing a backward-looking feature bank.\nOur approach introduces a dynamic merging technique for the feature bank, ensuring it remains compact and incurs minimal additional computational cost in comparison to StreamDiffusion.\n\n\n2.3Feature banksLong-term feature banks(Wu et al.,2019; Pan et al.,2021)have been used in video understanding as supportive context features to help reason the entire video.\nAdapting this concept, we employ feature banks to enhance video generation consistency.\nTo ensure our feature bank remains both informative and compact, we introduce a dynamic feature merging strategy.\nWhile token merging(Bolya et al.,2022; Bolya & Hoffman,2023)is a common method for merging similar features within single images, our approach extends this technique across video frames, differentiating it from traditional within-image operations.\n\n",
    "origin_citations": [
      "Instructpix2pix: Learning to follow image editing instructions.",
      "Prompt-to-prompt image editing with cross attention control.",
      "Plug-and-play diffusion features for text-driven image-to-image translation.",
      "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.",
      "Hierarchical text-conditional image generation with clip latents.",
      "High-resolution image synthesis with latent diffusion models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Structure and content-guided video synthesis with diffusion models.",
      "Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation.",
      "Zero-shot video editing using off-the-shelf image diffusion models.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Control-a-video: Controllable text-to-video generation with diffusion models.",
      "Text2video-zero: Text-to-image diffusion models are zero-shot video generators.",
      "Pix2video: Video editing using image diffusion.",
      "Fatezero: Fusing attentions for zero-shot text-based video editing.",
      "Tokenflow: Consistent diffusion features for consistent video editing.",
      "Fairy: Fast parallelized instruction-guided video-to-video synthesis.",
      "Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis.",
      "Anyv2v: A plug-and-play framework for any video-to-video editing tasks.",
      "Video editing via factorized diffusion distillation.",
      "Denoising diffusion implicit models.",
      "Rerender a video: Zero-shot text-guided video-to-video translation.",
      "Live2diff: Live stream translation via uni-directional attention in video diffusion models.",
      "Deepcache: Accelerating diffusion models for free.",
      "Cache me if you can: Accelerating diffusion models through block caching.",
      "Consistency models.",
      "On distillation of guided diffusion models.",
      "Latent consistency models: Synthesizing high-resolution images with few-step inference.",
      "Adversarial diffusion distillation.",
      "One-step diffusion with distribution matching distillation.",
      "Animatediff-lightning: Cross-model diffusion distillation.",
      "Streamdiffusion: A pipeline-level solution for real-time interactive generation.",
      "Long-term feature banks for detailed video understanding.",
      "Actor-context-actor relation network for spatio-temporal action localization.",
      "Token merging: Your vit but faster.",
      "Token merging for fast stable diffusion."
    ],
    "citations": [
      2235,
      236009,
      243990,
      228018,
      2558,
      2240,
      5926,
      490753,
      139303,
      492165,
      185878,
      259666,
      494137,
      494475,
      496009,
      469944,
      348373,
      346070,
      425598,
      287743,
      356595,
      304517,
      864337,
      274524,
      273315,
      5919,
      460643,
      322408,
      22514,
      426506,
      348493,
      358985,
      349613,
      336733,
      492160
    ],
    "extract_rate": 0.9722222222222222,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 91,
      "Final_Score": 88.2,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 91\nFinal_Score: 88"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 88,
      "Final_Score": 88.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 88\nFinal_Score: 88"
    },
    "overall_score": 88.5,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "AMkf7h7HER",
        "forum": "AMkf7h7HER",
        "number": 6311,
        "title": "Looking Backward: Streaming Video-to-Video Translation with Feature Banks",
        "normalized_title": "looking backward streaming video to video translation with feature banks",
        "abstract": "This paper introduces StreamV2V, a diffusion model that achieves real-time streaming video-to-video (V2V) translation with user prompts. \n  Unlike prior V2V methods using batches to process limited frames, we opt to process frames in a streaming fashion, to support unlimited frames.\n  At the heart of StreamV2V lies a backward-looking principle that relates the present to the past. \n  This is realized by maintaining a feature bank, which archives information from past frames.\n  For incoming frames, StreamV2V extends self-attention to include banked keys and values, and directly fuses similar past features into the output.\n  The feature bank is continually updated by merging stored and new features, making it compact yet informative.\n  StreamV2V stands out for its adaptability and efficiency, seamlessly integrating with image diffusion models without fine-tuning.\n  It can run 20 FPS on one A100 GPU, being 15$\\times$, 46$\\times$, 108$\\times$, and 158$\\times$ faster than FlowVid, CoDeF, Rerender, and TokenFlow, respectively. \n  Quantitative metrics and user studies confirm StreamV2V's exceptional ability to maintain temporal consistency.",
        "authors": [
          "Feng Liang",
          "Akio Kodaira",
          "Chenfeng Xu",
          "Masayoshi Tomizuka",
          "Kurt Keutzer",
          "Diana Marculescu"
        ],
        "decision": null,
        "pdate": 1737562459877,
        "cdate": 1727340013153
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "AMkf7h7HER",
          "forum": "AMkf7h7HER",
          "number": 6311,
          "title": "Looking Backward: Streaming Video-to-Video Translation with Feature Banks",
          "normalized_title": "looking backward streaming video to video translation with feature banks",
          "abstract": "This paper introduces StreamV2V, a diffusion model that achieves real-time streaming video-to-video (V2V) translation with user prompts. \n  Unlike prior V2V methods using batches to process limited frames, we opt to process frames in a streaming fashion, to support unlimited frames.\n  At the heart of StreamV2V lies a backward-looking principle that relates the present to the past. \n  This is realized by maintaining a feature bank, which archives information from past frames.\n  For incoming frames, StreamV2V extends self-attention to include banked keys and values, and directly fuses similar past features into the output.\n  The feature bank is continually updated by merging stored and new features, making it compact yet informative.\n  StreamV2V stands out for its adaptability and efficiency, seamlessly integrating with image diffusion models without fine-tuning.\n  It can run 20 FPS on one A100 GPU, being 15$\\times$, 46$\\times$, 108$\\times$, and 158$\\times$ faster than FlowVid, CoDeF, Rerender, and TokenFlow, respectively. \n  Quantitative metrics and user studies confirm StreamV2V's exceptional ability to maintain temporal consistency.",
          "authors": [
            "Feng Liang",
            "Akio Kodaira",
            "Chenfeng Xu",
            "Masayoshi Tomizuka",
            "Kurt Keutzer",
            "Diana Marculescu"
          ],
          "decision": null,
          "pdate": 1737562459877,
          "cdate": 1727340013153
        }
      ]
    }
  },
  {
    "paper_id": 40686,
    "title": "Restructuring Vector Quantization with the Rotation Trick",
    "abstract": "Vector Quantized Variational AutoEncoders (VQ-VAEs) are designed to compress a continuous input to a discrete latent space and reconstruct it with minimal distortion.\nThey operate by maintaining a set of vectors—often referred to as the codebook—and quantizing each encoder output to the nearest vector in the codebook.\nHowever, as vector quantization is non-differentiable, the gradient to the encoder flowsaroundthe vector quantization layer rather thanthroughit in a straight-through approximation.\nThis approximation may be undesirable as all information from the vector quantization operation is lost.\nIn this work, we propose a way to propagate gradients through the vector quantization layer of VQ-VAEs.\nWe smoothly transform each encoder output into its corresponding codebook vector via a rotation and rescaling linear transformation that is treated as a constant during backpropagation.\nAs a result, the relative magnitude and angle between encoder output and codebook vector becomes encoded into the gradient as it propagates through the vector quantization layer and back to the encoder.\nAcross 11 different VQ-VAE training paradigms, we find this restructuring improves reconstruction metrics, codebook utilization, and quantization error. Our code is available athttps://github.com/cfifty/rotation_trick.",
    "introduction": "Vector quantization(Gray,1984)is an approach to discretize a continuous vector space. It defines a finite set of vectors—referred to as the codebook—and maps any vector in the continuous vector space to the closest vector in the codebook.\nHowever, deep learning paradigms that use vector quantization are often difficult to train because replacing a vector with its closest codebook counterpart is a non-differentiable operation(Huh et al.,2023).\nThis characteristic was not an issue at its creation during the Renaissance of Information Theory for applications like noisy channel communication(Cover,1999); however in the era deep learning, it presents a challenge as gradients cannot directly flow through layers that use vector quantization during backpropagation.\n\n\nIn deep learning, vector quantization is largely used in the eponymous Vector Quantized-Variational AutoEncoder (VQ-VAE)(Van Den Oord et al.,2017). A VQ-VAE is an AutoEncoder with a vector quantization layer between the encoder’s output and decoder’s input, thereby quantizing the learned representation at the bottleneck. While VQ-VAEs are ubiquitous in state-of-the-art generative modeling(Rombach et al.,2022; Dhariwal et al.,2020; Brooks et al.,2024), their gradients cannot flow from the decoder to the encoder uninterrupted as they must pass through a non-differentiable vector quantization layer.\n\n\nA solution to the non-differentiability problem is to approximate gradients via a “straight-through estimator” (STE)(Bengio et al.,2013). During backpropagation, the STE copies and pastes the gradients from the decoder’s input to the encoder’s output, thereby skipping the quantization operation altogether.\nHowever, this approximation can lead to poor-performing models and codebook collapse: a phenomena where a large percentage of the codebook converge to zero norm and are unused by the model(Mentzer et al.,2023). Even if codebook collapse does not occur, the codebook is often under-utilized, thereby limiting the information capacity of the VQ-VAEs’s bottleneck(Dhariwal et al.,2020).\n\n\nFigure 1:Illustration of the rotation trick. In the forward pass, encoder outpute𝑒eitalic_eis rotated and rescaled toq1subscript𝑞1q_{1}italic_q start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT.\nNot shown is the rotation of the other encoder outputs to their corresponding codebook vectors. In the backward pass, the gradient atq1subscript𝑞1q_{1}italic_q start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTmoves toe𝑒eitalic_eso that the angle between∇q1ℒsubscript∇subscript𝑞1ℒ\\nabla_{q_{1}}\\mathcal{L}∇ start_POSTSUBSCRIPT italic_q start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT end_POSTSUBSCRIPT caligraphic_Landq1subscript𝑞1q_{1}italic_q start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTis preserved. Now, points within the same codebook region receive different gradients depending on their relative angle and magnitude to the codebook vector. For example, points with high angular distance can bepushedinto new codebook regions, thereby increasing codebook utilization.\n\n\nAlgorithm 1VQ-VAE with the Rotation Trickinput examplex𝑥xitalic_xe←Encoder⁢(x)←𝑒Encoder𝑥e\\leftarrow\\text{Encoder}(x)italic_e ← Encoder ( italic_x )q←←𝑞absentq\\leftarrowitalic_q ←nearest codebook vector toe𝑒eitalic_eR←←𝑅absentR\\leftarrowitalic_R ←rotation matrix that alignse𝑒eitalic_etoq𝑞qitalic_qq~←stop-gradient⁢[∥q∥∥e∥⁢R]⁢e←~𝑞stop-gradientdelimited-[]delimited-∥∥𝑞delimited-∥∥𝑒𝑅𝑒\\tilde{q}\\leftarrow\\text{stop-gradient}\\left[\\frac{{\\lVert q\\rVert}}{{\\lVert e%\n\\rVert}}R\\right]eover~ start_ARG italic_q end_ARG ← stop-gradient [ divide start_ARG ∥ italic_q ∥ end_ARG start_ARG ∥ italic_e ∥ end_ARG italic_R ] italic_ex~←Decoder(\\tilde{x}\\leftarrow\\text{Decoder}(over~ start_ARG italic_x end_ARG ← Decoder (q~~𝑞\\tilde{q}over~ start_ARG italic_q end_ARG))))loss←ℒ⁢(x,x~)←absentℒ𝑥~𝑥\\leftarrow\\mathcal{L}(x,\\tilde{x})← caligraphic_L ( italic_x , over~ start_ARG italic_x end_ARG )returnloss\n\n\nIn this work, we propose an alternate way to propagate gradients through the vector quantization layer in VQ-VAEs.\nFor a given encoder outpute𝑒eitalic_eand nearest codebook vectorq𝑞qitalic_q, we smoothly transforme𝑒eitalic_etoq𝑞qitalic_qvia a rotation and rescaling linear transformation and then send this output—rather than the direct result of the codebook lookup—to the decoder.\nAs the input to the decoder,q~~𝑞\\tilde{q}over~ start_ARG italic_q end_ARG, is now treated as a smooth linear transformation ofe𝑒eitalic_e, gradients flow back from the decoder to the encoder unimpeded.\nTo avoid differentiating through the rotation and rescaling, we treat both as constants with respect toe𝑒eitalic_eandq𝑞qitalic_q. We explain why this choice is necessary inSectionA.2.\nFollowing the convention ofKingma & Welling (2013), we call this restructuring “the rotation trick.” It is illustrated inFigure3and described inAlgorithm1.\n\n\nThe rotation trick does not change the output of the VQ-VAE in the forward pass.\nHowever, during the backward pass, it transports the gradient∇qℒsubscript∇𝑞ℒ\\nabla_{q}\\mathcal{L}∇ start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT caligraphic_Latq𝑞qitalic_qto become the gradient∇eℒsubscript∇𝑒ℒ\\nabla_{e}\\mathcal{L}∇ start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT caligraphic_Late𝑒eitalic_eso that the angle betweenq𝑞qitalic_qand∇qℒsubscript∇𝑞ℒ\\nabla_{q}\\mathcal{L}∇ start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT caligraphic_Lafterthe vector quantization layer equals the angle betweene𝑒eitalic_eand∇eℒsubscript∇𝑒ℒ\\nabla_{e}\\mathcal{L}∇ start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT caligraphic_Lbeforethe vector quantization layer. Preserving this angle encodes relative angular distances and magnitudes into the gradient and changes how points within the same codebook region are updated.\n\n\nThe STE applies the same update to all points within the same codebook region, maintaining their relative distances. However as we will show inSection4.3, the rotation trick can push points within the same codebook region farther apart—or pull them closer together—depending on the direction of the gradient vector.\nThe former capability can correspond to increased codebook usage while the latter to lower quantization error. In the context of lossy compression, both capabilities are desirable for reducing the distortion and increasing the information capacity of the vector quantization layer.\n\n\nWhen applied to several open-source VQ-VAE repositories, we find the rotation trick substantively improves reconstruction performance, increases codebook usage, and decreases the distance between encoder outputs and their corresponding codebook vectors. For instance, training the VQGAN fromRombach et al. (2022)on ImageNet(Deng et al.,2009)with the rotation trick improves reconstruction FID from 5.0 to 1.6, reconstruction IS from 141.5 to 190.3, increases codebook usage from 2% to 9%, and decreases quantization error by over an order of magnitude.\n\n",
    "related_work": "Many researchers have built upon the seminal work ofVan Den Oord et al. (2017)to improve VQ-VAE performance. While non-exhaustive, our review focuses on methods that address training instabilities caused by the vector quantization layer. We partition these efforts into two categories: (1) methods that sidestep the STE and (2) methods that improve codebook-model interactions.\n\n\nSidestepping the STE.Several prior works have sought to fix the problems caused by the STE by avoiding deterministic vector quantization.Baevski et al. (2019)employ the Gumbel-Softmax trick(Jang et al.,2016)to fit a categorical distribution over codebook vectors that converges to a one-hot distribution towards the end of training,Gautam et al. (2023)quantize using a a convex combination of codebook vectors, andTakida et al. (2022)employ stochastic quantization. Unlike the above that cast vector quantization as a distribution over codebook vectors,Huh et al. (2023)propose an alternating optimization where the encoder is optimized to output representations close to the codebook vectors while the decoder minimizes reconstruction loss from a fixed set of codebook vector inputs. While these approaches sidestep the training instabilities caused by the STE, they can introduce their own set of problems and complexities such as low codebook utilization at inference and the tuning of a temperature schedule(Zhang et al.,2023). As a result, many applications and research papers continue to employ VQ-VAEs that are trained using the STE(Rombach et al.,2022; Chang et al.,2022; Huang et al.,2023; Zhu et al.,2023; Dong et al.,2023).\n\n\nCodebook-Model Improvements.Another way to attack codebook collapse or under-utilization is to change the codebook lookup. Rather than use Euclidean distance,Yu et al. (2021)employ a cosine similarity measure,Goswami et al. (2024)a hyperbolic metric, andLee et al. (2022)stochastically sample codes as a function of the distance between the encoder output and codebook vectors. Another perspective examines the learning of the codebook.Kolesnikov et al. (2022)split high-usage codebook vectors,Dhariwal et al. (2020); Łańcucki et al. (2020); Zheng & Vedaldi (2023)resurrect low-usage codebook vectors throughout training,Chen et al. (2024)dynamically selects one ofm𝑚mitalic_mcodebooks for each datapoint, andMentzer et al. (2023); Zhao et al. (2024); Yu et al. (2023); Chiu et al. (2022)fix the codebook vectors to ana priorigeometry and train the model without learning the codebook at all. Other works propose loss penalties to encourage codebook utilization.Zhang et al. (2023)add a KL-divergence penalty between codebook utilization and a uniform distribution whileYu et al. (2023)add an entropy loss term to penalize low codebook utilization. While effective at targeting specific training difficulties, as each of these methods continue to use the STE, the training instability caused by this estimator persist. Most of our experiments inSection5implement a subset of these approaches, and we find that replacing the STE with the rotation trick further improves performance.\n\n",
    "origin_citations": [
      "Neural discrete representation learning.",
      "vq-wav2vec: Self-supervised learning of discrete speech representations.",
      "Categorical reparameterization with gumbel-softmax.",
      "Soft convex quantization: Revisiting vector quantization with convex optimization.",
      "Sq-vae: Variational bayes on discrete representation with self-annealed stochastic quantization.",
      "Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks.",
      "Regularized vector quantization for tokenized image synthesis.",
      "High-resolution image synthesis with latent diffusion models.",
      "Maskgit: Masked generative image transformer.",
      "Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization.",
      "Designing a better asymmetric vqgan for stablediffusion.",
      "Peco: Perceptual codebook for bert pre-training of vision transformers.",
      "Vector-quantized image modeling with improved vqgan.",
      "Hypervq: Mlr-based vector quantization in hyperbolic space.",
      "Autoregressive image generation using residual quantization.",
      "Uvim: A unified modeling approach for vision with learned guiding codes.",
      "Jukebox: A generative model for music.",
      "Robust training of vector quantized bottleneck models.",
      "Online clustered codebook.",
      "Balance of number of embedding and their dimensions in vector quantization.",
      "Finite scalar quantization: Vq-vae made simple.",
      "Image and video tokenization with binary spherical quantization.",
      "Language model beats diffusion–tokenizer is key to visual generation.",
      "Self-supervised learning with random-projection quantizer for speech recognition."
    ],
    "citations": [
      3018,
      2550,
      782,
      323070,
      149013,
      892849,
      76396,
      2240,
      89438,
      260882,
      306167,
      2522,
      341953,
      427222,
      518938,
      147782,
      8727,
      60850,
      467819,
      975754,
      478224,
      739410,
      321651,
      90524
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 88,
      "Final_Score": 89.9,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 95\nSynthesis_Analysis: 88\nFinal_Score: 90"
    },
    "overall_score": 88.30000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "GMwRl2e9Y1",
        "forum": "GMwRl2e9Y1",
        "number": 11319,
        "title": "Restructuring Vector Quantization with the Rotation Trick",
        "normalized_title": "restructuring vector quantization with the rotation trick",
        "abstract": "Vector Quantized Variational AutoEncoders (VQ-VAEs) are designed to compress a continuous input to a discrete latent space and reconstruct it with minimal distortion. \nThey operate by maintaining a set of vectors---often referred to as the codebook---and quantizing each encoder output to the nearest vector in the codebook. \nHowever, as vector quantization is non-differentiable, the gradient to the encoder flows _around_ the vector quantization layer rather than _through_ it in a straight-through approximation.\nThis approximation may be undesirable as all information from the vector quantization operation is lost. \nIn this work, we propose a way to propagate gradients through the vector quantization layer of VQ-VAEs. \nWe smoothly transform each encoder output into its corresponding codebook vector via a rotation and rescaling linear transformation that is treated as a constant during backpropagation. \nAs a result, the relative magnitude and angle between encoder output and codebook vector becomes encoded into the gradient as it propagates through the vector quantization layer and back to the encoder.\nAcross 11 different VQ-VAE training paradigms, we find this restructuring improves reconstruction metrics, codebook utilization, and quantization error.",
        "authors": [
          "Christopher Fifty",
          "Ronald Guenther Junkins",
          "Dennis Duan",
          "Aniketh Iyengar",
          "Jerry Weihong Liu",
          "Ehsan Amid",
          "Sebastian Thrun",
          "Christopher Re"
        ],
        "decision": null,
        "pdate": 1737562771947,
        "cdate": 1727457688818
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "GMwRl2e9Y1",
          "forum": "GMwRl2e9Y1",
          "number": 11319,
          "title": "Restructuring Vector Quantization with the Rotation Trick",
          "normalized_title": "restructuring vector quantization with the rotation trick",
          "abstract": "Vector Quantized Variational AutoEncoders (VQ-VAEs) are designed to compress a continuous input to a discrete latent space and reconstruct it with minimal distortion. \nThey operate by maintaining a set of vectors---often referred to as the codebook---and quantizing each encoder output to the nearest vector in the codebook. \nHowever, as vector quantization is non-differentiable, the gradient to the encoder flows _around_ the vector quantization layer rather than _through_ it in a straight-through approximation.\nThis approximation may be undesirable as all information from the vector quantization operation is lost. \nIn this work, we propose a way to propagate gradients through the vector quantization layer of VQ-VAEs. \nWe smoothly transform each encoder output into its corresponding codebook vector via a rotation and rescaling linear transformation that is treated as a constant during backpropagation. \nAs a result, the relative magnitude and angle between encoder output and codebook vector becomes encoded into the gradient as it propagates through the vector quantization layer and back to the encoder.\nAcross 11 different VQ-VAE training paradigms, we find this restructuring improves reconstruction metrics, codebook utilization, and quantization error.",
          "authors": [
            "Christopher Fifty",
            "Ronald Guenther Junkins",
            "Dennis Duan",
            "Aniketh Iyengar",
            "Jerry Weihong Liu",
            "Ehsan Amid",
            "Sebastian Thrun",
            "Christopher Re"
          ],
          "decision": null,
          "pdate": 1737562771947,
          "cdate": 1727457688818
        }
      ]
    }
  },
  {
    "paper_id": 25465,
    "title": "Demystify Mamba in Vision: A Linear Attention Perspective",
    "abstract": "Mamba is an effective state space model with linear computation complexity. It has recently shown impressive efficiency in dealing with high-resolution inputs across various vision tasks.\nIn this paper, we reveal that the powerful Mamba model shares surprising similarities with linear attention Transformer, which typically underperform conventional Transformer in practice.\nBy exploring the similarities and disparities between the effective Mamba and subpar linear attention Transformer, we provide comprehensive analyses to demystify the key factors behind Mamba’s success. Specifically, we reformulate the selective state space model and linear attention within a unified formulation, rephrasing Mamba as a variant of linear attention Transformer with six major distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. For each design, we meticulously analyze its pros and cons, and empirically evaluate its impact on model performance in vision tasks. Interestingly, the results highlight the forget gate and block design as the core contributors to Mamba’s success, while the other four designs are less crucial.\nBased on these findings, we propose aMamba-Inspired Linear Attention (MILA)model by incorporating the merits of these two key designs into linear attention. The resulting model outperforms various vision Mamba models in both image classification and high-resolution dense prediction tasks, while enjoying parallelizable computation and fast inference speed. Code is available athttps://github.com/LeapLabTHU/MLLA.",
    "introduction": "Recently, state space models, exemplified by Mamba, have rapidly gained wide research interest. In contrast to the quadratic complexity of prevailing Transformer models, the state-space-based Mamba offers effective sequence modeling with linear complexity. This crucial property allows Mamba to handle extremely long sequences with manageable computational costs, making it a promising architecture for both natural language processing[14,29]and visual recognition[57,31].\n\n\nHowever, Mamba is not the first model to achieve global modeling with linear complexity. Linear attention[26], an early work, was proposed as an computationally efficient alternative to the widely adopted Softmax attention[42], namely dot-product attention. Specifically, linear attention replaces the non-linear Softmax function in attention operation with linear normalization. This enables a change in computation order from(Q⁢K⊤)⁢V𝑄superscript𝐾top𝑉(QK^{\\top})V( italic_Q italic_K start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT ) italic_VtoQ⁢(K⊤⁢V)𝑄superscript𝐾top𝑉Q(K^{\\top}V)italic_Q ( italic_K start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT italic_V ), thus reducing computation complexity from𝒪⁢(N2)𝒪superscript𝑁2\\mathcal{O}(N^{2})caligraphic_O ( italic_N start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT )to𝒪⁢(N)𝒪𝑁\\mathcal{O}(N)caligraphic_O ( italic_N ). Despite its efficiency, previous works[4,39,15,16]proved that linear attention suffers from insufficient expressive power, making it impractical for real applications. Surprisingly, we find a very close relationship between the formulas of high-performance Mamba and subpar linear attention Transformer. Therefore, a compelling research question emerges:What factors contribute to Mamba’s success and its significant superiority to linear attention Transformer?\n\n\nIn this paper, we offer both theoretical and empirical analyses to unveil Mamba through the lens of linear attention Transformer. Specifically, we rewrite the formulas of selective state space model and linear attention within a unified formulation, depicting Mamba as a variation of linear attention Transformer with six distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. To demystify what factors lead to Mamba’s effectiveness, empirical studies on vision tasks are conducted to assess the impact of each special design. The results demonstrate that the forget gate and block design tend to be the two core contributors to Mamba’s superiority. While the block design can be easily adopted, the forget gate necessitates recurrent computation, which may not be well-suited for non-auto-regressive vision models. Therefore, we delve into the essence of the forget gate and verify that it can be replaced by suitable positional encoding in vision tasks. Based on our findings, we introduce the two core contributors or their alternatives to linear attention Transformer, presenting ourMamba-Inspired Linear Attention (MILA)model. Experimental results demonstrate that MILA achieves superior results to various vision Mamba models in both image classification and high-resolution dense prediction tasks, validating that linear attention can surpass Mamba with the merits of two core designs.\n\n\nOur main contributions and takeaways are as follows:•We reveal Mamba’s close relationship to linear attention Transformer:Mamba and linear attention Transformer can be formulated within a unified framework, with Mamba exhibiting six distinct designs compared to the conventional linear attention paradigm: input gate, forget gate, shortcut, no attention normalization, single-head and modified block design.•We provide detailed analyses of each special design and empirically validate that theforget gate and block design largely lead to Mamba’s superiority.Additionally, we demonstrate that the recurrent calculation of the forget gate might not be ideal for vision models. Instead, proper positional encoding can function as the forget gate in vision tasks, while preserving parallelizable computation and fast inference speed.•We develop a series of linear attention vision Transformer models named MILA, which inherit the core merits of Mamba and tend to be more suitable for vision tasks than the original Mamba model.\n\n\nFigure 1:Illustration of selective SSM in Mamba (eq.11) and single head linear attention (eq.12). It can be seen that selective SSM resembles single-head linear attention with additional input gate𝚫isubscript𝚫𝑖\\mathbf{\\Delta}_{i}bold_Δ start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, forget gate𝑨~isubscript~𝑨𝑖\\widetilde{{\\bm{A}}}_{i}over~ start_ARG bold_italic_A end_ARG start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTand shortcut𝑫⊙𝒙idirect-product𝑫subscript𝒙𝑖{\\bm{D}}\\odot{\\bm{x}}_{i}bold_italic_D ⊙ bold_italic_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, while omitting normalization𝑸i⁢𝒁isubscript𝑸𝑖subscript𝒁𝑖{\\bm{Q}}_{i}{\\bm{Z}}_{i}bold_italic_Q start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT bold_italic_Z start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT.\n\n",
    "related_work": "Vision Transformer and attention.Originating from natural language processing, Transformer and attention have been highly successful in vision tasks[32,3,36,37], demonstrating superiority to the conventional CNN models[18,7,12]. However, the quadratic complexity of widely adopted Softmax attention[42]poses challenges in handling high-resolution images. Numerous works have attempted to reduce the computational cost by introducing local attention windows[32,9,20]or sparsity[43,45,56]. Linear attention[26], another approach, inherently offers linear complexity𝒪⁢(N)𝒪𝑁\\mathcal{O}(N)caligraphic_O ( italic_N )and is capable of modeling long sequences. Despite its efficiency, previous works[4,39,47,17]have shown that linear attention always fails to deliver satisfactory results, limiting its applicability.\n\n\nMamba[14]is a recently proposed state space model that achieves effective sequence modeling with linear complexity. Motivated by its potential for modeling high-resolution images, many researchers try to apply Mamba to vision tasks[14,31,27,25,38,49,48,23]. For instance, VMamba[31]introduces a cross-scan module to enable 1D selective scanning in 2D image space. LocalMamba[25]utilizes local windows to enhance local modeling capability. EfficientVMamba[38]designs an atrous-based selective scan approach to enhance efficiency. In addition, MambaOut[51]analyzes whether Mamba is needed for vision, and explainability methods[1]have also been proposed.\n\n\nIn contrast to incorporating Mamba into vision, this paper reveals the surprising similarities between the formulas of inferior linear attention Transformer and powerful Mamba model. This interesting finding gives us the opportunity to demystify the key factors behind Mamba’s success.\n\n",
    "origin_citations": [
      "Swin transformer: Hierarchical vision transformer using shifted windows.",
      "End-to-end object detection with transformers.",
      "Deep incubation: Training large models by divide-and-conquering.",
      "Enat: Rethinking spatial-temporal interactions in token-based image synthesis.",
      "On the connection between local attention and dynamic depth-wise convolution.",
      "Demystify transformers & convolutions in modern image deep networks.",
      "What makes for hierarchical vision transformer?",
      "Attention is all you need.",
      "Cswin transformer: A general vision transformer backbone with cross-shaped windows.",
      "Neighborhood attention transformer.",
      "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions.",
      "Vision transformer with deformable attention.",
      "Biformer: Vision transformer with bi-level routing attention.",
      "Transformers are rnns: Fast autoregressive transformers with linear attention.",
      "Rethinking attention with performers.",
      "cosformer: Rethinking softmax in attention.",
      "Nyströmformer: A nyström-based algorithm for approximating self-attention.",
      "Agent attention: On the integration of softmax and linear attention.",
      "Mamba: Linear-time sequence modeling with selective state spaces.",
      "Vmamba: Visual state space model.",
      "Mamba-nd: Selective state space modeling for multi-dimensional data.",
      "Localmamba: Visual state space model with windowed selective scan.",
      "Efficientvmamba: Atrous selective scan for light weight visual mamba.",
      "Plainmamba: Improving non-hierarchical mamba in visual recognition.",
      "Diffusion models without attention.",
      "Zigma: A dit-style zigzag mamba diffusion model.",
      "Mambaout: Do we really need mamba for vision?",
      "The hidden attention of mamba models."
    ],
    "citations": [
      1501,
      7360,
      222277,
      1087344,
      194846,
      32865,
      372596,
      38,
      4780,
      63187,
      2577,
      256286,
      704886,
      798,
      8384,
      175109,
      119364,
      270362,
      274507,
      126226,
      98956,
      287742,
      993243,
      423937,
      425938,
      1046552,
      291750
    ],
    "extract_rate": 0.9642857142857143,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 88,
      "Final_Score": 88.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 88\nFinal_Score: 89"
    },
    "overall_score": 87.82,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "LvJ1R88KAk",
        "forum": "LvJ1R88KAk",
        "number": 4979,
        "title": "Demystify Mamba in Vision: A Linear Attention Perspective",
        "normalized_title": "demystify mamba in vision a linear attention perspective",
        "abstract": "Mamba is an effective state space model with linear computation complexity. It has recently shown impressive efficiency in dealing with high-resolution inputs across various vision tasks. In this paper, we reveal that the powerful Mamba model shares surprising similarities with linear attention Transformer, which typically underperform conventional Transformer in practice. By exploring the similarities and disparities between the effective Mamba and subpar linear attention Transformer, we provide comprehensive analyses to demystify the key factors behind Mamba’s success. Specifically, we reformulate the selective state space model and linear attention within a unified formulation, rephrasing Mamba as a variant of linear attention Transformer with six major distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. For each design, we meticulously analyze its pros and cons, and empirically evaluate its impact on model performance in vision tasks. Interestingly, the results highlight the forget gate and block design as the core contributors to Mamba’s success, while the other four designs are less crucial. Based on these findings, we propose a Mamba-\nInspired Linear Attention (MILA) model by incorporating the merits of these two key designs into linear attention. The resulting model outperforms various vision Mamba models in both image classification and high-resolution dense prediction tasks, while enjoying parallelizable computation and fast inference speed. Code is available at https://github.com/LeapLabTHU/MLLA.",
        "authors": [
          "Dongchen Han",
          "Ziyi Wang",
          "Zhuofan Xia",
          "Yizeng Han",
          "Yifan Pu",
          "Chunjiang Ge",
          "Jun Song",
          "Shiji Song",
          "Bo Zheng",
          "Gao Huang"
        ],
        "decision": null,
        "pdate": 1727287766591,
        "cdate": 1715491786004
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "LvJ1R88KAk",
          "forum": "LvJ1R88KAk",
          "number": 4979,
          "title": "Demystify Mamba in Vision: A Linear Attention Perspective",
          "normalized_title": "demystify mamba in vision a linear attention perspective",
          "abstract": "Mamba is an effective state space model with linear computation complexity. It has recently shown impressive efficiency in dealing with high-resolution inputs across various vision tasks. In this paper, we reveal that the powerful Mamba model shares surprising similarities with linear attention Transformer, which typically underperform conventional Transformer in practice. By exploring the similarities and disparities between the effective Mamba and subpar linear attention Transformer, we provide comprehensive analyses to demystify the key factors behind Mamba’s success. Specifically, we reformulate the selective state space model and linear attention within a unified formulation, rephrasing Mamba as a variant of linear attention Transformer with six major distinctions: input gate, forget gate, shortcut, no attention normalization, single-head, and modified block design. For each design, we meticulously analyze its pros and cons, and empirically evaluate its impact on model performance in vision tasks. Interestingly, the results highlight the forget gate and block design as the core contributors to Mamba’s success, while the other four designs are less crucial. Based on these findings, we propose a Mamba-\nInspired Linear Attention (MILA) model by incorporating the merits of these two key designs into linear attention. The resulting model outperforms various vision Mamba models in both image classification and high-resolution dense prediction tasks, while enjoying parallelizable computation and fast inference speed. Code is available at https://github.com/LeapLabTHU/MLLA.",
          "authors": [
            "Dongchen Han",
            "Ziyi Wang",
            "Zhuofan Xia",
            "Yizeng Han",
            "Yifan Pu",
            "Chunjiang Ge",
            "Jun Song",
            "Shiji Song",
            "Bo Zheng",
            "Gao Huang"
          ],
          "decision": null,
          "pdate": 1727287766591,
          "cdate": 1715491786004
        }
      ]
    }
  },
  {
    "paper_id": 27923,
    "title": "Adapting Pretrained ViTs with Convolution Injector for Visuo-Motor Control",
    "abstract": "Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias.\nHowever, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases.\nSuch absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduceConvolutionInjector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases.111Project page:https://godnpeter.github.io/CoIn222Code:https://github.com/dojeon-ai/CoIn",
    "introduction": "Developing intelligent robotic agents capable of precise visuo-motor control is an important area of research.\nA standard paradigm of developing such agents is to train the visual encoder and control policy end-to-end, using domain-specific control data(Levine et al.,2016).\nHowever, this approach limits the applicability of visuo-motor control policies in real-world scenarios due to the excessive amount of data required for learning and the lack of flexibility in adapting to new situations, such as unseen environments.\n\n\nIn the fields of computer vision and natural language processing, a large number of studies have shown that pretraining high-capacity models on large datasets demonstrate superior data efficiency and generalization capabilities compared to approaches trained from scratch(Dosovitskiy et al.,2021; Bommasani et al.,2021; Devlin et al.,2018; Brown et al.,2020).\nIn response, for visuo-motor control, there has been a growing interest in utilizing large visual encoders pretrained on extensive and diverse datasets(Radosavovic et al.,2022; Hansen et al.,2021; Majumdar et al.,2023).\n\n\nFigure 1:Avg. performance across 12 visuo-motor control tasks.Our model CoIn introduces convolutional inductive biases into ViTs, resulting in consistent performance improvements for various pretrained ViTs.\n\n\nFigure 2:Overall framework.(Stage 1) The advent of open-sourced, large-scale ViTs pretrained with extensive web-scale datasets provides generalized, ready-to-go visual representations. (Stage 2) To adapt these pretrained ViTs for visuo-motor control, we finetune them with an additional light-weight module, CoIn, enhancing the ViT’s ability to extract visual features beneficial for control, such as spatial locality and translation equivariance.\n\n\nFor visuo-motor control, Vision Transformers (ViT)(Dosovitskiy et al.,2021)emerges as an appealing choice as it achieved remarkable success in a wide range of computer vision tasks such as image classification(Dosovitskiy et al.,2021; Bao et al.,2022), object detection(Liu et al.,2021; Li et al.,2022)and semantic segmentation(Strudel et al.,2021; Kirillov et al.,2023).\nThe success of ViTs is attributed to their weak inductive bias, which significantly enhances model performance when scaled with a large pretraining dataset and model size(Naseer et al.,2021; Yu et al.,2021; Mao et al.,2022; Chu et al.,2021; Dehghani et al.,2023).\n\n\nNonetheless, although the weak inductive bias of ViTs is advantageous for scaling during the pretraining phase, this characteristic may hinder their effective adaptation for visuo-motor control.\nFor effective visuo-motor control, a visual encoder must (i) focus on the interaction area of interest, and (ii) track object and gripper positions with respect to their changes in locations.\nViTs inherently lack such properties due to their design.\nSuch limitations of ViTs can be addressed by incorporating two specific biases that are naturally present in convolutional layers: (i) a bias towards spatial locality, and (ii) a bias for translation equivariance.\n\n\nTo this end, we introduceConvolutionInjector (CoIn), a module designed to inject spatial locality and translation equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control.\nCoIn is a simple and lightweight add-on module (3.6% of additional parameters to a standard ViT-B/16) designed to exploit the strengths of pretrained ViTs while providing advantageous inductive biases essential for visual control tasks.\nSpecifically, CoIn extracts locality and translation equivariance-aware features through convolutional layers and integrates them into the ViT architecture using a cross-attention mechanism (see Figure3).\nThis integration enables the pretrained ViT to effectively leverage both its pretrained knowledge and newly obtained spatial prior features during adaptation for downstream control tasks.\nTherefore, CoIn eliminates the need to retrain pretrained ViTs from scratch with datasets and objectives specifically tailored for visual control applications.\n\n\nTo thoroughly evaluate the effectiveness of CoIn, we conduct extensive evaluations across 12 different visuo-motor control tasks within 3 distinct domains: Adroit(Rajeswaran et al.,2018), MetaWorld(Yu et al.,2020), and DMC(Tassa et al.,2018)for three different pretrained ViT visual encoders: CLIP(Radford et al.,2021), MVP(Radosavovic et al.,2022), and VC-1(Majumdar et al.,2023).\nOur results demonstrate that CoIn consistently enhances downstream control task performance across all environments and with all pretrained ViTs. Notably, when paired with CLIP, finetuning with CoIn achieved a substantial 11.3 point increase in mean success over finetuning the baseline CLIP model.\nThese findings suggest that the incorporation of locality and translation-equivariance-aware features plays a crucial role in enhancing the capabilities of ViTs for visuo-motor control tasks.\n\n\nIn summary, although ViTs gain advantages from large-scale pretraining due to their weak inductive bias, this same characteristic limits their adaptability for visuo-motor control tasks because of the absence of specific control-centric biases. Consequently, we introduce CoIn, a module which incorporates beneficial control-centric inductive biases, readily provided by convolutional layers, into large-scale pretrained ViTs. Our code is available athttps://godnpeter.github.io/CoIn.\n\n",
    "related_work": "2.1Pretrained Visual Encoders for ControlRecently, pretraining effective visual encoders for control by leveraging large, diverse datasets from the internet has gain much interest from the research community(Parisi et al.,2022; Nair et al.,2022; Radosavovic et al.,2022; Majumdar et al.,2023; Wang et al.,2022; Yuan et al.,2022; Shah & Kumar,2021). Specifically, PVR(Parisi et al.,2022)is among the initial investigations into the use of large pretrained visual encoders for control. It demonstrates that while doing behavior cloning, ResNet encoders(He et al.,2016)trained via self-supervised contrastive learning(He et al.,2020)can match the performance of state-based inputs.\nFurther advancements are seen in R3M(Nair et al.,2022), which employs a temporal contrastive objective to learn representations for robotic control and VIP(Ma et al.,2022), which focuses on learning visual representations which reflect the distance between states and goals. Similarly, MVP(Radosavovic et al.,2022)and VC-1(Majumdar et al.,2023)demonstrate the efficacy of ViTs pretrained with MAE(He et al.,2022)on extensive internet video and image data for robotic manipulations. As an alternative attempt, MOO(Stone et al.,2023)and RT-2(Brohan et al.,2023)investigate the application of vision-language models pretrained on broad internet data, for improved robotic control and emergent reasoning. Unlike previous research which mainly focus on the performance offrozenweights in different control tasks, our work delves into the effectiveness and challenges offinetuningViTs for control tasks, particularly in the context of imitation learning.\n\n\n2.2Integration of CNNs with Pretrained ViTs in Computer VisionThe integration of CNNs with pretrained ViTs to leverage their collective capabilities for various computer vision tasks has recently been investigated by the research community(Peng et al.,2021; Fang et al.,2023; Chen et al.,2022b; Ranftl et al.,2021; Hong et al.,2022).\nVitMatte(Yao et al.,2024)demonstrates the effectiveness of combining lightweight CNNs with a pretrained ViT for enhanced image matting.\nFor dense prediction, DPT(Ranftl et al.,2021)introduces a randomly initialized CNN decoder, and ViT-Adapter(Chen et al.,2022b)utilizes a CNN-based adapter which embeds local semantic features into pretrained ViTs. MIMDET(Fang et al.,2023)employs a compact CNN encoder before the patch embedding layer of ViT, creating a CNN-ViT hybrid feature extractor for object detection.Such hybrid models are particularly well-suited for visuo-motor control applications.\nThis approach naturally integrates spatial locality and translation equivariance biases which lack in ViTs, but are essential for visuo-motor control.\nThis eliminates the need for retraining with visuo-motor specific datasets for obtaining ViTs tailored towards visuo-motor control.\nHowever, despite active exploration in computer vision, the application of such convolutional bias integrated pretrained ViTs for visuo-motor control has been relatively unexplored (often relying on either standard ResNet or ViT models as previously mentioned in section2.1). Our research aims to address this gap, exploring how pretrained ViTs can be effectively adapted for control tasks while fully leveraging their well-generalized features.\n\n",
    "origin_citations": [
      "The unsurprising effectiveness of pre-trained vision models for control.",
      "R3m: A universal visual representation for robot manipulation.",
      "Real-world robot learning with masked visual pre-training.",
      "Where are we in the search for an artificial visual cortex for embodied intelligence?",
      "Vrl3: A data-driven framework for visual deep reinforcement learning.",
      "Pre-trained image encoder for generalizable visual reinforcement learning.",
      "Rrl: Resnet as representation for reinforcement learning.",
      "Deep residual learning for image recognition.",
      "Momentum contrast for unsupervised visual representation learning.",
      "Vip: Towards universal visual reward and representation via value-implicit pre-training.",
      "Masked autoencoders are scalable vision learners.",
      "Open-world object manipulation using pre-trained vision-language models.",
      "Rt-2: Vision-language-action models transfer web knowledge to robotic control.",
      "Conformer: Local features coupling global representations for visual recognition.",
      "Unleashing vanilla vision transformer with masked image modeling for object detection.",
      "Vision transformer adapter for dense predictions.",
      "Vision transformers for dense prediction.",
      "Representation separation for semantic segmentation with vision transformers.",
      "Vitmatte: Boosting image matting with pre-trained plain vision transformers."
    ],
    "citations": [
      2602,
      2603,
      460652,
      847764,
      175211,
      140396,
      7593,
      97,
      122,
      739554,
      2513,
      78991,
      7672,
      250176,
      65021,
      148626,
      507333,
      138496,
      258778
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 88,
      "Final_Score": 87.3,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 88\nFinal_Score: 87"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 88,
      "Final_Score": 88.1,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 92\nSynthesis_Analysis: 88\nFinal_Score: 88"
    },
    "overall_score": 87.78,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "CuiRGtVI55",
        "forum": "CuiRGtVI55",
        "number": 9250,
        "title": "Adapting Pretrained ViTs with Convolution Injector for Visuo-Motor Control",
        "normalized_title": "adapting pretrained vits with convolution injector for visuo motor control",
        "abstract": "Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias. However, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases. Such absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduce Convolution Injector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases.",
        "authors": [
          "Dongyoon Hwang",
          "Byungkun Lee",
          "Hojoon Lee",
          "Hyunseung Kim",
          "Jaegul Choo"
        ],
        "decision": null,
        "pdate": 1714610489085,
        "cdate": 1706870403830
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "CuiRGtVI55",
          "forum": "CuiRGtVI55",
          "number": 9250,
          "title": "Adapting Pretrained ViTs with Convolution Injector for Visuo-Motor Control",
          "normalized_title": "adapting pretrained vits with convolution injector for visuo motor control",
          "abstract": "Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias. However, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases. Such absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduce Convolution Injector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases.",
          "authors": [
            "Dongyoon Hwang",
            "Byungkun Lee",
            "Hojoon Lee",
            "Hyunseung Kim",
            "Jaegul Choo"
          ],
          "decision": null,
          "pdate": 1714610489085,
          "cdate": 1706870403830
        }
      ]
    }
  },
  {
    "paper_id": 15626,
    "title": "How Smooth Is Attention?",
    "abstract": "Self-attention and masked self-attention are at the heart of Transformers’ outstanding success.\nStill, our mathematical understanding of attention, in particular of its Lipschitz properties – which are key when it comes to analyzing robustness and expressive power – is incomplete.\nWe provide a detailed study of the Lipschitz constant of self-attention in several practical scenarios, discussing the impact of the sequence lengthn𝑛nitalic_nand layer normalization on the local Lipschitz constant of both unmasked and masked self-attention.\nIn particular, we show that for inputs of lengthn𝑛nitalic_nin any compact set, the Lipschitz constant of self-attention is bounded byn𝑛\\sqrt{n}square-root start_ARG italic_n end_ARGup to a constant factor and that this bound is tight for reasonable sequence lengths.\nWhen the sequence lengthn𝑛nitalic_nis too large for the previous bound to be tight, which we refer to as the mean-field regime, we provide an upper bound and a matching lower bound which are independent ofn𝑛nitalic_n.\nOur mean-field framework for masked self-attention is novel and of independent interest.\nOur experiments on pretrained and randomly initialized BERT and GPT-2 support our theoretical findings.",
    "introduction": "Introduced byVaswani et al. (2017), Transformers and their multi-head attention mechanism(Bahdanau et al.,2014)have significantly changed the machine learning landscape in just a few years, by becoming state-of-art models on a wide variety of tasks, from natural language processing(Brown et al.,2020; Radford et al.,2019; Wolf et al.,2019)to computer vision(Dosovitskiy et al.,2020; Zhao et al.,2020; Zhai et al.,2022; Lee et al.,2019).\nDespite this great empirical success, however, little is known from a theoretical perspective about the smoothness of Transformer architectures, particularly of self-attention, their main building block.\nWe tackle this problem by focusing on the Lipschitz properties of self-attention, especially on its local Lipschitz constant, which controls how fast the output can change with respect to the input in the neighborhood of each point of the domain.\n\n\nStudying the Lipschitz continuity of neural networks is particularly relevant for various questions(Rosca et al.,2020).\nIt provides guarantees of adversarial robustness, in an attack-agnostic way(Szegedy et al.,2013; Cisse et al.,2017; Tsuzuku et al.,2018; Anil et al.,2019; Weng et al.,2018).\nIdentifying inputs with a high local Lipschitz constant and understanding which local perturbations trigger the biggest change in the output also allows for robustifying the network, for example using adversarial training(Goodfellow et al.,2014; Miyato et al.,2015; Moosavi-Dezfooli et al.,2016; Kurakin et al.,2016).\nThe Lipschitz constant is also involved in generalization bounds(Sokolić et al.,2017; Neyshabur et al.,2017; Bartlett et al.,2017; von Luxburg & Bousquet,2004).\nFrom a different perspective, Lipschitz-constrained neural networks can be used to estimate Wasserstein distances(Peyré et al.,2019), enhance expressivity and improve the performance of deep models(Miyato et al.,2018; Dasoulas et al.,2021), and build invertible neural networks(Behrmann et al.,2019; Chen et al.,2019).\nFinally, bounding the Lipschitz constant of a neural network is an important step in the study of the associated neural ODE(Chen et al.,2018), in particular of its well-posedness(Lu et al.,2019; Geshkovski et al.,2023a,b).\n\n\nLipschitz continuity of feed-forward neural networks (FFNs) has been extensively studied and remains a hard problem.\nEstimating numerically the Lipschitz constant of a FFN is indeed NP-hard(Virmaux & Scaman,2018), and theoretical bounds appear to be much larger than the actual Lipschitz constant(Virmaux & Scaman,2018; Fazlyab et al.,2019; Latorre et al.,2020).\nThe main theoretical difficulty here is to handle the composition of several layers more accurately than just bounding it by the product of spectral norms of weight matrices, as done bySzegedy et al. (2013).\nStill, taken independently, each linear map or activation function has a known tight Lipschitz constant.\nThis isnotthe case for Transformers: the self-attention map has an involved non-linear structure, which makes the estimation of its local Lipschitz constant challenging and brings into play completely different approaches than for FFNs(Kim et al.,2021; Vuckovic et al.,2021).\n\n\n1.1ContributionsWe make the following contributions.•We derive a bound on the local Lipschitz constant of self-attention, which takes the formC⁢n𝐶𝑛C\\sqrt{n}italic_C square-root start_ARG italic_n end_ARGwithn𝑛nitalic_nthe sequence length of inputs andC𝐶Citalic_Ca constant factor that depends on the parameters of self-attention and on an upper boundR𝑅Ritalic_Ron the magnitude of tokens (Theorem 3.3).\nWe show that our bound is tight inn𝑛nitalic_nfor reasonable (i.e. not too large) sequence lengths (Proposition3.4).\nMoreover, in most Transformer architectures, the magnitudeR𝑅Ritalic_Ronly depends on the parameters of the network because of normalization layers (Subsection2.3).•We identify alarge radius regimethat is easier to analyze theoretically, withn𝑛nitalic_nfixed andR𝑅Ritalic_Rvery large.\nIn this regime and except for a measure-zero set of pathological configurations, we show that the local Lipschitz constant of self-attention is bounded byC⁢n𝐶𝑛C\\sqrt{n}italic_C square-root start_ARG italic_n end_ARGwithC𝐶Citalic_Ca constant that does not depend onR𝑅Ritalic_Ranymore (Theorem 3.7).•We also study the mean-field regime, where self-attention is modeled as a map on probability measures, which corresponds to the limitn→+∞→𝑛n\\to+\\inftyitalic_n → + ∞.\nIn this framework, we show that the upper bound obtained byGeshkovski et al. (2023a), which is of the formC⁢R2⁢eC⁢R2,𝐶superscript𝑅2superscript𝑒𝐶superscript𝑅2CR^{2}e^{CR^{2}},italic_C italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT italic_e start_POSTSUPERSCRIPT italic_C italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ,cannot be significantly improved (Proposition3.6), by finding aR𝑅Ritalic_R-indexed family of two-Dirac probability measures supported in the closed ballBRsubscript𝐵𝑅B_{R}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPTof center 0 and of radiusR𝑅Ritalic_Rwhose local Lipschitz constant grows likeC′2⁢R2⁢eC′⁢R2superscript𝐶′2superscript𝑅2superscript𝑒superscript𝐶′superscript𝑅2\\frac{C^{\\prime}}{2}R^{2}e^{C^{\\prime}R^{2}}divide start_ARG italic_C start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT end_ARG start_ARG 2 end_ARG italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT italic_e start_POSTSUPERSCRIPT italic_C start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_POSTSUPERSCRIPTwithC′≥C/16superscript𝐶′𝐶16C^{\\prime}\\geq C/16italic_C start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ≥ italic_C / 16.•We are the first to study the Lipschitz constant ofmaskedself-attention.\nWe introduce a novel mean-field framework for masked self-attention, where the order of points in input measures is encoded in a supplementary coordinate, and show both in the general regime, the large radius regime and the mean-field regime that similar bounds hold for masked self-attention as for unmasked self-attention (Section 4).•We compute numerically the local Lipschitz constant of unmasked and masked self-attention in a BERT model and a GPT-2 model, where inputs are text extracts, and observe a growth rate ofn1/4superscript𝑛14n^{1/4}italic_n start_POSTSUPERSCRIPT 1 / 4 end_POSTSUPERSCRIPTup to a constant factor, withn𝑛nitalic_nthe sequence length.\nThen, with the same networks, we build adversarial data in the input space of self-attention whose Lipschitz constant grows liken𝑛\\sqrt{n}square-root start_ARG italic_n end_ARG, which evidentiates the tightness of our bounds (Section5).\n\n\n1.2Related WorkRobustness and local Lipschitz constant estimation.Neural networks are vulnerable to adversarial attacks(Szegedy et al.,2013), and most of the methods proposed to measure and increase their robustness focus on specific attacks(Goodfellow et al.,2014; Papernot et al.,2016).\nIt turns out, however, that such methods can be defeated by well-chosen unseen attacks(Carlini & Wagner,2017).\nMeasures of robustness that are agnostic to attack methods have therefore been proposed, often relying on the notion of Lipschitz constant of networks(Szegedy et al.,2013; Leino et al.,2021; Tsuzuku et al.,2018).\nAs robustness lower bounds that rely on the (global) Lipschitz constant tend to be too loose, tighter constraints have been proposed involving the local Lipschitz constant(Hein & Andriushchenko,2017; Weng et al.,2018).\nThe problem of evaluating the local Lipschitz constant of deep networks is now at the heart of several recent articles(Tsuzuku et al.,2018; Leino et al.,2021), in particular for Transformers(Kim et al.,2021; Vuckovic et al.,2020; Geshkovski et al.,2023a; Catellier et al.,2023).\nFrom a more practical viewpoint, several Lipschitz-constrained variants of the Transformer architecture have been proposed, to increase robustness and reliability(Jia et al.,2023; Gupta & Verma,2023; Ye et al.,2023; Qi et al.,2023).Neural networks acting on measures.De Bie et al. (2019)andPevny & Kovarik (2019)are the first to define neural networks whose inputs are probability measures, followed by several other articles(Vuckovic et al.,2020; Zweig & Bruna,2021; Sander et al.,2022; Geshkovski et al.,2023a).\nModeling neural networks as maps on probability measures has multiple applications, such as studying Wasserstein regularity(Vuckovic et al.,2020; Geshkovski et al.,2023a), proving generalization bounds(Zweig & Bruna,2021)and doing a mean-field limit analysis of the dynamics of particles as they go through the network(Geshkovski et al.,2023a).\nThe mean-field approach is particularly suited to the case of Encoder-only Transformers(Devlin et al.,2018), as the self-attention map is permutation equivariant, i.e., ignores the order of vectors in its input.\nThis property can be leveraged to model any infinitely deep Encoder as a partial differential equation (PDE) on the space of measures(Sander et al.,2022), following the principle of neural ODEs(Chen et al.,2018).\nAnalyzing this PDE then provides information about the dynamics of tokens as they go through the Transformer, showing for instance the emergence of clusters(Geshkovski et al.,2023a,b).\nIn contrast, masked self-attention, which is crucial in Decoder-only(Liu et al.,2018)and Encoder-Decoder(Vaswani et al.,2017)architectures, is not permutation equivariant, so cannot be cast as naturally into a mean-field framework.Regularity of self-attention and its variants.Kim et al. (2021)show that the self-attention map is not globally Lipschitz continuous by deriving a lower bound on its Lipschitz constant restricted toBRnsuperscriptsubscript𝐵𝑅𝑛B_{R}^{n}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT.\nTheir lower bound grows quadratically withR𝑅Ritalic_R.\nTo gain regularity, they define a new self-attention map, called L2 self-attention, which is globally Lipschitz continuous on the set of inputs of lengthn𝑛nitalic_n, for alln≥1𝑛1n\\geq 1italic_n ≥ 1.Dasoulas et al. (2021)enforce the Lipschitz continuity of self-attention modules by normalizing the attention scores with a well-chosen normalization function.Geshkovski et al. (2023a)andVuckovic et al. (2020)prove a mean-field upper bound on the Lipschitz constant of self-attention onBRsubscript𝐵𝑅B_{R}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPT, by viewing self-attention as a map acting on probability measures.\nTheir upper bound grows more than exponentially withR𝑅Ritalic_Rso that the quadratic lower bound and the exponential upper bound put together provide a very loose estimation of the Lipschitz constant of self-attention on compact sets.\nFinally,Sander et al. (2022)propose a modification of the attention kernel that builds on the Sinkhorn-Knopp algorithm, and provide empirical evidence of the better properties of this new choice of kernel with respect to the classical one.\n\n\n1.3NotationsThe Euclidean norm onℝdsuperscriptℝ𝑑\\mathbb{R}^{d}blackboard_R start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPTis denoted|⋅|⋅\\lvert\\,\\cdot\\,\\rvert| ⋅ |.\nFor any vectorw∈ℝn𝑤superscriptℝ𝑛w\\in\\mathbb{R}^{n}italic_w ∈ blackboard_R start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT, we denotesoftmax⁡(w)≔(exp⁡(wi)/∑j=1dexp⁡(wj))1≤i≤n≔softmax𝑤subscriptsubscript𝑤𝑖superscriptsubscript𝑗1𝑑subscript𝑤𝑗1𝑖𝑛\\operatorname{\\mathrm{softmax}}(w)\\coloneqq\\left(\\exp(w_{i})/{\\sum_{j=1}^{d}%\n\\exp(w_{j})}\\right)_{1\\leq i\\leq n}roman_softmax ( italic_w ) ≔ ( roman_exp ( italic_w start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) / ∑ start_POSTSUBSCRIPT italic_j = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPT roman_exp ( italic_w start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT ) ) start_POSTSUBSCRIPT 1 ≤ italic_i ≤ italic_n end_POSTSUBSCRIPTthe Softmax operator, anddiag⁢(w)diag𝑤\\mathrm{diag}(w)roman_diag ( italic_w )the diagonal matrix such thatdiag⁢(w)i⁢i=widiagsubscript𝑤𝑖𝑖subscript𝑤𝑖\\mathrm{diag}(w)_{ii}=w_{i}roman_diag ( italic_w ) start_POSTSUBSCRIPT italic_i italic_i end_POSTSUBSCRIPT = italic_w start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT.\nFor any functiong:ℰ→ℱ:𝑔→ℰℱg\\colon\\mathcal{E}\\to\\mathcal{F}italic_g : caligraphic_E → caligraphic_Fand any subset𝒳⊂ℰ𝒳ℰ\\mathcal{X}\\subset\\mathcal{E}caligraphic_X ⊂ caligraphic_E, the restriction ofg𝑔gitalic_gto𝒳𝒳\\mathcal{X}caligraphic_Xis denotedg|𝒳g_{\\lvert\\mathcal{X}}italic_g start_POSTSUBSCRIPT | caligraphic_X end_POSTSUBSCRIPT.\nThe closed ball centered at 0 and of radiusR>0𝑅0R>0italic_R > 0is denotedBRsubscript𝐵𝑅B_{R}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPT.\nForφ,ψ:ℝ→ℝ:𝜑𝜓→ℝℝ\\varphi,\\psi\\colon\\mathbb{R}\\to\\mathbb{R}italic_φ , italic_ψ : blackboard_R → blackboard_Randa∈ℝ∪{+∞}𝑎ℝa\\in\\mathbb{R}\\cup\\{+\\infty\\}italic_a ∈ blackboard_R ∪ { + ∞ }we writeφ⁢(x)∼x→aψ⁢(x)subscriptsimilar-to→𝑥𝑎𝜑𝑥𝜓𝑥\\varphi(x)\\sim_{x\\to a}\\psi(x)italic_φ ( italic_x ) ∼ start_POSTSUBSCRIPT italic_x → italic_a end_POSTSUBSCRIPT italic_ψ ( italic_x )ifφ⁢(x)/ψ⁢(x)𝜑𝑥𝜓𝑥\\varphi(x)/\\psi(x)italic_φ ( italic_x ) / italic_ψ ( italic_x )is well-defined forx𝑥xitalic_xclose enough toa𝑎aitalic_a, andφ⁢(x)/ψ⁢(x)→x→a1subscript→→𝑥𝑎𝜑𝑥𝜓𝑥1\\varphi(x)/\\psi(x)\\to_{x\\to a}1italic_φ ( italic_x ) / italic_ψ ( italic_x ) → start_POSTSUBSCRIPT italic_x → italic_a end_POSTSUBSCRIPT 1.\n\n",
    "related_work": "Robustness and local Lipschitz constant estimation.Neural networks are vulnerable to adversarial attacks(Szegedy et al.,2013), and most of the methods proposed to measure and increase their robustness focus on specific attacks(Goodfellow et al.,2014; Papernot et al.,2016).\nIt turns out, however, that such methods can be defeated by well-chosen unseen attacks(Carlini & Wagner,2017).\nMeasures of robustness that are agnostic to attack methods have therefore been proposed, often relying on the notion of Lipschitz constant of networks(Szegedy et al.,2013; Leino et al.,2021; Tsuzuku et al.,2018).\nAs robustness lower bounds that rely on the (global) Lipschitz constant tend to be too loose, tighter constraints have been proposed involving the local Lipschitz constant(Hein & Andriushchenko,2017; Weng et al.,2018).\nThe problem of evaluating the local Lipschitz constant of deep networks is now at the heart of several recent articles(Tsuzuku et al.,2018; Leino et al.,2021), in particular for Transformers(Kim et al.,2021; Vuckovic et al.,2020; Geshkovski et al.,2023a; Catellier et al.,2023).\nFrom a more practical viewpoint, several Lipschitz-constrained variants of the Transformer architecture have been proposed, to increase robustness and reliability(Jia et al.,2023; Gupta & Verma,2023; Ye et al.,2023; Qi et al.,2023).\n\n\nNeural networks acting on measures.De Bie et al. (2019)andPevny & Kovarik (2019)are the first to define neural networks whose inputs are probability measures, followed by several other articles(Vuckovic et al.,2020; Zweig & Bruna,2021; Sander et al.,2022; Geshkovski et al.,2023a).\nModeling neural networks as maps on probability measures has multiple applications, such as studying Wasserstein regularity(Vuckovic et al.,2020; Geshkovski et al.,2023a), proving generalization bounds(Zweig & Bruna,2021)and doing a mean-field limit analysis of the dynamics of particles as they go through the network(Geshkovski et al.,2023a).\nThe mean-field approach is particularly suited to the case of Encoder-only Transformers(Devlin et al.,2018), as the self-attention map is permutation equivariant, i.e., ignores the order of vectors in its input.\nThis property can be leveraged to model any infinitely deep Encoder as a partial differential equation (PDE) on the space of measures(Sander et al.,2022), following the principle of neural ODEs(Chen et al.,2018).\nAnalyzing this PDE then provides information about the dynamics of tokens as they go through the Transformer, showing for instance the emergence of clusters(Geshkovski et al.,2023a,b).\nIn contrast, masked self-attention, which is crucial in Decoder-only(Liu et al.,2018)and Encoder-Decoder(Vaswani et al.,2017)architectures, is not permutation equivariant, so cannot be cast as naturally into a mean-field framework.\n\n\nRegularity of self-attention and its variants.Kim et al. (2021)show that the self-attention map is not globally Lipschitz continuous by deriving a lower bound on its Lipschitz constant restricted toBRnsuperscriptsubscript𝐵𝑅𝑛B_{R}^{n}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT.\nTheir lower bound grows quadratically withR𝑅Ritalic_R.\nTo gain regularity, they define a new self-attention map, called L2 self-attention, which is globally Lipschitz continuous on the set of inputs of lengthn𝑛nitalic_n, for alln≥1𝑛1n\\geq 1italic_n ≥ 1.Dasoulas et al. (2021)enforce the Lipschitz continuity of self-attention modules by normalizing the attention scores with a well-chosen normalization function.Geshkovski et al. (2023a)andVuckovic et al. (2020)prove a mean-field upper bound on the Lipschitz constant of self-attention onBRsubscript𝐵𝑅B_{R}italic_B start_POSTSUBSCRIPT italic_R end_POSTSUBSCRIPT, by viewing self-attention as a map acting on probability measures.\nTheir upper bound grows more than exponentially withR𝑅Ritalic_Rso that the quadratic lower bound and the exponential upper bound put together provide a very loose estimation of the Lipschitz constant of self-attention on compact sets.\nFinally,Sander et al. (2022)propose a modification of the attention kernel that builds on the Sinkhorn-Knopp algorithm, and provide empirical evidence of the better properties of this new choice of kernel with respect to the classical one.\n\n",
    "origin_citations": [
      "Intriguing properties of neural networks.",
      "Explaining and harnessing adversarial examples.",
      "Distillation as a defense to adversarial perturbations against deep neural networks.",
      "Towards evaluating the robustness of neural networks.",
      "Globally-robust neural networks.",
      "Lipschitz-margin training: Scalable certification of perturbation invariance for deep neural networks.",
      "Formal guarantees on the robustness of a classifier against adversarial manipulation.",
      "Evaluating the robustness of neural networks: An extreme value theory approach.",
      "The lipschitz constant of self-attention, 2021.",
      "A mathematical theory of attention.",
      "The emergence of clusters in self-attention dynamics.",
      "On the robustness of text vectorizers.",
      "Revisiting and exploring efficient fast adversarial training via law: Lipschitz regularization and auto weight averaging.",
      "Certvit: Certified robustness of pre-trained vision transformers.",
      "Mitigating transformer overconfidence via lipschitz regularization.",
      "Lipsformer: Introducing lipschitz continuity to vision transformers.",
      "Stochastic deep networks.",
      "Approximation capability of neural networks on spaces of probability measures and tree-structured domains.",
      "A functional perspective on learning symmetric functions with neural networks.",
      "Sinkformers: Transformers with doubly stochastic attention.",
      "Bert: Pre-training of deep bidirectional transformers for language understanding.",
      "Neural ordinary differential equations.",
      "A mathematical perspective on transformers.",
      "Generating wikipedia by summarizing long sequences.",
      "Attention is all you need.",
      "Lipschitz normalization for self-attention layers with application to graph neural networks."
    ],
    "citations": [
      314,
      892,
      455027,
      890,
      474704,
      534296,
      16343,
      913039,
      351081,
      201663,
      115512,
      76866,
      405556,
      492048,
      305055,
      53946,
      382021,
      269940,
      238845,
      513152,
      7,
      7970,
      349252,
      1133,
      38,
      216850
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 95,
      "Publication_Potential": 96,
      "Final_Score": 95.3,
      "raw_text": "Content_Quality: 95\nPublication_Potential: 96\nFinal_Score: 95"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 75,
      "Final_Score": 82.6,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 88\nSynthesis_Analysis: 75\nFinal_Score: 82"
    },
    "overall_score": 87.67999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "aP0H8A1ywk",
        "forum": "aP0H8A1ywk",
        "number": 9302,
        "title": "How Smooth Is Attention?",
        "normalized_title": "how smooth is attention",
        "abstract": "Self-attention and masked self-attention are at the heart of Transformers' outstanding success. Still, our mathematical understanding of attention, in particular of its Lipschitz properties — which are key when it comes to analyzing robustness and expressive power — is incomplete. We provide a detailed study of the Lipschitz constant of self-attention in several practical scenarios, discussing the impact of the sequence length $n$ and layer normalization on the local Lipschitz constant of both unmasked and masked self-attention. In particular, we show that for inputs of length $n$ in any compact set, the Lipschitz constant of self-attention is bounded by $\\sqrt{n}$ up to a constant factor and that this bound is tight for reasonable sequence lengths. When the sequence length $n$ is too large for the previous bound to be tight, which we refer to as the mean-field regime, we provide an upper bound and a matching lower bound which are independent of $n$. Our mean-field framework for masked self-attention is novel and of independent interest. Our experiments on pretrained and randomly initialized BERT and GPT-2 support our theoretical findings.",
        "authors": [
          "Valérie Castin",
          "Pierre Ablin",
          "Gabriel Peyré"
        ],
        "decision": null,
        "pdate": 1714610490491,
        "cdate": 1706870873705
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "aP0H8A1ywk",
          "forum": "aP0H8A1ywk",
          "number": 9302,
          "title": "How Smooth Is Attention?",
          "normalized_title": "how smooth is attention",
          "abstract": "Self-attention and masked self-attention are at the heart of Transformers' outstanding success. Still, our mathematical understanding of attention, in particular of its Lipschitz properties — which are key when it comes to analyzing robustness and expressive power — is incomplete. We provide a detailed study of the Lipschitz constant of self-attention in several practical scenarios, discussing the impact of the sequence length $n$ and layer normalization on the local Lipschitz constant of both unmasked and masked self-attention. In particular, we show that for inputs of length $n$ in any compact set, the Lipschitz constant of self-attention is bounded by $\\sqrt{n}$ up to a constant factor and that this bound is tight for reasonable sequence lengths. When the sequence length $n$ is too large for the previous bound to be tight, which we refer to as the mean-field regime, we provide an upper bound and a matching lower bound which are independent of $n$. Our mean-field framework for masked self-attention is novel and of independent interest. Our experiments on pretrained and randomly initialized BERT and GPT-2 support our theoretical findings.",
          "authors": [
            "Valérie Castin",
            "Pierre Ablin",
            "Gabriel Peyré"
          ],
          "decision": null,
          "pdate": 1714610490491,
          "cdate": 1706870873705
        }
      ]
    }
  },
  {
    "paper_id": 26443,
    "title": "4Diffusion: Multi-view Video Diffusion Model for 4D Generation",
    "abstract": "Current 4D generation methods have achieved noteworthy efficacy with the aid of advanced diffusion generative models. However, these methods lack multi-view spatial-temporal modeling and encounter challenges in integrating diverse prior knowledge from multiple diffusion models, resulting in inconsistent temporal appearance and flickers. In this paper, we propose a novel 4D generation pipeline, namely4Diffusion, aimed at generating spatial-temporally consistent 4D content from a monocular video. We first design a unified diffusion model tailored for multi-view video generation by incorporating a learnable motion module into a frozen 3D-aware diffusion model to capture multi-view spatial-temporal correlations. After training on a curated dataset, our diffusion model acquires reasonable temporal consistency and inherently preserves the generalizability and spatial consistency of the 3D-aware diffusion model. Subsequently, we propose 4D-aware Score Distillation Sampling loss, which is based on our multi-view video diffusion model, to optimize 4D representation parameterized by dynamic NeRF. This aims to eliminate discrepancies arising from multiple diffusion models, allowing for generating spatial-temporally consistent 4D content. Moreover, we devise an anchor loss to enhance the appearance details and facilitate the learning of dynamic NeRF. Extensive qualitative and quantitative experiments demonstrate that our method achieves superior performance compared to previous methods.",
    "introduction": "In recent years, diffusion models have significantly impacted the era of image, video, and 3D generation. With the support of large-scale text-to-image diffusion models[43,1]and 3D-aware diffusion models[44,29,52], many works[26,39,51,9,34,41,50,30,55,27]leverage Score Distillation Sampling (SDS)[38]to distill the prior knowledge from diffusion models to optimize a 3D shape parameterized by NeRF[35]or 3DGS[21]. Although they have attained faithful results, they only focus on creating static 3D shapes, neglecting the dynamics of objects in the real world.\n\n\nGenerating 4D content, i.e., dynamic 3D content, holds diverse applications in the virtual realm, including digital human, gaming, media, and AR/VR. The main challenge lies in creating 4D content with vivid motion and high-quality spatial-temporal consistency. The pioneering study MAV3D[47]introduces a two-stage method, which first learns a static 3D shape with a text-to-image diffusion model and then deforms the static 3D shape with a text-to-video diffusion model[46]. However, MAV3D encounters the Janus problem and generates 4D contents with poor appearance and motion[4]. To overcome these issues, the following works[4,66,65,28]employ multiple diffusion models for distinct purposes. Specifically, these methods leverage 3D-aware diffusion models[44,29]and text-to-image diffusion models[43]to achieve spatial consistency and visually appealing appearance. Akin to MAV3D, they utilize video diffusion models[2,54,46]to add motion to create 4D content.\n\n\nThe aforementioned methods utilize multiple diffusion models for 4D generation. As Fig.2illustrates, when diffusing images rendered from a 3D model, the 3D-aware diffusion model[44]generates multi-view images to address the spatial ambiguity. On the other hand, the 2D image diffusion model[43]produces a clean image with subtle details to refine appearance. The 2D video diffusion model[2]generates dynamic frames to ensure temporal consistency within the same viewpoint. However, there is no accurate guidance to ensure multi-view spatial-temporal consistency due to the lack of multi-view spatial-temporal modeling. Moreover, it is challenging to integrate diverse prior knowledge from multiple diffusion models, often leading to inconsistent temporal appearance and flickers as shown in the second row of Fig.7.\n\n\nFigure 2:Challenges.The denoised images from Stable Diffusion (SD)[43], MVDream[44], and ZeroScope[2]. These diffusion models can not provide multi-view spatial-temporal guidance and exhibit discrepancies, making their integration challenging.\n\n\nIn this paper, we present a novel 4D generation pipeline, namely4Diffusion, to create high-quality spatial-temporally consistent 4D content from a monocular video. Specifically, we propose a unified diffusion model, 4DM, to capture multi-view spatial-temporal correlations for multi-view video generation. To achieve this, we construct 4DM based on the powerful pre-trained 3D-aware diffusion model[52], which already ensures high-quality multi-view spatial consistency. We then seamlessly integrate a motion module into the 3D-aware diffusion model to extend the temporal modeling capability. Contrary to previous attempts[16,14]that typically demand extensive large-scale video datasets for tuning the motion module, 4DM achieves reasonable temporal consistency and captures multi-view spatial-temporal correlations after training on only hundreds of multi-view videos. Importantly, we keep the parameters of the 3D-aware diffusion model unchanged to preserve the generalization ability and spatial consistency. 4DM provides multi-view spatial-temporal guidance for 4D generation. Therefore, we propose 4D-aware SDS loss to distill prior knowledge from 4DM to optimize 4D content parameterized by dynamic NeRF. This approach eliminates discrepancies arising from multiple diffusion models and stabilizes the optimizing process. Moreover, we use 4DM to generate anchor videos conditioned on the input monocular video and devise an anchor loss to enhance the appearance details, facilitating the learning of dynamic NeRF. Finally, we generate 4D content with high-quality spatial-temporal consistency and vibrant motion coherence with the input video as shown in Fig.1. Qualitative and quantitative experiments demonstrate that our method achieves state-of-the-art performance on multi-view video generation and 4D generation from monocular videos.\n\n\nTo summarize, our contributions are as follows:1)We present4Diffusion, a novel 4D generation pipeline that generates high-quality spatial-temporal consistent 4D content from a monocular video with a multi-view video diffusion model.2)We propose a multi-view video diffusion model, 4DM, which provides multi-view spatial-temporal guidance for 4D generation. It trains on only hundreds of curated high-quality multi-view videos to capture multi-view spatial-temporal correlations.3)We combine 4D-aware SDS loss and an anchor loss based on 4DM to optimize dynamic NeRF, which stabilizes the training process and allows for generating high-quality 4D content.\n\n",
    "related_work": "Recent breakthroughs in multiple research domains have significantly accelerated progress in 4D generation task. Here, we discuss the most relevant fields, including 3D generation, video and 3D-aware diffusion models, and 4D generation.\n\n\n3D Generation.Recent studies in 3D generation can be classified into three categories: 3D generative methods[53,17,45,36,60,3,58,12], feed forward methods[18,49,23,67], and diffusion prior-based methods[26,39,51,9,34,41,50,30,55,27]. Inspired by the advancements in 2D content creation, 3D generative methods utilize the robust diffusion[53]or flow-based[60]backbone to generate 3D data represented by Signed Distance Function (SDF)[60], voxel grid[36], triplane[8,17,45], or weights of neural network[12]. However, these methods require time-consuming pre-training to fit each 3D data and are limited to creating a single category. Feed forward methods[18,23]adopt image features extracted from the pre-trained visual encoder DINO[7]to reconstruct 3D representations through a highly scalable and efficient transformer-based decoder. Although they can produce a 3D shape in a few seconds, they demand extensive training on large-scale 3D datasets, which is impractical with limited 4D datasets for 4D generation. Furthermore, diffusion prior-based methods distill prior knowledge from diffusion generative models via SDS[38]to optimize 3D representations, enabling the generation of high-quality 3D shapes with strong generalizability. In contrast to static 3D generation, our method focuses on creating 4D content.\n\n\nVideo and 3D-aware Diffusion Models.With the success of large-scale text-to-image diffusion models[43,1], recent works attempt to use diffusion models to generate more complex signals, including video and 3D. AnimateDiff[16]inserts a learnable motion module into the frozen text-to-image model for video generation, which preserves the efficacy of the text-to-image model while successfully modeling temporal information. Recent 3D-aware diffusion model Zero-1-to-3[29]adopts a stable diffusion model conditioned on relative camera pose and a single image for novel view synthesis. However, this method still suffers from the Janus problem and content drafting problem[44]due to the lack of explicit 3D modeling. Approaches like[31,44,59,32,52]leverage 3D-aware attention block to model the joint probability distribution of multi-view images, leading to spatially consistent generation. However, these approaches are incapable of producing multi-view consistent videos, due to the absence of temporal or spatial modeling.\n\n\nFigure 3:4Diffusion overview.Our method first trains a unified diffusion, named 4DM, by inserting a learnable motion module at the end of each frozen spatial module of ImageDream to capture multi-view spatial-temporal correlations. Given a monocular video and text prompt, 4DM can produce consistent multi-view videos. Then, we combine 4D-aware SDS and an anchor loss based on 4DM to optimize 4D content parameterized by Dynamic NeRF.\n\n\n4D Generation.Recently, several works have delved into 4D generation from various user-friendly prompts, such as text[47,4,66,28], a single image[65,66], and a monocular video[42,20,62]. The pioneering study MAV3D[47]proposes a two-stage method to optimize 4D representation, i.e., Hexplane[6], with both text-to-image and text-to-video diffusion models in a static-to-dynamic manner. To generate 4D contents with realistic appearance, Dream-in-4D[66]and 4D-fy[4]combine hybrid diffusion models. Specifically, they utilize 3D-aware and 2D diffusion guidance to learn a static 3D representation and incorporate video diffusion guidance to add motion. However, these diffusion models can not offer multi-view spatial-temporally consistent guidance and it is difficult to integrate diverse prior knowledge from multiple diffusion models, resulting in suboptimal results. In contrast to these approaches, we design a unified model to capture multi-view spatial-temporal correlations for 4D generation.\n\n\nSimilar to us,[20,42,63,62,57]generate 4D content from a monocular video. Consistent4D[20]introduces an interpolation-driven loss between two adjacent frames to enhance spatial-temporal consistency. However, Consistent4D lacks temporal modeling cross frames. DreamGaussian4D[42], 4DGen[62], and SC4D[57]combine 4D Gaussian Splatting[56,19]into 4D generation pipeline. Although they notably reduce optimization time, they may result in blurred appearance and inaccurate geometry due to the explicit characteristics of Gaussians. STAG4D[63]proposes a training-free strategy to generate sparse anchor multi-view videos for 4D generation. In contrast, we propose a multi-view video diffusion model to provide multi-view spatial-temporal consistency guidance for 4D generation.\n\n",
    "origin_citations": [
      "Rodin: A generative model for sculpting 3d digital avatars using diffusion.",
      "3dgen: Triplane latent diffusion for textured mesh generation.",
      "3d neural field generation using triplane diffusion.",
      "Diffrf: Rendering-guided 3d radiance field diffusion.",
      "Mosaic-sdf for 3d generative models.",
      "Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation.",
      "Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model.",
      "Hyperdiffusion: Generating implicit neural fields with weight-space diffusion.",
      "Lrm: Large reconstruction model for single image to 3d.",
      "Lgm: Large multi-view gaussian model for high-resolution 3d content creation.",
      "Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.",
      "Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers.",
      "Magic3d: High-resolution text-to-3d content creation.",
      "Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors.",
      "Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior.",
      "Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.",
      "Realfusion: 360° reconstruction of any object from a single image.",
      "Dreambooth3d: Subject-driven text-to-3d generation.",
      "Dreamgaussian: Generative gaussian splatting for efficient 3d content creation.",
      "Humangaussian: Text-driven 3d human generation with gaussian splatting.",
      "Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation.",
      "Consistent123: One image to highly consistent 3d asset using case-aware diffusion priors.",
      "Efficient geometry-aware 3d generative adversarial networks.",
      "Emerging properties in self-supervised vision transformers.",
      "Dreamfusion: Text-to-3d using 2d diffusion.",
      "High-resolution image synthesis with latent diffusion models.",
      "2023.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Mvdream: Multi-view diffusion for 3d generation.",
      "Syncdreamer: Generating multiview-consistent images from a single-view image.",
      "Consistnet: Enforcing 3d consistency for multi-view images diffusion.",
      "Wonder3d: Single image to 3d using cross-domain diffusion.",
      "Imagedream: Image-prompt multi-view diffusion for 3d generation.",
      "Text-to-4d dynamic scene generation.",
      "4d-fy: Text-to-4d generation using hybrid score distillation sampling.",
      "A unified approach for text-and image-guided 4d scene generation.",
      "Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models.",
      "Animate124: Animating one image to 4d dynamic scene.",
      "Dreamgaussian4d: Generative 4d gaussian splatting.",
      "Consistent4d: Consistent 360° dynamic object generation from monocular video.",
      "4dgen: Grounded 4d content generation with spatial-temporal consistency.",
      "Hexplane: A fast representation for dynamic scenes.",
      "Stag4d: Spatial-temporal anchored generative 4d gaussians.",
      "Sc4d: Sparse-controlled video-to-4d generation and motion transfer.",
      "4d gaussian splatting for real-time dynamic scene rendering.",
      "Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes."
    ],
    "citations": [
      221397,
      76870,
      223530,
      270311,
      245260,
      492427,
      295035,
      99387,
      294218,
      270342,
      245032,
      1009434,
      493794,
      493936,
      226975,
      494102,
      477674,
      22498,
      258271,
      477383,
      884251,
      2514,
      176110,
      2240,
      995727,
      185878,
      495077,
      1087300,
      129306,
      315924,
      313594,
      274286,
      212984,
      22093,
      22606,
      347883,
      23669,
      346294,
      295772,
      346260,
      213655,
      425405,
      448503,
      320337,
      273928
    ],
    "extract_rate": 0.9574468085106383,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 79,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 89,
      "Final_Score": 88.4,
      "raw_text": "Content_Coherence: 79\nCitation_Quality: 95\nSynthesis_Analysis: 89\nFinal_Score: 88"
    },
    "overall_score": 87.64,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "SFk7AMpyhx",
        "forum": "SFk7AMpyhx",
        "number": 9911,
        "title": "4Diffusion: Multi-view Video Diffusion Model for 4D Generation",
        "normalized_title": "4diffusion multi view video diffusion model for 4d generation",
        "abstract": "Current 4D generation methods have achieved noteworthy efficacy with the aid of advanced diffusion generative models. However, these methods lack multi-view spatial-temporal modeling and encounter challenges in integrating diverse prior knowledge from multiple diffusion models, resulting in inconsistent temporal appearance and flickers. In this paper, we propose a novel 4D generation pipeline, namely $\\textbf{4Diffusion}$, aimed at generating spatial-temporally consistent 4D content from a monocular video. We first design a unified diffusion model tailored for multi-view video generation by incorporating a learnable motion module into a frozen 3D-aware diffusion model to capture multi-view spatial-temporal correlations. After training on a curated dataset, our diffusion model acquires reasonable temporal consistency and inherently preserves the generalizability and spatial consistency of the 3D-aware diffusion model. Subsequently, we propose 4D-aware Score Distillation Sampling loss, which is based on our multi-view video diffusion model, to optimize 4D representation parameterized by dynamic NeRF. This aims to eliminate discrepancies arising from multiple diffusion models, allowing for generating spatial-temporally consistent 4D content. Moreover, we devise an anchor loss to enhance the appearance details and facilitate the learning of dynamic NeRF. Extensive qualitative and quantitative experiments demonstrate that our method achieves superior performance compared to previous methods.",
        "authors": [
          "Haiyu Zhang",
          "Xinyuan Chen",
          "Yaohui Wang",
          "Xihui Liu",
          "Yunhong Wang",
          "Yu Qiao"
        ],
        "decision": null,
        "pdate": 1727287924515,
        "cdate": 1715687781772
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "SFk7AMpyhx",
          "forum": "SFk7AMpyhx",
          "number": 9911,
          "title": "4Diffusion: Multi-view Video Diffusion Model for 4D Generation",
          "normalized_title": "4diffusion multi view video diffusion model for 4d generation",
          "abstract": "Current 4D generation methods have achieved noteworthy efficacy with the aid of advanced diffusion generative models. However, these methods lack multi-view spatial-temporal modeling and encounter challenges in integrating diverse prior knowledge from multiple diffusion models, resulting in inconsistent temporal appearance and flickers. In this paper, we propose a novel 4D generation pipeline, namely $\\textbf{4Diffusion}$, aimed at generating spatial-temporally consistent 4D content from a monocular video. We first design a unified diffusion model tailored for multi-view video generation by incorporating a learnable motion module into a frozen 3D-aware diffusion model to capture multi-view spatial-temporal correlations. After training on a curated dataset, our diffusion model acquires reasonable temporal consistency and inherently preserves the generalizability and spatial consistency of the 3D-aware diffusion model. Subsequently, we propose 4D-aware Score Distillation Sampling loss, which is based on our multi-view video diffusion model, to optimize 4D representation parameterized by dynamic NeRF. This aims to eliminate discrepancies arising from multiple diffusion models, allowing for generating spatial-temporally consistent 4D content. Moreover, we devise an anchor loss to enhance the appearance details and facilitate the learning of dynamic NeRF. Extensive qualitative and quantitative experiments demonstrate that our method achieves superior performance compared to previous methods.",
          "authors": [
            "Haiyu Zhang",
            "Xinyuan Chen",
            "Yaohui Wang",
            "Xihui Liu",
            "Yunhong Wang",
            "Yu Qiao"
          ],
          "decision": null,
          "pdate": 1727287924515,
          "cdate": 1715687781772
        }
      ]
    }
  },
  {
    "paper_id": 16771,
    "title": "Fast Timing-Conditioned Latent Audio Diffusion",
    "abstract": "Generating long-form 44.1kHz stereo audio from text prompts can be computationally demanding. Further, most previous works do not tackle that music and sound effects naturally vary in their duration.\nOur research focuses on the efficient generation of long-form, variable-length stereo music and sounds at 44.1kHz using text prompts with a generative model.\nStable Audio is based on latent diffusion, with its latent defined by a fully-convolutional variational autoencoder. It is conditioned on text prompts as well as timing embeddings, allowing for fine control over both the content and length of the generated music and sounds.\nStable Audio is capable of rendering stereo signals of up to 95 sec at 44.1kHz in 8 sec on an A100 GPU.\nDespite its compute efficiency and fast inference, it is one of the best in two public text-to-music and -audio benchmarks and, differently from state-of-the-art models, can generate music with structure and stereo sounds.",
    "introduction": "The introduction of diffusion-based generative models(Sohl-Dickstein et al.,2015; Ho et al.,2020)has\nlead to rapid improvements in the quality and controllability of generated images(Podell et al.,2023), video(Blattmann et al.,2023), and audio(Rouard & Hadjeres,2021; Liu et al.,2023a).\n\n\nOne challenge is that diffusion models working within the raw signal space tend to be computationally demanding during both training and inference. Diffusion models working in the latent space of a pre-trained autoencoder, termed “latent diffusion models”(Rombach et al.,2022), are significantly more compute efficient.\nWorking with a heavily downsampled latent representation of audio allows for much faster inference times compared to raw audio, and also allows generating long-form audio (e.g., 95 sec).\n\n\nAnother challenge with audio diffusion models is that those are usually trained to generate fixed-size outputs(Huang et al.,2023a), e.g., a model trained on 30 sec audio chunks will generate 30 sec outputs. This is an issue when training on and trying to generate audio of varying lengths, as is the case when generating full songs or sound effects.\nHence audio diffusion models are commonly trained on randomly cropped chunks from longer audios, cropped or padded to fit the diffusion model’s training length. With music, e.g., this causes the model to generate arbitrary sections of a song, which may start or end in the middle of a musical phrase.\n\n\nStable Audio is based on a latent diffusion model for audio conditioned on a text prompt as well as timing embeddings, allowing for control over the content and length of the generated music and sound effects. This additional timing conditioning allows us to generate audio of a specified (variable) length up to the training window length.\nDue to the compute efficient nature of latent diffusion modeling, it can generate long-form content in short inference times.\nIt can render up to 95 sec (our training window length) of stereo audio at 44.1kHz in 8 sec on an A100 GPU (40GB VRAM).\n\n\nThe commonly used metrics for generative audio are designed to evaluate short-form mono signals at 16kHz(Kilgour et al.,2018). Yet, our work focuses on generating long-form full-band stereo signals. We propose: (i) a Fréchet Distance based on OpenL3 embeddings(Cramer et al.,2019)to evaluate the plausibility of the generated long-form full-band stereo signals, (ii) a Kullback-Leibler divergence to evaluate the semantic correspondence between lengthy generated and reference audios up to 32kHz, and (iii) a CLAP score to evaluate how long-form full-band stereo audios adhere to the given text prompt.\nWe also conduct a qualitative study, assessing audio quality and text alignment, while also pioneering the assessment of musicality, stereo correctness, and musical structure.\nWe show that Stable Audio can obtain state-of-the-art results on long-form full-band stereo music and sound effects generation from text and timing inputs.\nWe also show that, differently from previous works, Stable Audio is also capable to generate structured music (with intro, development, outro) and stereo sound effects.\n\n\nCode to reproduce our model/metrics and demos is online111Model:https://github.com/Stability-AI/stable-audio-tools.Metrics:https://github.com/Stability-AI/stable-audio-metrics.Demo:https://stability-ai.github.io/stable-audio-demo..\n\n",
    "related_work": "Autoregressive models— WaveNet(Oord et al.,2016)autoregressively models quantized audio samples, but is slow during inference because it operates with waveforms.\nRecent autoregressive models addressed this by operating on a quantized latent space, enabling faster processing. Jukebox(Dhariwal et al.,2020)relies on a multi-scale approach to encode music into a sequence of quantized latents and subsequently models them using autoregressive transformers.\nRecent work such as MusicLM(Agostinelli et al.,2023)and MusicGen(Copet et al.,2023)utilize a similar approach and also autoregressively model quantized latent sequences.\nHowever, unlike Jukebox, such models are conditioned on text prompts rather than on artist, genre, and/or lyrics.\nAutoregressive models similar to MusicLM (AudioLM) and MusicGen (AudioGen) have also been used for sound synthesis(Borsos et al.,2023; Kreuk et al.,2022)and for generating music accompaniments from singing(Donahue et al.,2023).\nOur work is not based on autoregressive modeling.\n\n\nNon-autoregressive models— Parallel WaveNet(Oord et al.,2018)and adversarial audio synthesis(Donahue et al.,2018; Pasini & Schlüter,2022)were developed to tackle the computational inefficiencies inherent in autoregressive modeling.\nRecent works like VampNet(Garcia et al.,2023), StemGen(Parker et al.,2024)and MAGNeT(Ziv et al.,2024)are based on masked token modeling(Chang et al.,2022). These are for creating musical variations,\ngenerating additional stems for a given song, and to efficiently synthesize music and sounds, respectively.\nFlow-matching generative modeling(Vyas et al.,2023)was also recently introduced for speech and sounds synthesis.\nOur work is not based on any of the non-autoregressive models above.\n\n\nEnd-to-end diffusion models— CRASH(Rouard & Hadjeres,2021)was proposed for unconditional drums synthesis,\nDAG(Pascual et al.,2023)for class-conditional sounds synthesis, Noise2Music(Huang et al.,2023a)for text-conditional music synthesis, andMariani et al. (2023)built an end-to-end diffusion model capable of both music synthesis and source separation.\nOur work is also based on diffusion, albeit not in an end-to-end fashion. Rather, it involves latent diffusion due to its computational efficiency.\n\n\nSpectrogram diffusion models— Riffusion(Forsgren & Martiros,2022)fine-tuned Stable Diffusion to generate spectrograms from text prompts,Hawthorne et al. (2022)addressed MIDI-to-spectrogram generation, and\nCQT-Diff(Moliner et al.,2023)relied on CQT spectrograms for bandwidth extension, inpatining, and declipping.\nAn additional step is required to render waveforms from magnitude spectrograms.\nOur work is also based on diffusion, albeit it does not rely on spectrogram-based synthesis.\n\n\nLatent diffusion models— Moûsai(Schneider et al.,2023)and AudioLDM(Liu et al.,2023a)pioneered using latent diffusion for text-to-music and -audio.\nTheir main difference being that Moûsai decodes latents onto waveforms through a diffusion decoder, while AudioLDM decodes latents onto spectrograms which are then inverted to waveforms with HiFi-GAN(Kong et al.,2020).\nAudioLDM2(Liu et al.,2023b)extends AudioLDM to also synthesize speech by using a shared representation for music, audio, and speech to condition the latent diffusion model.\nJEN-1(Li et al.,2023)is anomnidirectionallatent diffusion model trained in a multitask fashion.\nJEN-1 Composer(Yao et al.,2023)is its extension for multi-track music generation.Levy et al. (2023)explored sampling-time guidance for both end-to-end and latent diffusion models.\nAll previous works constrain the latent to be normalized, often with a variational autoencoder (VAE). The exceptions being JEN-1, which runs over a dimensionality reduced latent that is normalized based on the mean and covariance, and Moûsai that simply uses a tanh.\nOur work is also based on latent diffusion, and we normalize latents by using a VAE. AppendixDincludes further discussion on related latent diffusion models.\n\n\nHigh sampling rate and stereo generation—\nMoûsai and JEN-1 generate 48kHz stereo music.\nAudioLDM2 can generate 48kHz mono music.Levy et al. (2023)generates 44.1kHz stereo music.\nNo other prior works generate music up to the standard specifications of commercial music (44.1kHz stereo). DAG and AudioLDM2 generate 48kHz mono sounds, and we are not aware of prior works tackling stereo sound synthesis.\nOur work focuses on generating 44.1kHz stereo music and sounds from text prompts.\n\n\nText embeddings— CLAP(Wu et al.,2023)and T5-like(Raffel et al.,2020; Ghosal et al.,2023)text embeddings are commonly used because of their open-source nature. CLAP relies on a contrastive (multimodal) language-audio pretraining, and T5 is a large language model. Further, MusicLM uses MuLan(Huang et al.,2022), that is also based on contrastive language-audio pretraining but on their private dataset.\nOur work relies on a CLAP-based model trained in a contrastive language-audio fashion on our dataset.\n\n\nFast generation of variable-length, long-form audio— Autoregressive models can generate long-form audio of variable length due to their sequential (one-sample-at-a-time generation) nature, but are slow at inference time.\nPrevious non-autoregressive models were trained to generate up to 20 sec long music(Parker et al.,2024).\nPrevious end-to-end and latent diffusion models were trained to generate up to 30 sec long music(Huang et al.,2023a; Levy et al.,2023; Lam et al.,2024), with the exception of Moûsai that was trained to generate 44 sec. Hence, previous works are either slow at inference time (autoregressive models) or cannot generate variable-length, long-form audio (the rest).\nOur work relies on latent diffusion to generate long-form (up to 95 sec), variable-length (controlled by the timing condition) stereo signals at 44.1kHz in 8 sec on an A100 GPU (40GB VRAM).\n\n\nTiming conditioning—\nThe use of learned embeddings to condition music generation models on timing information was introduced by Jukebox(Dhariwal et al.,2020), an autoregressive model conditioned with timing information on: (i) song duration, (ii) starting time of the training/generated audio sample within the song, and (iii) how much fraction of the song has elapsed.\nWe are not aware of previous works using timing conditioning for conditioning (latent) diffusion models.\nOur work employs timing conditioning to control the length of the generations, enabling our latent diffusion models to generate variable-length outputs.\n\n\nEvaluation metrics— The commonly used quantitative audio metrics were developed for evaluating short-form mono audio generations at 16kHz(Kilgour et al.,2018; Copet et al.,2023). Yet, our work focuses on generating long-form full-band stereo signals. OnlyPascual et al. (2023)explored quantitative metrics for evaluating full-band audio, although their focus was short-form mono signals. Our work explores new quantitative metrics to evaluate long-form full-band stereo generations.\nQualitative metrics assessing audio quality and text alignment are also prevalent in the literature(Dong et al.,2023; Copet et al.,2023; Ziv et al.,2024). Our work also explores additional qualitative metrics to evaluate musicality, stereo correctness, and musical structure.\n\n\nMultitask generative modeling—\nWhile generative models have traditionally focused on specific tasks like speech, music or sound synthesis, recent works showed success in addressing all these tasks simultaneously(Yang et al.,2023; Liu et al.,2023b).\nOur work relies on one model to generate both music and sound (not speech) given a text prompt.\n\n",
    "origin_citations": [
      "Wavenet: A generative model for raw audio.",
      "Jukebox: A generative model for music.",
      "Musiclm: Generating music from text.",
      "Simple and controllable music generation.",
      "Audiolm: a language modeling approach to audio generation.",
      "Audiogen: Textually guided audio generation.",
      "Singsong: Generating musical accompaniments from singing.",
      "Parallel wavenet: Fast high-fidelity speech synthesis.",
      "Adversarial audio synthesis.",
      "Musika! fast infinite waveform music generation.",
      "Vampnet: Music generation via masked acoustic token modeling.",
      "Stemgen: A music generation model that listens.",
      "Masked audio generation using a single non-autoregressive transformer.",
      "Maskgit: Masked generative image transformer.",
      "Audiobox: Unified audio generation with natural language prompts.",
      "Crash: Raw audio score-based generative modeling for controllable high-resolution drum sound synthesis.",
      "Full-band general audio synthesis with score-based diffusion.",
      "Noise2music: Text-conditioned music generation with diffusion models.",
      "Multi-source diffusion models for simultaneous music generation and separation.",
      "Riffusion - stable diffusion for real-time music generation.",
      "Multi-instrument music synthesis with spectrogram diffusion.",
      "Solving audio inverse problems with a diffusion model.",
      "Moûsai: Text-to-music generation with long-context latent diffusion.",
      "Audioldm: Text-to-audio generation with latent diffusion models.",
      "Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis.",
      "Audioldm 2: Learning holistic audio generation with self-supervised pretraining.",
      "Jen-1: Text-guided universal music generation with omnidirectional diffusion models.",
      "Jen-1 composer: A unified framework for high-fidelity multi-track music generation.",
      "Controllable music production with diffusion models and guidance gradients.",
      "Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation.",
      "Exploring the limits of transfer learning with a unified text-to-text transformer.",
      "Text-to-audio generation using instruction-tuned llm and latent diffusion model.",
      "Mulan: A joint embedding of music audio and natural language.",
      "Efficient neural music generation.",
      "Fréchet audio distance: A metric for evaluating music enhancement algorithms.",
      "Clipsonic: Text-to-audio synthesis with unlabeled videos and pretrained language-vision models.",
      "Uniaudio: An audio foundation model toward universal audio generation."
    ],
    "citations": [
      4722,
      8727,
      8033,
      305828,
      8029,
      882944,
      212291,
      398315,
      5904,
      68793,
      185896,
      270566,
      484346,
      89438,
      346910,
      193352,
      334168,
      490343,
      491107,
      411847,
      333880,
      212740,
      8037,
      355182,
      180191,
      180655,
      311529,
      296900,
      32464,
      9,
      52977,
      8030,
      258567,
      328860,
      50521,
      324311
    ],
    "extract_rate": 0.972972972972973,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 88,
      "Final_Score": 83.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 87,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 90,
      "Final_Score": 89.9,
      "raw_text": "Content_Coherence: 87\nCitation_Quality: 92\nSynthesis_Analysis: 90\nFinal_Score: 90"
    },
    "overall_score": 87.46000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "jOlO8t1xdx",
        "forum": "jOlO8t1xdx",
        "number": 3875,
        "title": "Fast Timing-Conditioned Latent Audio Diffusion",
        "normalized_title": "fast timing conditioned latent audio diffusion",
        "abstract": "Generating long-form 44.1kHz stereo audio from text prompts can be computationally demanding. Further, most previous works do not tackle that music and sound effects naturally vary in their duration. Our research focuses on the efficient generation of long-form, variable-length stereo music and sounds at 44.1kHz using text prompts with a generative model. It is based on latent diffusion, with its latent defined by a fully-convolutional variational autoencoder. The generative model is conditioned on text prompts as well as timing embeddings, allowing for fine control over both the content and length of the generated music and sounds. It is capable of rendering stereo signals of up to 95 sec at 44.1kHz in 8 sec on an A100 GPU. Despite its compute efficiency and fast inference, the proposed model is one of the best in two public text-to-music and -audio benchmarks and, differently from state-of-the-art models, can generate music with structure and stereo sounds.",
        "authors": [
          "Zach Evans",
          "CJ Carr",
          "Josiah Taylor",
          "Scott H. Hawley",
          "Jordi Pons"
        ],
        "decision": null,
        "pdate": 1714610348402,
        "cdate": 1706701606291
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "jOlO8t1xdx",
          "forum": "jOlO8t1xdx",
          "number": 3875,
          "title": "Fast Timing-Conditioned Latent Audio Diffusion",
          "normalized_title": "fast timing conditioned latent audio diffusion",
          "abstract": "Generating long-form 44.1kHz stereo audio from text prompts can be computationally demanding. Further, most previous works do not tackle that music and sound effects naturally vary in their duration. Our research focuses on the efficient generation of long-form, variable-length stereo music and sounds at 44.1kHz using text prompts with a generative model. It is based on latent diffusion, with its latent defined by a fully-convolutional variational autoencoder. The generative model is conditioned on text prompts as well as timing embeddings, allowing for fine control over both the content and length of the generated music and sounds. It is capable of rendering stereo signals of up to 95 sec at 44.1kHz in 8 sec on an A100 GPU. Despite its compute efficiency and fast inference, the proposed model is one of the best in two public text-to-music and -audio benchmarks and, differently from state-of-the-art models, can generate music with structure and stereo sounds.",
          "authors": [
            "Zach Evans",
            "CJ Carr",
            "Josiah Taylor",
            "Scott H. Hawley",
            "Jordi Pons"
          ],
          "decision": null,
          "pdate": 1714610348402,
          "cdate": 1706701606291
        }
      ]
    }
  },
  {
    "paper_id": 29394,
    "title": "Translation Equivariant Transformer Neural Processes",
    "abstract": "The effectiveness ofneural processes (NPs)in modelling posterior prediction maps—the mapping from data to posterior predictive distributions—has significantly improved since their inception. This improvement can be attributed to two principal factors:(1)advancements in the architecture of permutation invariant set functions, which are intrinsic to allNPs; and(2)leveraging symmetries present in the true posterior predictive map, which are problem dependent.Transformers are a notable development in permutation invariant set functions, and their utility withinNPshas been demonstrated through the family of models we refer to astransformer neural processes (TNPs). Despite significant interest inTNPs, little attention has been given to incorporating symmetries. Notably, the posterior prediction maps for data that are stationary—a common assumption in spatio-temporal modelling—exhibit translation equivariance. In this paper, we introduce of a new family oftranslation equivariant TNPs (TE-TNPs)that incorporatetranslation equivariance. Through an extensive range of experiments on synthetic and real-world spatio-temporal data, we demonstrate the effectiveness ofTE-TNPsrelative to their non-translation-equivariant counterparts and otherNPbaselines.",
    "introduction": "Transformers have emerged as an immensely effective architecture for natural language processing and computer vision tasks(Vaswani et al.,2017; Dosovitskiy et al.,2020). They have become the backbone for many state-of-the-art models—such ChatGPT(Achiam et al.,2023)and DALL-E(Betker et al.,2023)—owing to their ability to learn complex dependencies amongst input data.\nMore generally, transformers can be understood as permutation equivariant set functions. This abstraction has led to the deployment of transformers in domains beyond that of sequence modelling, including particle physics, molecular modelling, climate science, and Bayesian inference(Lee et al.,2019; Fuchs et al.,2020; Müller et al.,2021).\n\n\nNPs(Garnelo et al.,2018a,b)are a broad family of meta-learning models which learn the mapping from sets of observed datapoints to predictive stochastic processes(Foong et al.,2020). They are straightforward to train, handle off-the-grid data and missing observations with ease, and can be easily adapted for different data modalities.\nThis flexibility makes them an attractive choice for a wide variety of problem domains, including spatio-temporal modelling, healthcare, and few-shot learning(Jha et al.,2022).\nExchangeability in the predictive distribution with respect to the context set is achieved through the use of permutation invariant set functions, which, inNPs, map from the sets of observations to some representation space. Given the utility of transformers as set functions, it is natural to consider their use withinNPs. This gives rise toTNPs.\n\n\nThe family ofTNPsinclude theattentive NP (ANP)(Kim et al.,2019),diagonal TNP (TNP-D),autoregressive TNP (TNP-AR), andnon-diagonal TNP (TNP-ND)(Nguyen & Grover,2022), and thelatent-bottlenecked ANP (LBANP)(Kim et al.,2019).\nDespite a significant amount of interest inTNPsfrom the research community, there are certain properties that we may wish our model to possess that have not yet been addressed. In particular, for spatio-temporal problems the data is often roughly stationary, in which case it is desirable to equip our model with translation equivariance: if the data are translated in space or time, then the predictions of our model should be translated correspondingly.\nAlthough translation equivariance has been incorporated into other families ofNPmodels, such as theconvolutional conditional NP (ConvCNP)(Gordon et al.,2019)andrelational CNP (RCNP)(Huang et al.,2023), it is yet to be incorporated into theTNP. The key ingredient to achieving this is to establish effective translation equivariant attention layers that can be used in place of the standard attention layers within the transformer encoder.\nIn this paper, we develop theTE-TNP.\nOur contributions are as follows:1.We develop an effective method for incorporating translation equivariance into the attention mechanism of transformers, developing thetranslation equivariant multi-head self attention (TE-MHSA)andtranslation equivariant multi-head cross attention (TE-MHCA)operations. These operations replace standard MHSA and MHCA operations within transformer encoders to obtain a new family of translation equivariantTNPs.2.We use pseudo-tokens to reduce the quadratic computational complexity ofTE-TNPs, developingtranslation equivariant PT-TNPs (TE-PT-TNPs).3.We demonstrate the efficacy ofTE-TNPsrelative to existing NPs—including theConvCNPand theRCNP—on a number of synthetic and real-world spatio-temporal modelling problems.\n\n",
    "related_work": "Throughout this section, we will use the following notation. Let𝒳=ℝDx𝒳superscriptℝsubscript𝐷𝑥\\mathcal{X}=\\mathbb{R}^{D_{x}}caligraphic_X = blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝒴=ℝDy𝒴superscriptℝsubscript𝐷𝑦\\mathcal{Y}=\\mathbb{R}^{D_{y}}caligraphic_Y = blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT end_POSTSUPERSCRIPTdenote the input and output spaces, and let(𝐱,𝐲)∈𝒳×𝒴𝐱𝐲𝒳𝒴(\\mathbf{x},\\mathbf{y})\\in\\mathcal{X}\\times\\mathcal{Y}( bold_x , bold_y ) ∈ caligraphic_X × caligraphic_Ydenote an input–output pair. Let𝒮=⋃N=0∞(𝒳×𝒴)N𝒮superscriptsubscript𝑁0superscript𝒳𝒴𝑁\\mathcal{S}=\\bigcup_{N=0}^{\\infty}(\\mathcal{X}\\times\\mathcal{Y})^{N}caligraphic_S = ⋃ start_POSTSUBSCRIPT italic_N = 0 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ∞ end_POSTSUPERSCRIPT ( caligraphic_X × caligraphic_Y ) start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPTbe a collection of all finite data sets, which includes the empty set∅\\varnothing∅, the data set containing no data points.\nWe denote a context and target set with𝒟c,𝒟t∈𝒮subscript𝒟𝑐subscript𝒟𝑡𝒮\\mathcal{D}_{c},\\ \\mathcal{D}_{t}\\in\\mathcal{S}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ caligraphic_S, where|𝒟c|=Ncsubscript𝒟𝑐subscript𝑁𝑐|\\mathcal{D}_{c}|=N_{c}| caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT | = italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT,|𝒟t|=Ntsubscript𝒟𝑡subscript𝑁𝑡|\\mathcal{D}_{t}|=N_{t}| caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | = italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT.\nLet𝐗c∈ℝNc×Dxsubscript𝐗𝑐superscriptℝsubscript𝑁𝑐subscript𝐷𝑥\\mathbf{X}_{c}\\in\\mathbb{R}^{N_{c}\\times D_{x}}bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝐘c∈ℝNc×Dysubscript𝐘𝑐superscriptℝsubscript𝑁𝑐subscript𝐷𝑦\\mathbf{Y}_{c}\\in\\mathbb{R}^{N_{c}\\times D_{y}}bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT end_POSTSUPERSCRIPTbe the inputs and corresponding outputs of𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT, with𝐗t∈ℝNt×Dxsubscript𝐗𝑡superscriptℝsubscript𝑁𝑡subscript𝐷𝑥\\mathbf{X}_{t}\\in\\mathbb{R}^{N_{t}\\times D_{x}}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝐘t∈ℝNt×Dysubscript𝐘𝑡superscriptℝsubscript𝑁𝑡subscript𝐷𝑦\\mathbf{Y}_{t}\\in\\mathbb{R}^{N_{t}\\times D_{y}}bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_y end_POSTSUBSCRIPT end_POSTSUPERSCRIPTdefined analogously. We denote a single task asξ=(𝒟c,𝒟t)=((𝐗c,𝐘c),(𝐗t,𝐘t))𝜉subscript𝒟𝑐subscript𝒟𝑡subscript𝐗𝑐subscript𝐘𝑐subscript𝐗𝑡subscript𝐘𝑡\\xi=(\\mathcal{D}_{c},\\mathcal{D}_{t})=((\\mathbf{X}_{c},\\mathbf{Y}_{c}),(%\n\\mathbf{X}_{t},\\mathbf{Y}_{t}))italic_ξ = ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) = ( ( bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) , ( bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) ). Let𝒫⁢(𝒳)𝒫𝒳\\mathcal{P}(\\mathcal{X})caligraphic_P ( caligraphic_X )denote the collection of stochastic processes on𝒳𝒳\\mathcal{X}caligraphic_X.\n\n\n2.1Neural ProcessesNPs(Garnelo et al.,2018a,b)aim to learn the mapping from context sets𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTto ground truth posterior distributions over the target outputs,𝒟c↦p⁢(𝐘t|𝐗t,𝒟c)maps-tosubscript𝒟𝑐𝑝conditionalsubscript𝐘𝑡subscript𝐗𝑡subscript𝒟𝑐\\mathcal{D}_{c}\\mapsto p(\\mathbf{Y}_{t}|\\mathbf{X}_{t},\\mathcal{D}_{c})caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ↦ italic_p ( bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ), using meta-learning. This mapping is known as theposterior prediction mapπP:𝒮→𝒫⁢(𝒳):subscript𝜋𝑃→𝒮𝒫𝒳\\pi_{P}:\\mathcal{S}\\rightarrow\\mathcal{P}(\\mathcal{X})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT : caligraphic_S → caligraphic_P ( caligraphic_X ), whereP𝑃Pitalic_Pdenotes the ground truth stochastic process over functions mapping from𝒳𝒳\\mathcal{X}caligraphic_Xto𝒴𝒴\\mathcal{Y}caligraphic_Y.\nCommon to allNParchitectures is an encoder and decoder. The encoder maps from𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTand𝐗tsubscript𝐗𝑡\\mathbf{X}_{t}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTto some representation,e⁢(𝒟c,𝐗t)𝑒subscript𝒟𝑐subscript𝐗𝑡e(\\mathcal{D}_{c},\\mathbf{X}_{t})italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ).111In manyNParchitectures, including the originalconditional NP (CNP)andNP, the representation does not depend on the target inputs𝐗tsubscript𝐗𝑡\\mathbf{X}_{t}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT.The decoder takes as input the representation and target inputs𝐗tsubscript𝐗𝑡\\mathbf{X}_{t}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTand outputsd⁢(𝐗t,e⁢(𝒟c,𝐗t))𝑑subscript𝐗𝑡𝑒subscript𝒟𝑐subscript𝐗𝑡d(\\mathbf{X}_{t},e(\\mathcal{D}_{c},\\mathbf{X}_{t}))italic_d ( bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) ), which are the parameters of the predictive distribution over the target outputs𝐘tsubscript𝐘𝑡\\mathbf{Y}_{t}bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT:p⁢(𝐘t|𝐗t,𝒟c)=p⁢(𝐘t|d⁢(𝐗t,e⁢(𝒟c,𝐗t)))𝑝conditionalsubscript𝐘𝑡subscript𝐗𝑡subscript𝒟𝑐𝑝conditionalsubscript𝐘𝑡𝑑subscript𝐗𝑡𝑒subscript𝒟𝑐subscript𝐗𝑡p(\\mathbf{Y}_{t}|\\mathbf{X}_{t},\\mathcal{D}_{c})=p(\\mathbf{Y}_{t}|d(\\mathbf{X}%\n_{t},e(\\mathcal{D}_{c},\\mathbf{X}_{t})))italic_p ( bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) = italic_p ( bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | italic_d ( bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) ) ).\nAn important requirement of the predictive distribution is permutation invariance with respect to the elements of𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT.\nWe shall focus onCNPs(Garnelo et al.,2018a), which factorise the predictive distribution asp⁢(𝐘t|𝐗t,𝒟c)=∏n=1Ntp⁢(𝐲t,n|d⁢(𝐱t,n,e⁢(𝒟c,𝐱t,n)))𝑝conditionalsubscript𝐘𝑡subscript𝐗𝑡subscript𝒟𝑐superscriptsubscriptproduct𝑛1subscript𝑁𝑡𝑝conditionalsubscript𝐲𝑡𝑛𝑑subscript𝐱𝑡𝑛𝑒subscript𝒟𝑐subscript𝐱𝑡𝑛p(\\mathbf{Y}_{t}|\\mathbf{X}_{t},\\mathcal{D}_{c})=\\prod_{n=1}^{N_{t}}p(\\mathbf{%\ny}_{t,n}|d(\\mathbf{x}_{t,n},e(\\mathcal{D}_{c},\\mathbf{x}_{t,n})))italic_p ( bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) = ∏ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPT italic_p ( bold_y start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT | italic_d ( bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT , italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ) ) ).CNPsare trained by maximising the posterior predictive likelihood:ℒML=𝔼p⁢(ξ)⁢[∑n=1Ntlog⁡p⁢(𝐲t,n|d⁢(𝐱t,n,e⁢(𝒟c,𝐱t,n)))].subscriptℒMLsubscript𝔼𝑝𝜉delimited-[]superscriptsubscript𝑛1subscript𝑁𝑡𝑝conditionalsubscript𝐲𝑡𝑛𝑑subscript𝐱𝑡𝑛𝑒subscript𝒟𝑐subscript𝐱𝑡𝑛\\textstyle\\!\\!\\mathcal{L}_{\\text{ML}}\\!=\\!\\mathbb{E}_{p(\\xi)}\\Big{[}\\!\\sum_{n=%\n1}^{N_{t}}\\log p(\\mathbf{y}_{t,n}|d(\\mathbf{x}_{t,n},e(\\mathcal{D}_{c},\\mathbf%\n{x}_{t,n})))\\Big{]}.caligraphic_L start_POSTSUBSCRIPT ML end_POSTSUBSCRIPT = blackboard_E start_POSTSUBSCRIPT italic_p ( italic_ξ ) end_POSTSUBSCRIPT [ ∑ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPT roman_log italic_p ( bold_y start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT | italic_d ( bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT , italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ) ) ) ] .(1)Here, the expectation is taken with respect to the distribution of tasksp⁢(ξ)𝑝𝜉p(\\xi)italic_p ( italic_ξ ). As shown inFoong et al. (2020), the global maximum is achieved if and only if the model recovers the ground-truth posterior prediction map.\nWhen training aCNP, we often approximate the expectation with an average over the finite number of tasks available.\n\n\n2.2TransformersA useful perspective is to understand transformers as a permutation equivariant set functionf𝑓fitalic_f.222Note that not all permutation equivariant set functions can be represented by transformers. For example, the family of informers(Garnelo & Czarnecki,2023)cannot be represented by transformers, yet are permutation equivariant set functions. However, transformers are universal approximators of permutation equivariant set functions(Lee et al.,2019; Wagstaff et al.,2022).They take as input a set ofN𝑁Nitalic_Ntokens,𝐙0∈ℝN×Dzsuperscript𝐙0superscriptℝ𝑁subscript𝐷𝑧\\mathbf{Z}^{0}\\in\\mathbb{R}^{N\\times D_{z}}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N × italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPT, output a set ofN𝑁Nitalic_Ntokens of the same cardinality:f:(ℝDz)N→(ℝDz)N:𝑓→superscriptsuperscriptℝsubscript𝐷𝑧𝑁superscriptsuperscriptℝsubscript𝐷𝑧𝑁f\\colon(\\mathbb{R}^{D_{z}})^{N}\\rightarrow(\\mathbb{R}^{D_{z}})^{N}italic_f : ( blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPT ) start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT → ( blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPT ) start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT. If the input set is permuted, then the output set is permuted accordingly:f⁢(𝐳1,…,𝐳N)n=f⁢(𝐳σ⁢(1),…,𝐳σ⁢(N))σ⁢(n)𝑓subscriptsubscript𝐳1…subscript𝐳𝑁𝑛𝑓subscriptsubscript𝐳𝜎1…subscript𝐳𝜎𝑁𝜎𝑛f(\\mathbf{z}_{1},\\ldots,\\mathbf{z}_{N})_{n}=f(\\mathbf{z}_{\\sigma(1)},\\ldots,%\n\\mathbf{z}_{\\sigma(N)})_{\\sigma(n)}italic_f ( bold_z start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , … , bold_z start_POSTSUBSCRIPT italic_N end_POSTSUBSCRIPT ) start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT = italic_f ( bold_z start_POSTSUBSCRIPT italic_σ ( 1 ) end_POSTSUBSCRIPT , … , bold_z start_POSTSUBSCRIPT italic_σ ( italic_N ) end_POSTSUBSCRIPT ) start_POSTSUBSCRIPT italic_σ ( italic_n ) end_POSTSUBSCRIPTfor all permutationsσ∈𝕊N𝜎superscript𝕊𝑁\\sigma\\in\\mathbb{S}^{N}italic_σ ∈ blackboard_S start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPTofN𝑁Nitalic_Nelements. At the core of each layer of the transformer architecture is themulti-head self attention (MHSA)operation(Vaswani et al.,2017). Let𝐙ℓ∈ℝN×Dzsuperscript𝐙ℓsuperscriptℝ𝑁subscript𝐷𝑧\\mathbf{Z}^{\\ell}\\in\\mathbb{R}^{N\\times D_{z}}bold_Z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N × italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPTdenote the input set to theℓℓ\\ellroman_ℓ-th MHSA operation. The MHSA operation updates then𝑛nitalic_nthtoken𝐳nℓsubscriptsuperscript𝐳ℓ𝑛\\mathbf{z}^{\\ell}_{n}bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPTas𝐳~nℓ=cat⁡({∑m=1Nαhℓ⁢(𝐳nℓ,𝐳mℓ)⁢𝐳mℓT⁢𝐖V,hℓ}h=1Hℓ)⁢𝐖Oℓsubscriptsuperscript~𝐳ℓ𝑛catsuperscriptsubscriptsuperscriptsubscript𝑚1𝑁subscriptsuperscript𝛼ℓℎsubscriptsuperscript𝐳ℓ𝑛subscriptsuperscript𝐳ℓ𝑚superscriptsubscriptsuperscript𝐳ℓ𝑚𝑇subscriptsuperscript𝐖ℓ𝑉ℎℎ1superscript𝐻ℓsubscriptsuperscript𝐖ℓ𝑂\\textstyle\\!\\!\\tilde{\\mathbf{z}}^{\\ell}_{n}\\!=\\!\\operatorname{cat}\\!\\Big{(}%\n\\Big{\\{}\\sum_{m=1}^{N}\\alpha^{\\ell}_{h}(\\mathbf{z}^{\\ell}_{n},\\mathbf{z}^{\\ell%\n}_{m}){\\mathbf{z}^{\\ell}_{m}\\!}^{T}\\mathbf{W}^{\\ell}_{V,h}\\Big{\\}}_{h=1}^{H^{%\n\\ell}}\\Big{)}\\mathbf{W}^{\\ell}_{O}over~ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT = roman_cat ( { ∑ start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT italic_α start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT ( bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_V , italic_h end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_h = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_H start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ) bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_O end_POSTSUBSCRIPT(2)wherecatcat\\operatorname{cat}roman_catdenotes the concatenation operation across the last dimension. Here,𝐖V,hℓ∈ℝDz×DVsubscriptsuperscript𝐖ℓ𝑉ℎsuperscriptℝsubscript𝐷𝑧subscript𝐷𝑉\\mathbf{W}^{\\ell}_{V,h}\\in\\mathbb{R}^{D_{z}\\times D_{V}}bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_V , italic_h end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT end_POSTSUPERSCRIPTand𝐖Oℓ∈ℝHℓ⁢DV×Dzsubscriptsuperscript𝐖ℓ𝑂superscriptℝsuperscript𝐻ℓsubscript𝐷𝑉subscript𝐷𝑧\\mathbf{W}^{\\ell}_{O}\\in\\mathbb{R}^{H^{\\ell}D_{V}\\times D_{z}}bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_O end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_H start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_V end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPTare the value and projection weight matrices, whereHℓsuperscript𝐻ℓH^{\\ell}italic_H start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPTdenotes the number of ‘heads’ in layerℓℓ\\ellroman_ℓ. Note that permutation equivariance is achieved through the permutation invariant summation operator. As this is the only mechanism through which the tokens interact with each other, permutation equivariance for the overall model is ensured. The attention mechanism,αhℓsubscriptsuperscript𝛼ℓℎ\\alpha^{\\ell}_{h}italic_α start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT, is implemented asαhℓ⁢(𝐳nℓ,𝐳mℓ)=e𝐳nℓT⁢𝐖Q,hℓ⁢[𝐖K,hℓ]T⁢𝐳mℓ∑m=1Ne𝐳nℓT⁢𝐖Q,hℓ⁢[𝐖K,hℓ]T⁢𝐳mℓsubscriptsuperscript𝛼ℓℎsubscriptsuperscript𝐳ℓ𝑛subscriptsuperscript𝐳ℓ𝑚superscript𝑒superscriptsubscriptsuperscript𝐳ℓ𝑛𝑇subscriptsuperscript𝐖ℓ𝑄ℎsuperscriptdelimited-[]subscriptsuperscript𝐖ℓ𝐾ℎ𝑇subscriptsuperscript𝐳ℓ𝑚superscriptsubscript𝑚1𝑁superscript𝑒superscriptsubscriptsuperscript𝐳ℓ𝑛𝑇subscriptsuperscript𝐖ℓ𝑄ℎsuperscriptdelimited-[]subscriptsuperscript𝐖ℓ𝐾ℎ𝑇subscriptsuperscript𝐳ℓ𝑚\\alpha^{\\ell}_{h}(\\mathbf{z}^{\\ell}_{n},\\mathbf{z}^{\\ell}_{m})=\\frac{e^{{%\n\\mathbf{z}^{\\ell}_{n}}^{T}\\mathbf{W}^{\\ell}_{Q,h}\\left[\\mathbf{W}^{\\ell}_{K,h}%\n\\right]^{T}\\mathbf{z}^{\\ell}_{m}}}{\\sum_{m=1}^{N}e^{{\\mathbf{z}^{\\ell}_{n}}^{T%\n}\\mathbf{W}^{\\ell}_{Q,h}\\left[\\mathbf{W}^{\\ell}_{K,h}\\right]^{T}\\mathbf{z}^{%\n\\ell}_{m}}}italic_α start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT ( bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) = divide start_ARG italic_e start_POSTSUPERSCRIPT bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_Q , italic_h end_POSTSUBSCRIPT [ bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_K , italic_h end_POSTSUBSCRIPT ] start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT end_POSTSUPERSCRIPT end_ARG start_ARG ∑ start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT italic_e start_POSTSUPERSCRIPT bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_Q , italic_h end_POSTSUBSCRIPT [ bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_K , italic_h end_POSTSUBSCRIPT ] start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT end_POSTSUPERSCRIPT end_ARG(3)where𝐖Q,hℓ∈ℝDz×DQ⁢Ksubscriptsuperscript𝐖ℓ𝑄ℎsuperscriptℝsubscript𝐷𝑧subscript𝐷𝑄𝐾\\mathbf{W}^{\\ell}_{Q,h}\\in\\mathbb{R}^{D_{z}\\times D_{QK}}bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_Q , italic_h end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_Q italic_K end_POSTSUBSCRIPT end_POSTSUPERSCRIPTand𝐖K,hℓ∈ℝDz×DQ⁢Ksubscriptsuperscript𝐖ℓ𝐾ℎsuperscriptℝsubscript𝐷𝑧subscript𝐷𝑄𝐾\\mathbf{W}^{\\ell}_{K,h}\\in\\mathbb{R}^{D_{z}\\times D_{QK}}bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_K , italic_h end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT × italic_D start_POSTSUBSCRIPT italic_Q italic_K end_POSTSUBSCRIPT end_POSTSUPERSCRIPTare the query and key weight matrices. The softmax-normalisation ensures that∑m=1Nαhℓ⁢(𝐳nℓ,𝐳mℓ)=1⁢∀n,h,ℓsuperscriptsubscript𝑚1𝑁subscriptsuperscript𝛼ℓℎsubscriptsuperscript𝐳ℓ𝑛subscriptsuperscript𝐳ℓ𝑚1for-all𝑛ℎℓ\\sum_{m=1}^{N}\\alpha^{\\ell}_{h}(\\mathbf{z}^{\\ell}_{n},\\mathbf{z}^{\\ell}_{m})=1%\n\\ \\forall n,h,\\ell∑ start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT italic_α start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT ( bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) = 1 ∀ italic_n , italic_h , roman_ℓ.\nOften, conditional independencies amongst the set of tokens—in the sense that the set{𝐳nℓ}ℓ=1ℓ=Lsubscriptsuperscriptsubscriptsuperscript𝐳ℓ𝑛ℓ𝐿ℓ1\\{\\mathbf{z}^{\\ell}_{n}\\}^{\\ell=L}_{\\ell=1}{ bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT } start_POSTSUPERSCRIPT roman_ℓ = italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ = 1 end_POSTSUBSCRIPTdo not depend on the set{𝐳mℓ}ℓ=1ℓ=Lsubscriptsuperscriptsubscriptsuperscript𝐳ℓ𝑚ℓ𝐿ℓ1\\{\\mathbf{z}^{\\ell}_{m}\\}^{\\ell=L}_{\\ell=1}{ bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT } start_POSTSUPERSCRIPT roman_ℓ = italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ = 1 end_POSTSUBSCRIPTgiven some other set of tokens for somen,m∈{1,…,N}𝑛𝑚1…𝑁n,\\ m\\in\\{1,\\ldots,N\\}italic_n , italic_m ∈ { 1 , … , italic_N }—are desirable. Whilst this is typically achieved through masking, if the same set of tokens are conditioned on for everyn𝑛nitalic_n, then it is more computationally efficient to usemulti-head cross attention (MHCA)operations together withMHSAoperations than it is to directly computeEquation2. TheMHCAoperation updates then𝑛nitalic_nthtoken𝐳nℓsubscriptsuperscript𝐳ℓ𝑛\\mathbf{z}^{\\ell}_{n}bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPTusing the set of tokens{𝐳^mℓ}m=1Msuperscriptsubscriptsubscriptsuperscript^𝐳ℓ𝑚𝑚1𝑀\\{\\hat{\\mathbf{z}}^{\\ell}_{m}\\}_{m=1}^{M}{ over^ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPTas𝐳~nℓ=cat⁡({∑m=1Mαhℓ⁢(𝐳nℓ,𝐳^mℓ)⁢𝐳^mℓT⁢𝐖V,hℓ}h=1Hℓ)⁢𝐖Oℓ.subscriptsuperscript~𝐳ℓ𝑛catsuperscriptsubscriptsuperscriptsubscript𝑚1𝑀subscriptsuperscript𝛼ℓℎsubscriptsuperscript𝐳ℓ𝑛subscriptsuperscript^𝐳ℓ𝑚superscriptsubscriptsuperscript^𝐳ℓ𝑚𝑇subscriptsuperscript𝐖ℓ𝑉ℎℎ1superscript𝐻ℓsubscriptsuperscript𝐖ℓ𝑂\\textstyle\\!\\!\\tilde{\\mathbf{z}}^{\\ell}_{n}\\!=\\!\\operatorname{cat}\\!\\Big{(}%\n\\Big{\\{}\\sum_{m=1}^{M}\\alpha^{\\ell}_{h}(\\mathbf{z}^{\\ell}_{n},\\hat{\\mathbf{z}}%\n^{\\ell}_{m}){\\hat{\\mathbf{z}}^{\\ell}_{m}\\!}^{T}\\mathbf{W}^{\\ell}_{V,h}\\Big{\\}}%\n_{h=1}^{H^{\\ell}}\\Big{)}\\mathbf{W}^{\\ell}_{O}.over~ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT = roman_cat ( { ∑ start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT italic_α start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_h end_POSTSUBSCRIPT ( bold_z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , over^ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) over^ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_V , italic_h end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_h = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_H start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT end_POSTSUPERSCRIPT ) bold_W start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_O end_POSTSUBSCRIPT .(4)Note that all tokens updated in this manner are conditionally independent of each other given{𝐳^mℓ}m=1Msuperscriptsubscriptsubscriptsuperscript^𝐳ℓ𝑚𝑚1𝑀\\{\\hat{\\mathbf{z}}^{\\ell}_{m}\\}_{m=1}^{M}{ over^ start_ARG bold_z end_ARG start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT } start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT.\nWe discuss this in more detail inAppendixB.MHCAoperations are at the core of the pseudo-token-based transformers such as the perceiver(Jaegle et al.,2021)andinduced set transformer (IST)(Lee et al.,2019). We describe these differences in the following section.MHSAandMHCAoperations are used in combination with layer-normalisation operations and pointwise MLPs to obtainMHSAandMHCAblocks. Unless stated otherwise, we shall adopt the order used byVaswani et al. (2017).\n\n\n2.3Pseudo-Token-Based TransformersPseudo-token based transformers reduce the quadratic computational complexity of the standard transformer through the use of pseudo-tokens. Concretely, let𝐔∈ℝM×Dz𝐔superscriptℝ𝑀subscript𝐷𝑧\\mathbf{U}\\in\\mathbb{R}^{M\\times D_{z}}bold_U ∈ blackboard_R start_POSTSUPERSCRIPT italic_M × italic_D start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT end_POSTSUPERSCRIPTdenote an initial set ofM≪Nmuch-less-than𝑀𝑁M\\ll Nitalic_M ≪ italic_Ntokens we call pseudo-tokens. There are two established methods for incorporating information about the set of observed tokens (𝐙)\\mathbf{Z})bold_Z )into these pseudo-tokens in a computationally efficient manner: the perceiver-style approach ofJaegle et al. (2021)and theISTstyle approach ofLee et al. (2019). The perceiver-style approach iterates between applyingMHCA⁡(𝐔ℓ,𝐙ℓ)MHCAsuperscript𝐔ℓsuperscript𝐙ℓ\\operatorname{MHCA}(\\mathbf{U}^{\\ell},\\mathbf{Z}^{\\ell})roman_MHCA ( bold_U start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT , bold_Z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT )andMHSA⁡(𝐔ℓ)MHSAsuperscript𝐔ℓ\\operatorname{MHSA}(\\mathbf{U}^{\\ell})roman_MHSA ( bold_U start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT ), outputting a set ofM𝑀Mitalic_Mpseudo-tokens, and has a computational complexity of𝒪⁢(M⁢N)𝒪𝑀𝑁\\mathcal{O}\\left(MN\\right)caligraphic_O ( italic_M italic_N )at each layer. TheIST-style approach iterates between applyingMHCA⁡(𝐔ℓ,𝐙ℓ)MHCAsuperscript𝐔ℓsuperscript𝐙ℓ\\operatorname{MHCA}(\\mathbf{U}^{\\ell},\\mathbf{Z}^{\\ell})roman_MHCA ( bold_U start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT , bold_Z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT )andMHCA⁡(𝐙ℓ,𝐔ℓ)MHCAsuperscript𝐙ℓsuperscript𝐔ℓ\\operatorname{MHCA}(\\mathbf{Z}^{\\ell},\\mathbf{U}^{\\ell})roman_MHCA ( bold_Z start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT , bold_U start_POSTSUPERSCRIPT roman_ℓ end_POSTSUPERSCRIPT ), outputting a set ofN𝑁Nitalic_Ntokens andM𝑀Mitalic_Mpseudo-tokens, and also has a computational complexity of𝒪⁢(M⁢N)𝒪𝑀𝑁\\mathcal{O}\\left(MN\\right)caligraphic_O ( italic_M italic_N )at each layer. We provide illustrations these differencesAppendixC.\n\n\n2.4Transformer Neural ProcessesGiven the utility of transformers as set functions, it is natural to consider their use in the encoder of aNP—we describe this family ofNPsasTNPs. Let𝐙c0∈ℝNc×Dsubscriptsuperscript𝐙0𝑐superscriptℝsubscript𝑁𝑐𝐷\\mathbf{Z}^{0}_{c}\\in\\mathbb{R}^{N_{c}\\times D}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT × italic_D end_POSTSUPERSCRIPTdenote the initial set-of-token representation of each input-output pair(𝐱c,n,𝐲c,n)∈𝒟csubscript𝐱𝑐𝑛subscript𝐲𝑐𝑛subscript𝒟𝑐(\\mathbf{x}_{c,n},\\mathbf{y}_{c,n})\\in\\mathcal{D}_{c}( bold_x start_POSTSUBSCRIPT italic_c , italic_n end_POSTSUBSCRIPT , bold_y start_POSTSUBSCRIPT italic_c , italic_n end_POSTSUBSCRIPT ) ∈ caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT, and𝐙t,n0∈ℝNt×Dsubscriptsuperscript𝐙0𝑡𝑛superscriptℝsubscript𝑁𝑡𝐷\\mathbf{Z}^{0}_{t,n}\\in\\mathbb{R}^{N_{t}\\times D}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT × italic_D end_POSTSUPERSCRIPTdenote the initial set-of-token representation of each input𝐱t,n∈𝐗tsubscript𝐱𝑡𝑛subscript𝐗𝑡\\mathbf{x}_{t,n}\\in\\mathbf{X}_{t}bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ∈ bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT. The encodinge⁢(𝒟c,𝐗t)𝑒subscript𝒟𝑐subscript𝐗𝑡e(\\mathcal{D}_{c},\\mathbf{X}_{t})italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT )ofTNPsis is achieved by passing the union of initial context and target tokens,𝐙0=[𝐙c0,𝐙t0]superscript𝐙0subscriptsuperscript𝐙0𝑐subscriptsuperscript𝐙0𝑡\\mathbf{Z}^{0}=[\\mathbf{Z}^{0}_{c},\\ \\mathbf{Z}^{0}_{t}]bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT = [ bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ], through a transformer-style architecture, and keeping only the output tokens corresponding to the target inputs,𝐙tLsubscriptsuperscript𝐙𝐿𝑡\\mathbf{Z}^{L}_{t}bold_Z start_POSTSUPERSCRIPT italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT.The specific transformer-style architecture is unique to eachTNPvariant. However, they generally consist of MHSA operations acting on the context tokens and MHCA operations acting to update the target tokens, given the context tokens.333As discussed inSection2.2, this is often implemented as a single MHSA operation with masking operating.The combination of MHSA and MHCA operations is a permutation invariant function with respect to the context tokens. We provide an illustration of this inFigure1(a). Enforcing these conditional independencies ensures that the final target token𝐳t,nLsubscriptsuperscript𝐳𝐿𝑡𝑛\\mathbf{z}^{L}_{t,n}bold_z start_POSTSUPERSCRIPT italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPTdepends only on𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTand𝐱t,nsubscript𝐱𝑡𝑛\\mathbf{x}_{t,n}bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT, i.e.[e⁢(𝒟c,𝐗t)]n=e⁢(𝒟c,𝐱t,n)subscriptdelimited-[]𝑒subscript𝒟𝑐subscript𝐗𝑡𝑛𝑒subscript𝒟𝑐subscript𝐱𝑡𝑛\\left[e(\\mathcal{D}_{c},\\mathbf{X}_{t})\\right]_{n}=e(\\mathcal{D}_{c},\\mathbf{x%\n}_{t,n})[ italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) ] start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT = italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ). This is required for the factorisation of the predictive distributionp⁢(𝐘t|𝐗t,𝒟c)=∏n=1Ntp⁢(𝐲t,n|d⁢(𝐱t,n,e⁢(𝒟c,𝐱t,n)))𝑝conditionalsubscript𝐘𝑡subscript𝐗𝑡subscript𝒟𝑐superscriptsubscriptproduct𝑛1subscript𝑁𝑡𝑝conditionalsubscript𝐲𝑡𝑛𝑑subscript𝐱𝑡𝑛𝑒subscript𝒟𝑐subscript𝐱𝑡𝑛p(\\mathbf{Y}_{t}|\\mathbf{X}_{t},\\mathcal{D}_{c})=\\prod_{n=1}^{N_{t}}p(\\mathbf{%\ny}_{t,n}|d(\\mathbf{x}_{t,n},e(\\mathcal{D}_{c},\\mathbf{x}_{t,n})))italic_p ( bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) = ∏ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPT italic_p ( bold_y start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT | italic_d ( bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT , italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT ) ) ). We denotepseudo-token TNPs (PT-TNPs)as the family ofTNPswhich use pseudo-token based transformers. Currently, this family is restricted to theLBANP, which uses a perceiver-style architecture; however, it is straightforward to use an IST-style architecture instead.\n\n\n2.5Translation Equivariance𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT𝐗tsubscript𝐗𝑡\\mathbf{X}_{t}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTMLP⁡(𝐗c,𝐘c)MLPsubscript𝐗𝑐subscript𝐘𝑐\\operatorname{MLP}(\\mathbf{X}_{c},\\mathbf{Y}_{c})roman_MLP ( bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )MLP⁡(𝐗t)MLPsubscript𝐗𝑡\\operatorname{MLP}(\\mathbf{X}_{t})roman_MLP ( bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT )𝐙c0subscriptsuperscript𝐙0𝑐\\mathbf{Z}^{0}_{c}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT𝐙t0subscriptsuperscript𝐙0𝑡\\mathbf{Z}^{0}_{t}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTMHSA⁡(𝐙c0)MHSAsubscriptsuperscript𝐙0𝑐\\operatorname{MHSA}(\\mathbf{Z}^{0}_{c})roman_MHSA ( bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )…MHSA⁡(𝐙cL−1)MHSAsubscriptsuperscript𝐙𝐿1𝑐\\operatorname{MHSA}(\\mathbf{Z}^{L-1}_{c})roman_MHSA ( bold_Z start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )MHCA⁡(𝐙t0,𝐙c1)MHCAsubscriptsuperscript𝐙0𝑡subscriptsuperscript𝐙1𝑐\\operatorname{MHCA}(\\mathbf{Z}^{0}_{t},\\mathbf{Z}^{1}_{c})roman_MHCA ( bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_Z start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )…MHCA(𝐙tL−1\\operatorname{MHCA}(\\mathbf{Z}^{L-1}_{t}roman_MHCA ( bold_Z start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT,𝐙cL)\\mathbf{Z}^{L}_{c})bold_Z start_POSTSUPERSCRIPT italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )e⁢(𝒟c,𝐗t)𝑒subscript𝒟𝑐subscript𝐗𝑡e(\\mathcal{D}_{c},\\mathbf{X}_{t})italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT )(a)TNP.te−MHSA⁡(𝐙c0,𝐗c0)teMHSAsubscriptsuperscript𝐙0𝑐subscriptsuperscript𝐗0𝑐\\operatorname{te-MHSA}(\\mathbf{Z}^{0}_{c},\\mathbf{X}^{0}_{c})start_OPFUNCTION roman_te - roman_MHSA end_OPFUNCTION ( bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )…te−MHSA⁡(𝐙cL−1,𝐗cL−1)teMHSAsubscriptsuperscript𝐙𝐿1𝑐subscriptsuperscript𝐗𝐿1𝑐\\operatorname{te-MHSA}(\\mathbf{Z}^{L-1}_{c},\\mathbf{X}^{L-1}_{c})start_OPFUNCTION roman_te - roman_MHSA end_OPFUNCTION ( bold_Z start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )te−MHCA⁡(𝐙t0,𝐙c1,𝐗t0,𝐗c0)teMHCAsubscriptsuperscript𝐙0𝑡subscriptsuperscript𝐙1𝑐subscriptsuperscript𝐗0𝑡subscriptsuperscript𝐗0𝑐\\operatorname{te-MHCA}(\\mathbf{Z}^{0}_{t},\\mathbf{Z}^{1}_{c},\\mathbf{X}^{0}_{t%\n},\\mathbf{X}^{0}_{c})start_OPFUNCTION roman_te - roman_MHCA end_OPFUNCTION ( bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_Z start_POSTSUPERSCRIPT 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )…te−MHCA(𝐙tL−1\\operatorname{te-MHCA}(\\mathbf{Z}^{L-1}_{t}start_OPFUNCTION roman_te - roman_MHCA end_OPFUNCTION ( bold_Z start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT,𝐙cL,𝐗tL−1,𝐗cL−1)\\mathbf{Z}^{L}_{c},\\mathbf{X}^{L-1}_{t},\\mathbf{X}^{L-1}_{c})bold_Z start_POSTSUPERSCRIPT italic_L end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_X start_POSTSUPERSCRIPT italic_L - 1 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )e⁢(𝒟c,𝐗t)𝑒subscript𝒟𝑐subscript𝐗𝑡e(\\mathcal{D}_{c},\\mathbf{X}_{t})italic_e ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT )𝐙c0subscriptsuperscript𝐙0𝑐\\mathbf{Z}^{0}_{c}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT𝐙t0subscriptsuperscript𝐙0𝑡\\mathbf{Z}^{0}_{t}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTMLP⁡(𝐘c)MLPsubscript𝐘𝑐\\operatorname{MLP}(\\mathbf{Y}_{c})roman_MLP ( bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT )𝐘csubscript𝐘𝑐\\mathbf{Y}_{c}bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT𝐗c0=𝐗csubscriptsuperscript𝐗0𝑐subscript𝐗𝑐\\mathbf{X}^{0}_{c}=\\mathbf{X}_{c}bold_X start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT = bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT𝐗t0=𝐗tsubscriptsuperscript𝐗0𝑡subscript𝐗𝑡\\mathbf{X}^{0}_{t}=\\mathbf{X}_{t}bold_X start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT = bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT(b)TE-TNP.Figure 1:Block diagrams illustrating theTNPandTE-TNPencoder architectures. For both models, we pass individual datapoints through pointwise MLPs to obtain the initial token representations,𝐙c0subscriptsuperscript𝐙0𝑐\\mathbf{Z}^{0}_{c}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTand𝐙t0subscriptsuperscript𝐙0𝑡\\mathbf{Z}^{0}_{t}bold_Z start_POSTSUPERSCRIPT 0 end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT. These are then passed through multiple attention layers, with the context tokens interacting with the target tokens through cross-attention. The output of the encoder depends on𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTand𝐗tsubscript𝐗𝑡\\mathbf{X}_{t}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT. TheTE-TNPencoder updates the input locations at each layer, in addition to the tokens.Figure 2:Average log-likelihood (↑bold-↑\\boldsymbol{\\mathbf{\\uparrow}}bold_↑) on the test datasets for the synthetic 1-D regression experiment.ΔΔ\\Deltaroman_Δdenotes the amount by which the range from which the context and target inputs and sampled from is shifted at test time. Standard errors are shown.Here, we provide new theoretical results which show the importance of translation equivariance as an inductive bias inNPs. In particular, we first show thatif, and only if,the ground-truth stochastic process is stationary, the corresponding predictive map is translation equivariant (Theorem2.1). Second, we show the importance of translation equivariance in the ability of our models to generalise to settings outside of the training distribution (Theorem2.2), for whichFigure3provides some intuition.Let𝖳𝝉subscript𝖳𝝉\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPTdenote a translation by𝝉∈ℝDx𝝉superscriptℝsubscript𝐷𝑥\\boldsymbol{\\mathbf{\\tau}}\\in\\mathbb{R}^{D_{x}}bold_italic_τ ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT.\nFor a data set𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_S,𝖳𝝉⁢𝒟∈𝒮subscript𝖳𝝉𝒟𝒮\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}\\mathcal{D}\\in\\mathcal{S}sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT caligraphic_D ∈ caligraphic_Stranslates the data set by adding𝝉𝝉\\boldsymbol{\\mathbf{\\tau}}bold_italic_τto all inputs.\nFor a functionf:𝒳→Z:𝑓→𝒳𝑍f\\colon\\mathcal{X}\\to Zitalic_f : caligraphic_X → italic_Z,𝖳𝝉⁢fsubscript𝖳𝝉𝑓\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}fsansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT italic_ftranslatesf𝑓fitalic_fby producing a new function𝒳→Z→𝒳𝑍\\mathcal{X}\\to Zcaligraphic_X → italic_Zsuch that𝖳𝝉⁢f⁢(𝐱)=f⁢(𝐱−𝝉)subscript𝖳𝝉𝑓𝐱𝑓𝐱𝝉\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}f(\\mathbf{x})=f(\\mathbf{x}-\\boldsymbol{%\n\\mathbf{\\tau}})sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT italic_f ( bold_x ) = italic_f ( bold_x - bold_italic_τ )for all𝐱∈ℝDx𝐱superscriptℝsubscript𝐷𝑥\\mathbf{x}\\in\\mathbb{R}^{D_{x}}bold_x ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT.\nFor a stochastic processμ∈𝒫⁢(𝒳)𝜇𝒫𝒳\\mu\\in\\mathcal{P}(\\mathcal{X})italic_μ ∈ caligraphic_P ( caligraphic_X ),𝖳𝝉⁢(μ)subscript𝖳𝝉𝜇\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}(\\mu)sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT ( italic_μ )denotes the pushforward measure of pushingμ𝜇\\muitalic_μthrough𝖳𝝉subscript𝖳𝝉\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT.\nAprediction mapπ𝜋\\piitalic_πis a mappingπ:𝒮→𝒫⁢(𝒳):𝜋→𝒮𝒫𝒳\\pi\\colon\\mathcal{S}\\to\\mathcal{P}(\\mathcal{X})italic_π : caligraphic_S → caligraphic_P ( caligraphic_X )from data sets𝒮𝒮\\mathcal{S}caligraphic_Sto stochastic processes𝒫⁢(𝒳)𝒫𝒳\\mathcal{P}(\\mathcal{X})caligraphic_P ( caligraphic_X ).\nPrediction maps are mathematical models of neural processes.\nSay thata prediction mapπ𝜋\\piitalic_πis translation equivariantif𝖳𝝉∘π=π∘𝖳𝝉subscript𝖳𝝉𝜋𝜋subscript𝖳𝝉\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}\\circ\\pi=\\pi\\circ\\mathsf{T}_{%\n\\boldsymbol{\\mathbf{\\tau}}}sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT ∘ italic_π = italic_π ∘ sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPTfor all translations𝝉∈ℝDx𝝉superscriptℝsubscript𝐷𝑥\\boldsymbol{\\mathbf{\\tau}}\\in\\mathbb{R}^{D_{x}}bold_italic_τ ∈ blackboard_R start_POSTSUPERSCRIPT italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT end_POSTSUPERSCRIPT.The ground-truth stochastic processP𝑃Pitalic_Pis stationary if and only if the prediction mapπPsubscript𝜋𝑃\\pi_{P}italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTis translation equivariant.Foong et al. (2020)provide a simple proof of the “only if”-direction.\nWe provide a rigorous proof in both directions.\nConsider𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_S.\nFormally defineπP⁢(𝒟)subscript𝜋𝑃𝒟\\pi_{P}(\\mathcal{D})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D )by integratingP𝑃Pitalic_Pagainst a densityπP′⁢(𝒟)superscriptsubscript𝜋𝑃′𝒟\\pi_{P}^{\\prime}(\\mathcal{D})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D )that depends on𝒟𝒟\\mathcal{D}caligraphic_D, sod⁢πP⁢(𝒟)=πP′⁢(𝒟)⁢d⁢Pdsubscript𝜋𝑃𝒟superscriptsubscript𝜋𝑃′𝒟d𝑃\\mathrm{d}\\pi_{P}(\\mathcal{D})=\\pi_{P}^{\\prime}(\\mathcal{D})\\,\\mathrm{d}Proman_d italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ) = italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D ) roman_d italic_P.444Intuitively,πP′⁢(𝒟)⁢(f)=p⁢(𝒟|f)/p⁢(𝒟)superscriptsubscript𝜋𝑃′𝒟𝑓𝑝conditional𝒟𝑓𝑝𝒟\\pi_{P}^{\\prime}(\\mathcal{D})(f)=p(\\mathcal{D}|f)/p(\\mathcal{D})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D ) ( italic_f ) = italic_p ( caligraphic_D | italic_f ) / italic_p ( caligraphic_D ), soπP′⁢(𝒟)⁢(f)superscriptsubscript𝜋𝑃′𝒟𝑓\\pi_{P}^{\\prime}(\\mathcal{D})(f)italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D ) ( italic_f )is the modelling assumption that specifies how observations are generated from the ground-truth stochastic process.\nA simple example isπP′⁢(𝒟)⁢(f)∝∏(𝐱,𝐲)∈𝒟𝒩⁢(𝐲∣f⁢(𝐱),σ2)proportional-tosuperscriptsubscript𝜋𝑃′𝒟𝑓subscriptproduct𝐱𝐲𝒟𝒩conditional𝐲𝑓𝐱superscript𝜎2\\pi_{P}^{\\prime}(\\mathcal{D})(f)\\propto\\prod_{(\\mathbf{x},\\mathbf{y})\\in%\n\\mathcal{D}}\\mathcal{N}(\\mathbf{y}\\mid f(\\mathbf{x}),\\sigma^{2})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D ) ( italic_f ) ∝ ∏ start_POSTSUBSCRIPT ( bold_x , bold_y ) ∈ caligraphic_D end_POSTSUBSCRIPT caligraphic_N ( bold_y ∣ italic_f ( bold_x ) , italic_σ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ), which adds independent Gaussian noise with varianceσ2superscript𝜎2\\sigma^{2}italic_σ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT.Assume thatπP′⁢(∅)∝1proportional-tosubscriptsuperscript𝜋′𝑃1\\pi^{\\prime}_{P}(\\varnothing)\\propto 1italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( ∅ ) ∝ 1, soπP⁢(∅)=Psubscript𝜋𝑃𝑃\\pi_{P}(\\varnothing)=Pitalic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( ∅ ) = italic_P.\nSay thatπP′superscriptsubscript𝜋𝑃′\\pi_{P}^{\\prime}italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPTistranslation invariantif, for all𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_Sand𝝉∈𝒳𝝉𝒳\\boldsymbol{\\mathbf{\\tau}}\\in\\mathcal{X}bold_italic_τ ∈ caligraphic_X,πP′⁢(𝖳𝝉⁢𝒟)∘𝖳𝝉=πP′⁢(𝒟)superscriptsubscript𝜋𝑃′subscript𝖳𝝉𝒟subscript𝖳𝝉superscriptsubscript𝜋𝑃′𝒟\\pi_{P}^{\\prime}(\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}\\mathcal{D})\\circ%\n\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}=\\pi_{P}^{\\prime}(\\mathcal{D})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT caligraphic_D ) ∘ sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT = italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ( caligraphic_D )P𝑃Pitalic_P–almost surely.555For example, the usual Gaussian likelihood is translation invariant:𝒩⁢(𝐲∣(𝖳𝝉⁢f)⁢(𝐱+𝝉),σ2)=𝒩⁢(𝐲∣f⁢(𝝉),σ2).𝒩conditional𝐲subscript𝖳𝝉𝑓𝐱𝝉superscript𝜎2𝒩conditional𝐲𝑓𝝉superscript𝜎2\\mathcal{N}(\\mathbf{y}\\mid(\\mathsf{T}_{\\boldsymbol{\\mathbf{\\tau}}}f)(\\mathbf{x%\n}+\\boldsymbol{\\mathbf{\\tau}}),\\sigma^{2})=\\mathcal{N}(\\mathbf{y}\\mid f(%\n\\boldsymbol{\\mathbf{\\tau}}),\\sigma^{2}).caligraphic_N ( bold_y ∣ ( sansserif_T start_POSTSUBSCRIPT bold_italic_τ end_POSTSUBSCRIPT italic_f ) ( bold_x + bold_italic_τ ) , italic_σ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ) = caligraphic_N ( bold_y ∣ italic_f ( bold_italic_τ ) , italic_σ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT ) .Theorem 2.1.(1) The ground-truth stochastic processP𝑃Pitalic_Pis stationary andπP′subscriptsuperscript𝜋′𝑃\\pi^{\\prime}_{P}italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTis translation invariant if and only if\n(2)πPsubscript𝜋𝑃\\pi_{P}italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTis translation equivariant.SeeAppendixEfor the proof.\nIf the ground-truth stochastic process is stationary, it is helpful to build translation equivariance into the neural process:\nthis greatly reduces the model space to search over, which can significantly improve data efficiency(Foong et al.,2020).\nIn addition, it is possible to show that translation equivariantNPsgeneralise spatially.\nWe formalise this in the following theorem, which we present in the one-dimensional setting (Dx=1subscript𝐷𝑥1D_{x}=1italic_D start_POSTSUBSCRIPT italic_x end_POSTSUBSCRIPT = 1) for notational simplicity, and we provide an illustration of these ideas inFigure3.Definitions for theorem.For a stochastic processf∼μsimilar-to𝑓𝜇f\\sim\\muitalic_f ∼ italic_μwithμ∈𝒫⁢(𝒳)𝜇𝒫𝒳\\mu\\in\\mathcal{P}(\\mathcal{X})italic_μ ∈ caligraphic_P ( caligraphic_X ), for every𝐱∈ℝN𝐱superscriptℝ𝑁\\mathbf{x}\\in\\mathbb{R}^{N}bold_x ∈ blackboard_R start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT, denote the distribution of(f⁢(x1),…,f⁢(xN))𝑓subscript𝑥1…𝑓subscript𝑥𝑁(f(x_{1}),\\ldots,f(x_{N}))( italic_f ( italic_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ) , … , italic_f ( italic_x start_POSTSUBSCRIPT italic_N end_POSTSUBSCRIPT ) )byP𝐱⁢μsubscript𝑃𝐱𝜇P_{\\mathbf{x}}\\muitalic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_μ.\nWe now define the notion of thereceptive field.\nFor two vectors of inputs𝐱1∈ℝN1subscript𝐱1superscriptℝsubscript𝑁1\\mathbf{x}_{1}\\in\\mathbb{R}^{N_{1}}bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝐱2∈ℝN2subscript𝐱2superscriptℝsubscript𝑁2\\mathbf{x}_{2}\\in\\mathbb{R}^{N_{2}}bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT end_POSTSUPERSCRIPT, andR>0𝑅0R>0italic_R > 0, let𝐱1|𝐱2,Revaluated-atsubscript𝐱1subscript𝐱2𝑅\\mathbf{x}_{1}|_{\\mathbf{x}_{2},R}bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT | start_POSTSUBSCRIPT bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , italic_R end_POSTSUBSCRIPTbe the subvector of𝐱1subscript𝐱1\\mathbf{x}_{1}bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPTwith inputs at most distanceR𝑅Ritalic_Raway from any input in𝐱2subscript𝐱2\\mathbf{x}_{2}bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT.\nSimilarly, for a data set𝒟=(𝐱,𝐲)∈𝒮𝒟𝐱𝐲𝒮\\mathcal{D}=(\\mathbf{x},\\mathbf{y})\\in\\mathcal{S}caligraphic_D = ( bold_x , bold_y ) ∈ caligraphic_S, let𝒟|𝐱2,R∈𝒮evaluated-at𝒟subscript𝐱2𝑅𝒮\\mathcal{D}|_{\\mathbf{x}_{2},R}\\in\\mathcal{S}caligraphic_D | start_POSTSUBSCRIPT bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , italic_R end_POSTSUBSCRIPT ∈ caligraphic_Sbe the subset of data points of𝒟𝒟\\mathcal{D}caligraphic_Dwith inputs at most distanceR𝑅Ritalic_Raway from any input in𝐱2subscript𝐱2\\mathbf{x}_{2}bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT.\nWith these definitions, say that astochastic processf∼μsimilar-to𝑓𝜇f\\sim\\muitalic_f ∼ italic_μwithμ∈𝒫⁢(𝒳)𝜇𝒫𝒳\\mu\\in\\mathcal{P}(\\mathcal{X})italic_μ ∈ caligraphic_P ( caligraphic_X )has receptive fieldR>0𝑅0R>0italic_R > 0if, for allN1,N2∈ℕsubscript𝑁1subscript𝑁2ℕN_{1},N_{2}\\in\\mathbb{N}italic_N start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_N start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ∈ blackboard_N,𝐱1∈ℝN1subscript𝐱1superscriptℝsubscript𝑁1\\mathbf{x}_{1}\\in\\mathbb{R}^{N_{1}}bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT end_POSTSUPERSCRIPT, and𝐱2∈ℝN2subscript𝐱2superscriptℝsubscript𝑁2\\mathbf{x}_{2}\\in\\mathbb{R}^{N_{2}}bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,f(𝐱2)∣f(𝐱1)=df(𝐱2)∣f(𝐱1|𝐱2,12⁢Rf(\\mathbf{x}_{2})\\mid f(\\mathbf{x}_{1})\\overset{\\text{d}}{=}f(\\mathbf{x}_{2})%\n\\mid f(\\mathbf{x}_{1}|_{\\mathbf{x}_{2},\\frac{1}{2}R}italic_f ( bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) ∣ italic_f ( bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ) overd start_ARG = end_ARG italic_f ( bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ) ∣ italic_f ( bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT | start_POSTSUBSCRIPT bold_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , divide start_ARG 1 end_ARG start_ARG 2 end_ARG italic_R end_POSTSUBSCRIPT).\nIntuitively,f𝑓fitalic_fonly has local dependencies.\nMoreover, say that aprediction mapπ:𝒮→𝒫⁢(𝒳):𝜋→𝒮𝒫𝒳\\pi\\colon\\mathcal{S}\\to\\mathcal{P}(\\mathcal{X})italic_π : caligraphic_S → caligraphic_P ( caligraphic_X )has receptive fieldR>0𝑅0R>0italic_R > 0if, for all𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_S,N∈ℕ𝑁ℕN\\in\\mathbb{N}italic_N ∈ blackboard_N, and𝐱∈ℝN𝐱superscriptℝ𝑁\\mathbf{x}\\in\\mathbb{R}^{N}bold_x ∈ blackboard_R start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT,P𝐱⁢π⁢(𝒟)=P𝐱⁢π⁢(𝒟|𝐱,12⁢R)subscript𝑃𝐱𝜋𝒟subscript𝑃𝐱𝜋evaluated-at𝒟𝐱12𝑅P_{\\mathbf{x}}\\pi(\\mathcal{D})=P_{\\mathbf{x}}\\pi(\\mathcal{D}|_{\\mathbf{x},%\n\\frac{1}{2}R})italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π ( caligraphic_D ) = italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π ( caligraphic_D | start_POSTSUBSCRIPT bold_x , divide start_ARG 1 end_ARG start_ARG 2 end_ARG italic_R end_POSTSUBSCRIPT ).\nIntuitively, predictions by the neural processπ𝜋\\piitalic_πare only influenced by context points at most distance12⁢R12𝑅\\frac{1}{2}Rdivide start_ARG 1 end_ARG start_ARG 2 end_ARG italic_Raway.666For example, iff𝑓fitalic_fis a Gaussian process with a kernel compactly supported on[−12⁢R,12⁢R]12𝑅12𝑅[-\\frac{1}{2}R,\\frac{1}{2}R][ - divide start_ARG 1 end_ARG start_ARG 2 end_ARG italic_R , divide start_ARG 1 end_ARG start_ARG 2 end_ARG italic_R ], then the mapping𝒟↦p⁢(f∣𝒟)maps-to𝒟𝑝conditional𝑓𝒟\\mathcal{D}\\mapsto p(f\\mid\\mathcal{D})caligraphic_D ↦ italic_p ( italic_f ∣ caligraphic_D )is a prediction map which (a) has receptive fieldR𝑅Ritalic_Rand (b) maps to stochastic processes with receptive fieldR𝑅Ritalic_R.Theorem 2.2.Letπ1,π2:𝒮→𝒫⁢(𝒳):subscript𝜋1subscript𝜋2→𝒮𝒫𝒳\\pi_{1},\\pi_{2}\\colon\\mathcal{S}\\rightarrow\\mathcal{P}(\\mathcal{X})italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_π start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT : caligraphic_S → caligraphic_P ( caligraphic_X )be translation equivariant prediction maps with receptive fieldR>0𝑅0R>0italic_R > 0. Assume that, for all𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_S,π1⁢(𝒟)subscript𝜋1𝒟\\pi_{1}(\\mathcal{D})italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( caligraphic_D )andπ2⁢(𝒟)subscript𝜋2𝒟\\pi_{2}(\\mathcal{D})italic_π start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ( caligraphic_D )also have receptive fieldR>0𝑅0R>0italic_R > 0. Letϵ>0italic-ϵ0\\epsilon>0italic_ϵ > 0and fixN∈ℕ𝑁ℕN\\in\\mathbb{N}italic_N ∈ blackboard_N. Assume that, for all𝐱∈⋃n=1N[0,2⁢R]n𝐱superscriptsubscript𝑛1𝑁superscript02𝑅𝑛\\mathbf{x}\\in\\bigcup_{n=1}^{N}[0,2R]^{n}bold_x ∈ ⋃ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT [ 0 , 2 italic_R ] start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPTand𝒟∈𝒮∩⋃n=0∞([0,2⁢R]×ℝ)n𝒟𝒮superscriptsubscript𝑛0superscript02𝑅ℝ𝑛\\mathcal{D}\\in\\mathcal{S}\\cap\\bigcup_{n=0}^{\\infty}\\left([0,2R]\\times\\mathbb{R%\n}\\right)^{n}caligraphic_D ∈ caligraphic_S ∩ ⋃ start_POSTSUBSCRIPT italic_n = 0 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ∞ end_POSTSUPERSCRIPT ( [ 0 , 2 italic_R ] × blackboard_R ) start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT,KL[P𝐱π1(𝒟)||P𝐱π2(𝒟))]≤ϵ.\\mathrm{KL}\\left[{P_{\\mathbf{x}}\\pi_{1}(\\mathcal{D})}||{P_{\\mathbf{x}}\\pi_{2}(%\n\\mathcal{D}))}\\right]\\leq\\epsilon.roman_KL [ italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( caligraphic_D ) | | italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ( caligraphic_D ) ) ] ≤ italic_ϵ .(5)Then, for allM>0𝑀0M>0italic_M > 0,𝐱∈⋃n=1N[0,M]n𝐱superscriptsubscript𝑛1𝑁superscript0𝑀𝑛\\mathbf{x}\\in\\bigcup_{n=1}^{N}[0,M]^{n}bold_x ∈ ⋃ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPT [ 0 , italic_M ] start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT, and𝒟∈𝒮∩⋃n=0∞([0,M]×ℝ)n𝒟𝒮superscriptsubscript𝑛0superscript0𝑀ℝ𝑛\\mathcal{D}\\in\\mathcal{S}\\cap\\bigcup_{n=0}^{\\infty}\\left([0,M]\\times\\mathbb{R}%\n\\right)^{n}caligraphic_D ∈ caligraphic_S ∩ ⋃ start_POSTSUBSCRIPT italic_n = 0 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ∞ end_POSTSUPERSCRIPT ( [ 0 , italic_M ] × blackboard_R ) start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT,KL[P𝐱π1(𝒟)||P𝐱π2(𝒟))]≤⌈2M/R⌉ϵ.\\mathrm{KL}\\left[{P_{\\mathbf{x}}\\pi_{1}(\\mathcal{D})}||{P_{\\mathbf{x}}\\pi_{2}(%\n\\mathcal{D}))}\\right]\\leq\\lceil 2M/R\\rceil\\epsilon.roman_KL [ italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ( caligraphic_D ) | | italic_P start_POSTSUBSCRIPT bold_x end_POSTSUBSCRIPT italic_π start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT ( caligraphic_D ) ) ] ≤ ⌈ 2 italic_M / italic_R ⌉ italic_ϵ .(6)xcsubscript𝑥𝑐x_{c}italic_x start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTxtsubscript𝑥𝑡x_{t}italic_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTR𝑅Ritalic_RR𝑅Ritalic_R(a)For a model with receptive fieldR>0𝑅0R>0italic_R > 0,\na context point atxcsubscript𝑥𝑐x_{c}italic_x start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTinfluences predictions at target inputs only limitedly far away.\nConversely, a prediction at a target inputxtsubscript𝑥𝑡x_{t}italic_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTis influenced by context points only limitedly far away.xcsubscript𝑥𝑐x_{c}italic_x start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPTxtsubscript𝑥𝑡x_{t}italic_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTtraining rangeR𝑅Ritalic_RR𝑅Ritalic_RR𝑅Ritalic_RTE(b)If a model is translation equivariant,\nthen all context points and targets inputs can simultaneously be shifted left or right\nwithout changing the output of the model.\nIntuitively, this means that triangles in the figures can just be “shifted left or right”.Figure 3:Translation equivariance in combination with a limited receptive field (see (a)) can help generalisation performance.\nConsider a translation equivariant (TE) model which performs well within atraining range(see (b)).\nConsider a prediction for a target input outside the training range (right triangle in (b)).\nIf the model has receptive fieldR>0𝑅0R>0italic_R > 0and the training range is bigger thanR𝑅Ritalic_R, then TE can be used to “shift that prediction back into the training range” (see (b)).\nSince the model performs well within the training range, the model also performs well for the target input outside the training range.SeeAppendixEfor the proof.\nThe notion of receptive field is natural to CNNs and corresponds to the usual notion of the receptive field.\nThe notion is also inherent to transformers that adopt sliding window attention: the size of the window multiplied by the number of transformer layers gives the receptive field of the model.\nIntuitively, this theorem states that, if (a) the ground-truth stochastic process and our model are translation equivariant and (b) everything has receptive field sizeR>0𝑅0R>0italic_R > 0, then, whenever our model is accurate on[0,2⁢R]02𝑅[0,2R][ 0 , 2 italic_R ], it is also accurate on any bigger interval[0,M]0𝑀[0,M][ 0 , italic_M ].\nNote that this theorem accounts for dependencies between target points, so it also applies to latent-variable neural processes(Garnelo et al.,2018b)and Gaussian neural processes(Bruinsma et al.,2021; Markou et al.,2022).\nIn practice, we do not explicitly equip our transformer neural processes with a fixed receptive field size by using sliding window attention, although this would certainly be possible.\nThe main purpose of this theorem is to elucidate the underlying principles that enable translation equivariant neural processes to generalise.\n\n",
    "origin_citations": [
      "Conditional neural processes.",
      "Neural processes.",
      "Meta-learning stationary stochastic process prediction with convolutional neural processes.",
      "Exploring the space of key-value-query models with intention.",
      "Set transformer: A framework for attention-based permutation-invariant neural networks.",
      "Universal approximation of functions on sets.",
      "Attention is all you need.",
      "Perceiver: General perception with iterative attention.",
      "The Gaussian neural process.",
      "Practical conditional neural processes via tractable dependent predictions."
    ],
    "citations": [
      5945,
      7224,
      202311,
      261706,
      444051,
      372680,
      38,
      4862,
      39077,
      502950
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 88,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 92,
      "Final_Score": 88.0,
      "raw_text": "Content_Coherence: 88\nCitation_Quality: 85\nSynthesis_Analysis: 92\nFinal_Score: 88"
    },
    "overall_score": 87.4,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "pftXzp6Yn3",
        "forum": "pftXzp6Yn3",
        "number": 5681,
        "title": "Translation Equivariant Transformer Neural Processes",
        "normalized_title": "translation equivariant transformer neural processes",
        "abstract": "The effectiveness of neural processes (NPs) in modelling posterior prediction maps---the mapping from data to posterior predictive distributions---has significantly improved since their inception. This improvement can be attributed to two principal factors: (1) advancements in the architecture of permutation invariant set functions, which are intrinsic to all NPs; and (2) leveraging symmetries present in the true posterior predictive map, which are problem dependent. Transformers are a notable development in permutation invariant set functions, and their utility within NPs has been demonstrated through the family of models we refer to as TNPs. Despite significant interest in TNPs, little attention has been given to incorporating symmetries. Notably, the posterior prediction maps for data that are stationary---a common assumption in spatio-temporal modelling---exhibit translation equivariance. In this paper, we introduce of a new family of translation equivariant TNPs that incorporate *translation equivariance*. Through an extensive range of experiments on synthetic and real-world spatio-temporal data, we demonstrate the effectiveness of TE-TNPs relative to their non-translation-equivariant counterparts and other NP baselines.",
        "authors": [
          "Matthew Ashman",
          "Cristiana Diaconu",
          "Junhyuck Kim",
          "Lakee Sivaraya",
          "Stratis Markou",
          "James Requeima",
          "Wessel P Bruinsma",
          "Richard E. Turner"
        ],
        "decision": null,
        "pdate": 1714610395996,
        "cdate": 1706800702995
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "pftXzp6Yn3",
          "forum": "pftXzp6Yn3",
          "number": 5681,
          "title": "Translation Equivariant Transformer Neural Processes",
          "normalized_title": "translation equivariant transformer neural processes",
          "abstract": "The effectiveness of neural processes (NPs) in modelling posterior prediction maps---the mapping from data to posterior predictive distributions---has significantly improved since their inception. This improvement can be attributed to two principal factors: (1) advancements in the architecture of permutation invariant set functions, which are intrinsic to all NPs; and (2) leveraging symmetries present in the true posterior predictive map, which are problem dependent. Transformers are a notable development in permutation invariant set functions, and their utility within NPs has been demonstrated through the family of models we refer to as TNPs. Despite significant interest in TNPs, little attention has been given to incorporating symmetries. Notably, the posterior prediction maps for data that are stationary---a common assumption in spatio-temporal modelling---exhibit translation equivariance. In this paper, we introduce of a new family of translation equivariant TNPs that incorporate *translation equivariance*. Through an extensive range of experiments on synthetic and real-world spatio-temporal data, we demonstrate the effectiveness of TE-TNPs relative to their non-translation-equivariant counterparts and other NP baselines.",
          "authors": [
            "Matthew Ashman",
            "Cristiana Diaconu",
            "Junhyuck Kim",
            "Lakee Sivaraya",
            "Stratis Markou",
            "James Requeima",
            "Wessel P Bruinsma",
            "Richard E. Turner"
          ],
          "decision": null,
          "pdate": 1714610395996,
          "cdate": 1706800702995
        }
      ]
    }
  },
  {
    "paper_id": 29546,
    "title": "Sparse High Rank Adapters",
    "abstract": "Low Rank Adaptation (LoRA) has gained massive attention in the recent generative AI research. One of the main advantages of LoRA is its ability to be fused with pretrained models, adding no overhead during inference. However, from a mobile deployment standpoint, we can either avoid inference overhead in the fused mode but lose the ability to switch adapters rapidly, or suffer significant (up to 30% higher) inference latency while enabling rapid switching in the unfused mode. LoRA also exhibits concept-loss when multiple adapters are used concurrently. In this paper, we propose Sparse High Rank Adapters (SHiRA), a new paradigm which incurs no inference overhead, enables rapid switching, and significantly reduces concept-loss. Specifically, SHiRA can be trained by directly tuning only1111-2%percent22\\%2 %of the base model weights while leaving others unchanged. This results in a highly sparse adapter which can be switched directly in the fused mode. We further provide theoretical and empirical insights on how high sparsity in SHiRA can aid multi-adapter fusion by reducing concept loss. Our extensive experiments on LVMs and LLMs demonstrate that finetuning only a small fraction of the parameters in the base model significantly outperforms LoRA while enabling both rapid switching and multi-adapter fusion. Finally, we provide a latency- and memory-efficient SHiRA implementation based on Parameter-Efficient Finetuning (PEFT) Library which trains at nearly the same speed as LoRA while consuming up to16%percent1616\\%16 %lower peak GPU memory, thus making SHiRA easy to adopt for practical use cases. To demonstrate rapid switching benefits during inference, we show that loading SHiRA on a base model can be5×5\\times5 ×-16×16\\times16 ×faster than LoRA fusion on a CPU.¶",
    "introduction": "Low Rank Adaptation (LoRA)[13]is an established technique to tune the behavior of large generative models such as Large Language Models (LLMs)[30,29]and Stable Diffusion[24,22]. As the name suggests, LoRA requires very few parameters since it trains low rank projection weights that consume very low memory during the finetuning process while producing excellent results. Moreover, these low rank weights can be fused analytically into the base model, thereby incurring no additional overhead during inference.\n\n\nDespite its success, there are still several limitations of low rank adaptation methods. First, if LoRA parameters are fused into the corresponding pretrained base model weights, they modify the entire weight tensor. Therefore, deploying LoRA on large models such as LLaMA-1/2 (7B+ parameters) or Stable Diffusion (1.5B+ parameters) on mobile devices would require changing a large number of weights during inference. Consequently, for mobile scenarios, if an application requiresrapid adapter switching, existing low rank methods would incur a significant memory and latency cost. This is a major deployment challenge because, unlike large GPUs, local memory of small AI accelerators is limited and cannot store all weights at the same time. These challenges can be partially addressed by running LoRA in unfused mode; however, unfused inference can incur as high as𝟑𝟎%percent30\\bm{30\\%}bold_30 bold_%additional latencycompared to the base model[1](see section2.1for details). This increased inference time in unfused mode and time for adapter switching significantly hampers user experience; hence, this is an important problem which has been a focus of recent research by various industries[9].\nSecond, LoRA has a well-known limitation calledconcept losswhen using multiple concurrent adapters, e.g., combining multiple style transfer adapters, etc. Specifically, it has been well documented[34,26,8]that a simple additive merging of multiple LoRA adapters leads to concept loss of one or more adapters. Finally, recent literature also contributes important theoretical and empirical knowledge towards the value ofhigh rank adapters. For instance, Kalajdzievski[16]shows that the high rank adapters can greatly outperform low rank adapters when used with correct scaling factors. This calls for further investigation into whether other high rank adapters would significantly outperform LoRA.\n\n\nFigure 1:SparseHighRankAdapters (SHiRA): Changing about1111-2%percent22\\%2 %weights of the pretrained generative model is often sufficient to achieve high performance. Due to its extreme sparsity, SHiRA enables rapid switching and also reduced concept loss during multi-adapter fusion. In contrast, LoRA modifies majority of parameters when fused, thus prohibiting rapid switching on mobile devices, and also experiences concept loss during multi-adapter fusion. For LoRA, elephant for single “paintings” adapter case has artifacts (extra/broken tusks); bird and knight for multi-adapter case lose “paintings” concept and keep only the “blue fire” effects. SHiRA does not experience these issues.\n\n\nIn view of the above, we address the followingkey problemsin this paper: (i) How can we perform rapid switching for fused adapters? (ii) Is there a simpler solution for multi-adapter fusion to reduce concept loss? (iii) Can we build high rank adapters that have high expressive power without significantly increasing the training or inference costs?\n\n\nTo this end, we proposeSparseHighRankAdapters (SHiRA), a single solution to all three problems above. SHiRA is a highly sparse but a high rank adapter which relies on training only a very small subset of parameters from the original pretrained network. One of the crucial insights we demonstrate is that even finetuning merely1111-2%percent22\\%2 %parameters of the pretrained generative model is sufficient to achieve high performance on many adapter tasks (see Fig.1).\nHowever, unlike LoRA layers that modify all parameters in the weight tensors in the fused mode, SHiRA still keeps a very low percentage of parameters that need to be switched, thus enabling rapid switching at inference time. Moreover, since the pretrained weights are huge, SHiRA being a very sparse adapter greatly aids multi-adapter fusion by significantly reducing concept loss. Finally, we theoretically and emprically analyze the high rank vs. sparsity properties of SHiRA and why that helps with adapter performance.\n\n\nOverall, we make the followingkey contributions:•We propose SHiRA, a new high rank adapter paradigm to demonstrate that changing as few as1111-2%percent22\\%2 %parameters of the original network is sufficient for adaptation. Our crucial insight is that even the most basic masking criteria (to identify the top1111-2%percent22\\%2 %parameters) enable SHiRA to significantly outperform LoRA on diverse vision and language tasks.•SHiRA enables on-devicerapid adapter switchingand provides a natural multi-adapter fusion technique due to high sparsity, thus, significantly reducingconcept loss. We also theoretically analyze SHiRA through the lens ofhigh rank adaptationvs. sparsity.•We conduct extensive experiments on LLMs (LLaMA-7B, LLaMAv2-7B) and LVMs (Stable Diffusion, SDXL) where we demonstrate that SHiRA significantly outperforms LoRA on both single- and multi-adapter tasks. On LLMs, we show that SHiRA achieves up to2.7%percent2.72.7\\%2.7 %better accuracy than LoRA on commonsense reasoning. SHiRA also complements advanced variants of LoRA such as DoRA[20]and can be easily applied on top of them.•Finally, on the training side, we provide a PEFT-based latency- and memory-efficient implementation for SHiRA which trains nearly as fast as standard LoRA while consuming16%percent1616\\%16 %lower peak GPU memory. Beyond PEFT, we provide a simple way to turn any trainer into SHiRA finetuning. For inference, we demonstrate that SHiRA weights can be loaded on a CPU up to5×5\\times5 ×-16×16\\times16 ×faster than equivalent LoRA fusing, thereby enabling rapid switching.\n\n\nThe rest of this paper is organized as follows: section2presents the background and related work. We propose SHiRA in section3while describing its theoretical properties in section4. We then conduct extensive experiments for SHiRA in section5. Finally, we discuss the key findings in section6and conclude the paper in section7.\n\n",
    "related_work": "LoRA, its variants, and sparse adapters.Many LoRA variants exist in literature: DoRA[20], LoRA+[11], VeRA[17], LoRA-FA[35], RS-LoRA[16], among many others. The crucial difference between this literature and our work is that we develop a high rank adapter without increasing training and inference costs. Also, for such methods, the final fused adapter still updates all elements in the pretrained weight tensor, thus prohibiting rapid switching. Moreover, for completeness, we will also show that SHiRA is orthogonal to and can be applied on top of some of the latest, more advanced LoRA variants such as DoRA[20]while preserving the benefits of rapid switching.\n\n\nA few other LoRA variants have also explored a combination of sparsity and low rank adaptation. Examples include RoSA[21], SoRA[6], Sparse-Adapters[12], etc. Among these, Sparse-Adapters[12]explores the use of popular pruning techniques (e.g., SNIP[19]) to prune out adapters to improve their efficiency. SoRA[6]proposes an adaptive rank version of LoRA by gating elements of down and up projection layers and pruning out the zero entries at inference. Finally, RoSA[21]combines a sparse adapter with a low rank one to achieve some high rank benefits. However, since they combine their method with LoRA, the fused adapter weight still overwrites the entire pretrained weight tensor.\n\n\nPartial Finetuning.Our work is most closely related to partial finetuning techniques that were mostly proposed in the pre-LoRA era[36,28,3,33,10]. These methods use a mix of fixed sparse masks[28]or learned masks[36,10]to finetune a pretrained network.\nNote that, these techniques have been mostly explored for relativelysmalllanguage models, andnotfor recent LLMs and diffusion models. Since the LoRA models exploded in popularity, it has been unclear if other sparse finetuning techniques would achieve comparable results to LoRA on generic adapter tasks, particularly in the vision domain. Onesignificantlimitation of partial finetuning, as opposed to LoRA-based methods, is itshigh GPU memory consumption, making itimpracticalto be used for large generative models. Consequently, the reduced memory consumption for finetuning was a key factor to LoRA’s success and its widespread adoption. To this end, we provide a memory- and latency-efficient PEFT-based implementation for SHiRA which trains as efficiently as LoRA, thus requiring significantly lower memory consumption compared to prior partial finetuning techniques. Further, we explore the effectiveness of sparse finetuning on both large language and vision models and provide a detailed analysis on rapid switching and multi-adapter fusion of the high rank adapters.\n\n\nA notable concurrent work is SpIEL[4]which scales partial finetuning to modern LLMs and also has a PEFT implementation that results in comparable speed and memory as LoRA. The main differences between SpIEL and SHiRA are as follows: (i) SpIEL works with dynamic masks while SHiRA uses a static mask. (ii) Dynamic mask in SpIEL requires users to install custom sparse linear layer kernels for the GPUs. In contrast, SHiRA does not require installing any custom kernels and directly works with native Pytorch. Hence, SHiRA’s biggest advantage is its ease of training/inference deployment. (iii) We also analyze multi-adapter fusion properties, e.g., impact of sparsity on orthogonality between adapters, which were not discussed in SpIEL. (iv) Finally, SHiRA demonstrates its effectiveness on both vision and language tasks, whereas SpIEL only discusses the language tasks.\n\n\nMulti-Adapter Fusion.Existing Multi-adapter fusion methods focus on preventing concept loss[8,34,26]. However, these methods usually either just use the base LoRA as it is (and then perform some non-trivial postprocessing on them)[34,26], or some create some minor variants[8]. In contrast, we introduce a new adapter for the concept loss problem where multiple concepts naturally do not interfere with each other. In that respect, our work is orthogonal to the prior multi-adapter fusion work since our adapter can be further postprocessed using such techniques.\n\n\nFigure 2:(a) LoRA when fused into the pretrained model modifies all weights and prevents rapid adapter switching. (b) SHiRA does not require additional weights during training but finetunes very few pretrained weights. Our approach relies on a sparse mask for gradient-masking during training. We show that finetuning as low as1111-2%percent22\\%2 %parameters is sufficient to achieve high accuracy.\n\n",
    "origin_citations": [
      "Dora: Weight-decomposed low-rank adaptation.",
      "Lora+: Efficient low rank adaptation of large models.",
      "Vera: Vector-based random matrix adaptation.",
      "Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning.",
      "A rank stabilization scaling factor for fine-tuning with lora.",
      "Rosa: Accurate parameter-efficient fine-tuning via robust adaptation.",
      "Sparse low-rank adaptation of pre-trained language models.",
      "Sparseadapter: An easy approach for improving the parameter-efficiency of adapters.",
      "Snip: Single-shot network pruning based on connection sensitivity.",
      "Masking as an efficient alternative to finetuning for pretrained language models.",
      "Training neural networks with fixed sparse masks.",
      "Composable sparse fine-tuning for cross-lingual transfer.",
      "Raise a child in large language model: Towards effective and generalizable fine-tuning.",
      "Parameter-efficient transfer learning with diff pruning.",
      "Scaling sparse fine-tuning to large language models.",
      "Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models.",
      "Language models are super mario: Absorbing abilities from homologous models as a free lunch.",
      "Ziplora: Any subject in any style by effectively merging loras."
    ],
    "citations": [
      97191,
      466147,
      315382,
      181330,
      22829,
      484302,
      25001,
      460031,
      443644,
      191234,
      339328,
      340587,
      170626,
      108490,
      123181,
      256978,
      295635,
      24158
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 89,
      "Synthesis_Analysis": 92,
      "Final_Score": 88.1,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 89\nSynthesis_Analysis: 92\nFinal_Score: 88"
    },
    "overall_score": 87.22,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "6hY60tkiEK",
        "forum": "6hY60tkiEK",
        "number": 12426,
        "title": "Sparse High Rank Adapters",
        "normalized_title": "sparse high rank adapters",
        "abstract": "Low Rank Adaptation (LoRA) has gained massive attention in the recent generative AI research. One of the main advantages of LoRA is its ability to be fused with  pretrained models, adding no overhead during inference. However, from a mobile deployment standpoint, we can either avoid inference overhead in the fused mode but lose the ability to switch adapters rapidly, or suffer significant (up to 30% higher) inference latency while enabling rapid switching in the unfused mode. LoRA also exhibits concept-loss when multiple adapters are used concurrently. In this paper, we propose Sparse High Rank Adapters (SHiRA), a new paradigm which incurs no inference overhead, enables rapid switching, and significantly reduces concept-loss. Specifically, SHiRA can be trained by directly tuning only 1-2% of the base model weights while leaving others unchanged. This results in a highly sparse adapter which can be switched directly in the fused mode. We further provide theoretical and empirical insights on how high sparsity in SHiRA can aid multi-adapter fusion by reducing concept loss. Our extensive experiments on LVMs and LLMs demonstrate that finetuning only a small fraction of the parameters in the base model significantly outperforms LoRA while enabling both rapid switching and multi-adapter fusion. Finally, we provide a latency- and memory-efficient SHiRA implementation based on Parameter-Efficient Finetuning (PEFT) Library which trains at nearly the same speed as LoRA while consuming up to 16% lower peak GPU memory, thus making SHiRA easy to adopt for practical use cases. To demonstrate rapid switching benefits during inference, we show that loading SHiRA on a base model can be 5x-16x faster than LoRA fusion on a CPU.",
        "authors": [
          "Kartikeya Bhardwaj",
          "Nilesh Prasad Pandey",
          "Sweta Priyadarshi",
          "Viswanath Ganapathy",
          "Shreya Kadambi",
          "Rafael Esteves",
          "Shubhankar Borse",
          "Paul Whatmough",
          "Risheek Garrepalli",
          "Mart Van Baalen",
          "Harris Teague",
          "Markus Nagel"
        ],
        "decision": null,
        "pdate": 1727288006242,
        "cdate": 1715721884277
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "6hY60tkiEK",
          "forum": "6hY60tkiEK",
          "number": 12426,
          "title": "Sparse High Rank Adapters",
          "normalized_title": "sparse high rank adapters",
          "abstract": "Low Rank Adaptation (LoRA) has gained massive attention in the recent generative AI research. One of the main advantages of LoRA is its ability to be fused with  pretrained models, adding no overhead during inference. However, from a mobile deployment standpoint, we can either avoid inference overhead in the fused mode but lose the ability to switch adapters rapidly, or suffer significant (up to 30% higher) inference latency while enabling rapid switching in the unfused mode. LoRA also exhibits concept-loss when multiple adapters are used concurrently. In this paper, we propose Sparse High Rank Adapters (SHiRA), a new paradigm which incurs no inference overhead, enables rapid switching, and significantly reduces concept-loss. Specifically, SHiRA can be trained by directly tuning only 1-2% of the base model weights while leaving others unchanged. This results in a highly sparse adapter which can be switched directly in the fused mode. We further provide theoretical and empirical insights on how high sparsity in SHiRA can aid multi-adapter fusion by reducing concept loss. Our extensive experiments on LVMs and LLMs demonstrate that finetuning only a small fraction of the parameters in the base model significantly outperforms LoRA while enabling both rapid switching and multi-adapter fusion. Finally, we provide a latency- and memory-efficient SHiRA implementation based on Parameter-Efficient Finetuning (PEFT) Library which trains at nearly the same speed as LoRA while consuming up to 16% lower peak GPU memory, thus making SHiRA easy to adopt for practical use cases. To demonstrate rapid switching benefits during inference, we show that loading SHiRA on a base model can be 5x-16x faster than LoRA fusion on a CPU.",
          "authors": [
            "Kartikeya Bhardwaj",
            "Nilesh Prasad Pandey",
            "Sweta Priyadarshi",
            "Viswanath Ganapathy",
            "Shreya Kadambi",
            "Rafael Esteves",
            "Shubhankar Borse",
            "Paul Whatmough",
            "Risheek Garrepalli",
            "Mart Van Baalen",
            "Harris Teague",
            "Markus Nagel"
          ],
          "decision": null,
          "pdate": 1727288006242,
          "cdate": 1715721884277
        }
      ]
    }
  },
  {
    "paper_id": 49535,
    "title": "Remasking Discrete Diffusion Models with Inference-Time Scaling",
    "abstract": "Part of the success of diffusion models stems from their ability to perform iterative refinement, i.e., repeatedly correcting outputs during generation. However, modern masked discrete diffusion lacks this capability: when a token is generated, it cannot be updated again, even when it introduces an error. Here, we address this limitation by introducing the remasking diffusion model (ReMDM) sampler, a method that can be applied to pretrained masked diffusion models in a principled way and that is derived from a discrete diffusion model with a custom remasking backward process. Most interestingly, ReMDM endows discrete diffusion with a form of inference-time compute scaling. By increasing the number of sampling steps, ReMDM generates natural language outputs that approach the quality of autoregressive models, whereas when the computation budget is limited, ReMDM better maintains quality. ReMDM also improves sample quality of masked diffusion models for discretized images, and in scientific domains such as molecule design, ReMDM facilitates diffusion guidance and pushes the Pareto frontier of controllability relative to classical masking and uniform noise diffusion. We provide the code along with a blog post on the project page:https://remdm.github.io",
    "introduction": "Diffusion models have gained significant traction as algorithms for generating high-quality images and videos(Sohl-Dickstein et al.,2015; Song & Ermon,2019; Ho et al.,2020).\nPart of the success of diffusion stems from its ability to perform iterative refinement—repeatedly modifying outputs and fixing errors over multiple steps of generation—which makes diffusion models especially effective at guided generation(Dhariwal & Nichol,2021; Ho & Salimans,2022), fast sampling(Song et al.,2020a; Salimans & Ho,2022; Song et al.,2023), and supports inference-time scaling to improve sample quality(Ma et al.,2025).\n\n\nDiscrete counterparts of diffusion models(Austin et al.,2021)have also been steadily improving in quality on tasks such as language modeling and biological sequence design(Lou et al.,2024; Sahoo et al.,2024; Schiff et al.,2024). However, modern discrete diffusion—especially the most performant kind that relies on masking(Ou et al.,2024; Shi et al.,2024; Sahoo et al.,2024)—lacks the fundamental ability to iteratively refine outputs.\nOnce a discrete token is unmasked,\nit cannot be updated again, even if it introduces an error.\nThe lack of error correction in turn sets limits on controllable generation, sampling speed, and sample quality.\n\n\nHere, we address the inability of masked diffusion models to perform iterative refinement by introducing a new sampler that supports remasking during generation. Our method, the remasking diffusion model (ReMDM) sampler, is simple and allows users to directly specify the probability of remasking a token at each time step.\nWe augment the sampler with components that range from nucleus sampling to remasking schedules, significantly boosting performance.\n\n\nFigure 1:Our family of masked diffusion processes allow for more flexible generation with remasking of already decoded tokens.\nThis improves sample quality and further closes the gap to AR models.(Left)An illustrative example of errors fixed byReMDM.\nThe first two tokens can be “They sell” or “She sells”, but due to the independence of the parallelized decoding processes, “She sell” is decoded.\nSuch mistakes can be corrected by remasking samplers.(Right)MAUVE scores on OpenWebText.\nMDLM is fromSahoo et al. (2024). FB and DFM denote the forward-backward(Campbell et al.,2022)and discrete flow matching(Gat et al.,2024)correctors, respectively.\n\n\nOur approach for deriving the sampler is rooted in probabilistic modeling—-we show that our method corresponds to ancestral sampling in a discrete diffusion model (also called ReMDM) characterized by a remasking backward process. The ReMDM model admits an objective that is a rescaled version of the MDLM objective(Sahoo et al.,2024)—this suggests using the ReMDM sampler on top of pre-trained MDLMs, which we find to work well.\nWe complement our analysis of the sampler by also showing that it can be interpreted as a predictor-corrector technique(Song et al.,2020b; Campbell et al.,2022).\n\n\nMost interestingly, the ReMDM sampler endows discrete diffusion with a form of inference-time compute scaling. By increasing the number of denoising steps on language modeling tasks, ReMDM generates samples with significantly higher sample quality metrics than any previous diffusion model and almost matches the performance of an autoregressive (AR) model with the same architecture. Conversely, when we reduce sampling steps to increase speed, performance degrades less than other samplers on text and image generation. Lastly, ReMDM models are more amenable to guidance(Schiff et al.,2024)—in molecule design experiments, ReMDM pushes the Pareto frontier of novelty and the desired molecular property relative to alternatives relying on masking or uniform noise diffusion.\n\n\nContributionsWe summarize our contributions as follows:1.We introduce the remasking diffusion model (ReMDM) sampler and a number of add-on components that bring performant iterative refinement via remasking to masked diffusion models.2.We show that our method is a form of ancestral sampling in a probabilistic model whose ELBO is similar to that of classical masked diffusion. This analysis suggests using our sampler on top of pre-trained models, which we find to work well.3.Across the domains of natural language, discretized images, and molecule string representations, we demonstrate empirically that ReMDM endows masked diffusion with inference-time scaling that improves sample quality with more computation, and that also enhances controlled generation.\n\n",
    "related_work": "Discrete Diffusion ModelsDiffusion models(Sohl-Dickstein et al.,2015; Ho et al.,2020)are characterized by parametric modelspθsubscript𝑝𝜃p_{\\theta}italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTtrained to reverse a fixed noising processq𝑞qitalic_qthat maps clean data𝐱𝐱\\mathbf{x}bold_xto increasingly noisy latent variables𝐳tsubscript𝐳𝑡\\mathbf{z}_{t}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT, fort∈[0,1]𝑡01t\\in[0,1]italic_t ∈ [ 0 , 1 ].Austin et al. (2021); Sahoo et al. (2024); Shi et al. (2024)extend this framework to discrete signals.\nFormally, we define𝐱∈{0,1}|V|⊂Δ|V|𝐱superscript01𝑉superscriptΔ𝑉\\mathbf{x}\\in\\{0,1\\}^{|V|}\\subset\\Delta^{|V|}bold_x ∈ { 0 , 1 } start_POSTSUPERSCRIPT | italic_V | end_POSTSUPERSCRIPT ⊂ roman_Δ start_POSTSUPERSCRIPT | italic_V | end_POSTSUPERSCRIPTas one-hot vectors, where|V|𝑉|V|| italic_V |is the vocabulary size andΔ|V|superscriptΔ𝑉\\Delta^{|V|}roman_Δ start_POSTSUPERSCRIPT | italic_V | end_POSTSUPERSCRIPTis the simplex over this vocabulary.\nFor finite-time processes, we letT𝑇Titalic_Tbe the number of time steps.\nFor each time stept⁢(i)=iT∈[0,1]𝑡𝑖𝑖𝑇01t(i)=\\frac{i}{T}\\in[0,1]italic_t ( italic_i ) = divide start_ARG italic_i end_ARG start_ARG italic_T end_ARG ∈ [ 0 , 1 ](i.e., fori=0,1,…,T𝑖01…𝑇i=0,1,\\ldots,Titalic_i = 0 , 1 , … , italic_T), each intermediate latent variable𝐳t⁢(i)∈{0,1}|V|subscript𝐳𝑡𝑖superscript01𝑉\\mathbf{z}_{t(i)}\\in\\{0,1\\}^{|V|}bold_z start_POSTSUBSCRIPT italic_t ( italic_i ) end_POSTSUBSCRIPT ∈ { 0 , 1 } start_POSTSUPERSCRIPT | italic_V | end_POSTSUPERSCRIPThas the marginal:q⁢(𝐳t∣𝐱)=Cat⁢(𝐳t;αt⁢𝐱+(1−αt)⁢𝝅)∈Δ|V|,𝑞conditionalsubscript𝐳𝑡𝐱Catsubscript𝐳𝑡subscript𝛼𝑡𝐱1subscript𝛼𝑡𝝅superscriptΔ𝑉q(\\mathbf{z}_{t}\\mid\\mathbf{x})=\\mathrm{Cat}(\\mathbf{z}_{t};\\alpha_{t}\\mathbf{%\nx}+(1-\\alpha_{t})\\bm{\\pi})\\in\\Delta^{|V|},italic_q ( bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∣ bold_x ) = roman_Cat ( bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ; italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT bold_x + ( 1 - italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) bold_italic_π ) ∈ roman_Δ start_POSTSUPERSCRIPT | italic_V | end_POSTSUPERSCRIPT ,(1)where𝝅𝝅\\bm{\\pi}bold_italic_πis a user-specified prior\nand where we have dropped the explicit dependence oft𝑡titalic_toni𝑖iitalic_i. The predefined scheduleαt∈[0,1]subscript𝛼𝑡01\\alpha_{t}\\in[0,1]italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ [ 0 , 1 ]is monotonically decreasing int𝑡titalic_t.\nFor sequences ofL𝐿Litalic_Ltokens, we denote clean / noisy sequences as𝐱(1:L)superscript𝐱:1𝐿\\mathbf{x}^{(1:L)}bold_x start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT/𝐳t(1:L)superscriptsubscript𝐳𝑡:1𝐿\\mathbf{z}_{t}^{(1:L)}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPTand each tokenℓ∈{1,…,L}ℓ1…𝐿\\ell\\in\\{1,\\ldots,L\\}roman_ℓ ∈ { 1 , … , italic_L }as𝐱(ℓ)superscript𝐱ℓ\\mathbf{x}^{(\\ell)}bold_x start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT/𝐳t(ℓ)superscriptsubscript𝐳𝑡ℓ\\mathbf{z}_{t}^{(\\ell)}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT.\n\n\nMasked Diffusion Language ModelsA special case of this formulation, known as ‘masking’ or ‘absorbing state’ discrete diffusion, sets the limiting distribution𝝅=𝒎,𝝅𝒎\\bm{\\pi}=\\bm{m},bold_italic_π = bold_italic_m ,a one-hot vector centered on a special[MASK]token.\nLettings⁢(i)=i−1T𝑠𝑖𝑖1𝑇s(i)=\\frac{i-1}{T}italic_s ( italic_i ) = divide start_ARG italic_i - 1 end_ARG start_ARG italic_T end_ARGbe the time step directly precedingt𝑡titalic_t, these posteriors take the form:q⁢(𝐳s∣𝐳t,𝐱)=Cat⁢(𝐳s;αs−αt1−αt⁢𝐱+1−αs1−αt⁢𝐳t).𝑞conditionalsubscript𝐳𝑠subscript𝐳𝑡𝐱Catsubscript𝐳𝑠subscript𝛼𝑠subscript𝛼𝑡1subscript𝛼𝑡𝐱1subscript𝛼𝑠1subscript𝛼𝑡subscript𝐳𝑡q(\\mathbf{z}_{s}\\mid\\mathbf{z}_{t},\\mathbf{x})=\\mathrm{Cat}\\Big{(}\\mathbf{z}_{%\ns};\\frac{\\alpha_{s}-\\alpha_{t}}{1-\\alpha_{t}}\\mathbf{x}+\\frac{1-\\alpha_{s}}{1-%\n\\alpha_{t}}\\mathbf{z}_{t}\\Big{)}.italic_q ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_x ) = roman_Cat ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT ; divide start_ARG italic_α start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT - italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_ARG start_ARG 1 - italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_ARG bold_x + divide start_ARG 1 - italic_α start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_ARG start_ARG 1 - italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_ARG bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) .(2)Importantly, despite the recent success of MDLM(Sahoo et al.,2024)and other absorbing state discrete diffusion models, these models suffer from a key drawback, which we call thefailure to remaskproperty.\nNamely, the posterior formula in (2) implies that any unmasked token𝐳t≠𝒎subscript𝐳𝑡𝒎\\mathbf{z}_{t}\\neq\\bm{m}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ≠ bold_italic_mmust remain unchanged throughout the entire denoising process, for alls≤t,𝑠𝑡s\\leq t,italic_s ≤ italic_t ,whichSahoo et al. (2024)formalize asq⁢(𝐳s∣𝐳t,𝐱)=Cat⁢(𝐳s;𝐳t)𝑞conditionalsubscript𝐳𝑠subscript𝐳𝑡𝐱Catsubscript𝐳𝑠subscript𝐳𝑡q(\\mathbf{z}_{s}\\mid\\mathbf{z}_{t},\\mathbf{x})=\\mathrm{Cat}(\\mathbf{z}_{s};%\n\\mathbf{z}_{t})italic_q ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_x ) = roman_Cat ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT ; bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT )if𝐳t≠𝒎subscript𝐳𝑡𝒎\\mathbf{z}_{t}\\neq\\bm{m}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ≠ bold_italic_m.\nOnce a token is decoded, this prediction is ‘locked-in’, a limitation that these diffusion models share with AR models.\n\n\nAs inSahoo et al. (2024), a denoising neural network𝐱θsubscript𝐱𝜃\\mathbf{x}_{\\theta}bold_x start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTis used for parameterizingpθ⁢(𝐳s|𝐳t)=q⁢(𝐳s|𝐳t,𝐱θ⁢(𝐳t))subscript𝑝𝜃conditionalsubscript𝐳𝑠subscript𝐳𝑡𝑞conditionalsubscript𝐳𝑠subscript𝐳𝑡subscript𝐱𝜃subscript𝐳𝑡p_{\\theta}(\\mathbf{z}_{s}|\\mathbf{z}_{t})=q(\\mathbf{z}_{s}|\\mathbf{z}_{t},%\n\\mathbf{x}_{\\theta}(\\mathbf{z}_{t}))italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT | bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) = italic_q ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT | bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) )and trained using the following objective:ℒ=𝔼𝐳0∼q⁢(𝐳0∣𝐱)⁢[−log⁡pθ⁢(𝐱∣𝐳0)]ℒsubscript𝔼similar-tosubscript𝐳0𝑞conditionalsubscript𝐳0𝐱delimited-[]subscript𝑝𝜃conditional𝐱subscript𝐳0\\displaystyle\\mathcal{L}=\\mathbb{E}_{\\mathbf{z}_{0}\\sim q(\\mathbf{z}_{0}\\mid%\n\\mathbf{x})}\\bigg{[}-\\log p_{\\theta}(\\mathbf{x}\\mid\\mathbf{z}_{0})\\bigg{]}caligraphic_L = blackboard_E start_POSTSUBSCRIPT bold_z start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT ∼ italic_q ( bold_z start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT ∣ bold_x ) end_POSTSUBSCRIPT [ - roman_log italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_x ∣ bold_z start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT ) ](3)+𝔼t∈{1T,…,1}⁢𝔼𝐳t∼q⁢(𝐳t∣𝐱)⁢T⁢[αt−αs1−αt⁢log⁡(𝐱θ⁢(𝐳t)⊤⁢𝐱)]subscript𝔼𝑡1𝑇…1subscript𝔼similar-tosubscript𝐳𝑡𝑞conditionalsubscript𝐳𝑡𝐱𝑇delimited-[]subscript𝛼𝑡subscript𝛼𝑠1subscript𝛼𝑡subscript𝐱𝜃superscriptsubscript𝐳𝑡top𝐱\\displaystyle+\\mathbb{E}_{t\\in\\{\\frac{1}{T},\\ldots,1\\}}\\mathbb{E}_{\\mathbf{z}_%\n{t}\\sim q(\\mathbf{z}_{t}\\mid\\mathbf{x})}T\\bigg{[}\\frac{\\alpha_{t}-\\alpha_{s}}{%\n1-\\alpha_{t}}\\log(\\mathbf{x}_{\\theta}(\\mathbf{z}_{t})^{\\top}\\mathbf{x})\\bigg{]}+ blackboard_E start_POSTSUBSCRIPT italic_t ∈ { divide start_ARG 1 end_ARG start_ARG italic_T end_ARG , … , 1 } end_POSTSUBSCRIPT blackboard_E start_POSTSUBSCRIPT bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∼ italic_q ( bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∣ bold_x ) end_POSTSUBSCRIPT italic_T [ divide start_ARG italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT - italic_α start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_ARG start_ARG 1 - italic_α start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_ARG roman_log ( bold_x start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT bold_x ) ]\n\n\nDiscrete GuidanceAn important property of diffusion models is their amenability to controlled generation through the mechanisms of classifier-based (CBG;Dhariwal & Nichol (2021)) and classifier-free guidance (CFG;Nichol et al. (2021); Ho & Salimans (2022)).\nRecently, these mechanisms have been extended to discrete diffusion(Nisonoff et al.,2024; Schiff et al.,2024).\nLetting,γ𝛾\\gammaitalic_γrepresent a hyperparameter that controls guidance strength andy𝑦yitalic_ythe label of the desired class output, following the presentation inSchiff et al. (2024), we implement discrete CBG (D-CBG) by drawing samples for𝐳s(ℓ)superscriptsubscript𝐳𝑠ℓ\\mathbf{z}_{s}^{(\\ell)}bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPTfrom a tempered distribution:pγ⁢(𝐳s(ℓ)∣𝐳t(1:L),y)∝pϕ⁢(y∣𝐳s(ℓ),𝐳t(1:L))γ⋅pθ⁢(𝐳s(ℓ)∣𝐳t(1:L)),proportional-tosuperscript𝑝𝛾conditionalsuperscriptsubscript𝐳𝑠ℓsuperscriptsubscript𝐳𝑡:1𝐿𝑦⋅subscript𝑝italic-ϕsuperscriptconditional𝑦superscriptsubscript𝐳𝑠ℓsuperscriptsubscript𝐳𝑡:1𝐿𝛾subscript𝑝𝜃conditionalsuperscriptsubscript𝐳𝑠ℓsuperscriptsubscript𝐳𝑡:1𝐿p^{\\gamma}(\\mathbf{z}_{s}^{(\\ell)}\\mid\\mathbf{z}_{t}^{(1:L)},y)\\propto p_{\\phi%\n}(y\\mid\\mathbf{z}_{s}^{(\\ell)},\\mathbf{z}_{t}^{(1:L)})^{\\gamma}\\cdot p_{\\theta%\n}(\\mathbf{z}_{s}^{(\\ell)}\\mid\\mathbf{z}_{t}^{(1:L)}),italic_p start_POSTSUPERSCRIPT italic_γ end_POSTSUPERSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT , italic_y ) ∝ italic_p start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPT ( italic_y ∣ bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT , bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT ) start_POSTSUPERSCRIPT italic_γ end_POSTSUPERSCRIPT ⋅ italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT ) ,(4)wherepϕsubscript𝑝italic-ϕp_{\\phi}italic_p start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPTis a classifier trained on noisy sequences and evaluated on sequences where theℓthsuperscriptℓth\\ell^{\\text{th}}roman_ℓ start_POSTSUPERSCRIPT th end_POSTSUPERSCRIPTtoken in𝐳t(1:L)superscriptsubscript𝐳𝑡:1𝐿\\mathbf{z}_{t}^{(1:L)}bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPTis replaced by𝐳s(ℓ)superscriptsubscript𝐳𝑠ℓ\\mathbf{z}_{s}^{(\\ell)}bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT.\nFor discrete CFG (D-CFG), we sample frompγ⁢(𝐳s(ℓ)∣𝐳t(1:L),y)∝pθ⁢(𝐳s(ℓ)∣𝐳t(1:L),y)γ⋅pθ⁢(𝐳s(ℓ)∣𝐳t(1:L))(1−γ),\\begin{split}&p^{\\gamma}(\\mathbf{z}_{s}^{(\\ell)}\\mid\\mathbf{z}_{t}^{(1:L)},y)%\n\\propto\\\\\n&~{}~{}~{}~{}~{}~{}~{}~{}p_{\\theta}(\\mathbf{z}_{s}^{(\\ell)}\\mid\\mathbf{z}_{t}^%\n{(1:L)},y)^{\\gamma}\\cdot p_{\\theta}(\\mathbf{z}_{s}^{(\\ell)}\\mid\\mathbf{z}_{t}^%\n{(1:L)})^{(1-\\gamma)},\\end{split}start_ROW start_CELL end_CELL start_CELL italic_p start_POSTSUPERSCRIPT italic_γ end_POSTSUPERSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT , italic_y ) ∝ end_CELL end_ROW start_ROW start_CELL end_CELL start_CELL italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT , italic_y ) start_POSTSUPERSCRIPT italic_γ end_POSTSUPERSCRIPT ⋅ italic_p start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( bold_z start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( roman_ℓ ) end_POSTSUPERSCRIPT ∣ bold_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( 1 : italic_L ) end_POSTSUPERSCRIPT ) start_POSTSUPERSCRIPT ( 1 - italic_γ ) end_POSTSUPERSCRIPT , end_CELL end_ROW(5)where the same denoising network𝐱θsubscript𝐱𝜃\\mathbf{x}_{\\theta}bold_x start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTis used to model both the conditional and unconditional distributions, with the class label simply masked-out for the latter.\n\n",
    "origin_citations": [
      "Deep unsupervised learning using nonequilibrium thermodynamics.",
      "Denoising diffusion probabilistic models.",
      "Structured denoising diffusion models in discrete state-spaces.",
      "Simple and effective masked diffusion language models.",
      "Simplified and generalized masked diffusion for discrete data.",
      "Diffusion models beat gans on image synthesis.",
      "Glide: Towards photorealistic image generation and editing with text-guided diffusion models.",
      "Classifier-free diffusion guidance.",
      "Unlocking guidance for discrete state-space diffusion and flow models.",
      "Simple guidance mechanisms for discrete diffusion models."
    ],
    "citations": [
      5918,
      4827,
      372220,
      1088134,
      1088536,
      165,
      2559,
      5922,
      630344,
      1091077
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 92,
      "Final_Score": 88.5,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 92\nFinal_Score: 88.5"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 90,
      "Final_Score": 85.9,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 85\nSynthesis_Analysis: 90\nFinal_Score: 86"
    },
    "overall_score": 86.94,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2025/Conference",
        "venue": "NeurIPS",
        "year": "2025",
        "openreview_id": "IJryQAOy0p",
        "forum": "IJryQAOy0p",
        "number": 23437,
        "title": "Remasking Discrete Diffusion Models with Inference-Time Scaling",
        "normalized_title": "remasking discrete diffusion models with inference time scaling",
        "abstract": "Part of the success of diffusion models stems from their ability to perform iterative refinement, i.e., repeatedly correcting outputs during generation. However, modern masked discrete diffusion lacks this capability: when a token is generated, it cannot be updated again, even when it introduces an error. Here, we address this limitation by introducing the remasking diffusion model (ReMDM) sampler, a method that can be applied to pretrained masked diffusion models in a principled way and that is derived from a discrete diffusion model with a custom remasking backward process. Most interestingly, ReMDM endows discrete diffusion with a form of inference-time compute scaling. By increasing the number of sampling steps, ReMDM generates natural language outputs that approach the quality of autoregressive models, whereas when the computation budget is limited, ReMDM better maintains quality. ReMDM also improves sample quality of masked diffusion models for discretized images, and in scientific domains such as molecule design, ReMDM facilitates diffusion guidance and pushes the Pareto frontier of controllability relative to classical masking and uniform noise diffusion. When applied to large pretrained diffusion language models, ReMDM boosts the model’s performance on downstream tasks requiring factual knowledge grasp and reasoning ability.",
        "authors": [
          "Guanghan Wang",
          "Yair Schiff",
          "Subham Sekhar Sahoo",
          "Volodymyr Kuleshov"
        ],
        "decision": null,
        "pdate": 1758217398908,
        "cdate": 1746992298966
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2025/Conference",
          "venue": "NeurIPS",
          "year": "2025",
          "openreview_id": "IJryQAOy0p",
          "forum": "IJryQAOy0p",
          "number": 23437,
          "title": "Remasking Discrete Diffusion Models with Inference-Time Scaling",
          "normalized_title": "remasking discrete diffusion models with inference time scaling",
          "abstract": "Part of the success of diffusion models stems from their ability to perform iterative refinement, i.e., repeatedly correcting outputs during generation. However, modern masked discrete diffusion lacks this capability: when a token is generated, it cannot be updated again, even when it introduces an error. Here, we address this limitation by introducing the remasking diffusion model (ReMDM) sampler, a method that can be applied to pretrained masked diffusion models in a principled way and that is derived from a discrete diffusion model with a custom remasking backward process. Most interestingly, ReMDM endows discrete diffusion with a form of inference-time compute scaling. By increasing the number of sampling steps, ReMDM generates natural language outputs that approach the quality of autoregressive models, whereas when the computation budget is limited, ReMDM better maintains quality. ReMDM also improves sample quality of masked diffusion models for discretized images, and in scientific domains such as molecule design, ReMDM facilitates diffusion guidance and pushes the Pareto frontier of controllability relative to classical masking and uniform noise diffusion. When applied to large pretrained diffusion language models, ReMDM boosts the model’s performance on downstream tasks requiring factual knowledge grasp and reasoning ability.",
          "authors": [
            "Guanghan Wang",
            "Yair Schiff",
            "Subham Sekhar Sahoo",
            "Volodymyr Kuleshov"
          ],
          "decision": null,
          "pdate": 1758217398908,
          "cdate": 1746992298966
        }
      ]
    }
  },
  {
    "paper_id": 44473,
    "title": "Weighted-Reward Preference Optimization for Implicit Model Fusion",
    "abstract": "While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2% against GPT-4-0314 on Arena-Hard. Our code is available athttps://github.com/SLIT-AI/WRPO.",
    "introduction": "Combining the strengths of multiple Large Language Models (LLMs) can potentially enhance the capabilities of individual models. Model ensemble techniques(Jiang et al.,2023b; Wang et al.,2024b)aggregate predictions from several models to improve overall performance and robustness over a single model. However, this approach requires substantial computational resources, as all models must remain active during inference. The Mixture of Experts (MoE)(Komatsuzaki et al.,2023; Feng et al.,2024; Sukhbaatar et al.,2024)leverages sparse expert networks to boost capacity by activating only a subset of parameters. Despite reduced activation, MoEs still incur significant memory overhead, as all parameters must be maintained. Model merging(Wortsman et al.,2022; Matena & Raffel,2022; Yadav et al.,2023), which combines independently trained instances of the same model through arithmetic operations, allows a single model to be maintained during inference. While more efficient, this method is restricted to models with identical architectures and sizes.\n\n\nAnother approach is to fuse these LLMs into a target model through multi-teacher knowledge distillation(Wan et al.,2024a;b; Shi et al.,2024). Unlike traditional knowledge distillation(Gou et al.,2021), which usually leverages diverse sources (e.g., logits, features, and relations) of knowledge from teacher models, this method relies exclusively on the probabilistic distribution matrices generated by source LLMs to transfer knowledge to the target model.\nWe refer to this method asexplicit model fusion(EMF) because it involves a well-defined knowledge transfer process. While applicable to models with varying architectures and sizes, and without increasing memory overhead during inference, this approach presents notable challenges such as vocabulary alignment and the merging of distribution matrices from different LLMs. These issues complicate model fusion, reduce its efficiency, and may introduce noise and errors and affect the fusion results.\n\n\nThis work aims to enhance the capabilities of a single LLM by implicitly learning from robust open-source LLMs, a process we termimplicit model fusion(IMF). The concept of IMF has been widely utilized to improve the performance of weaker models. For instance, a weak model can be boosted through fine-tuning with outputs from stronger LLMs(Ranaldi & Freitas,2024; Tian et al.,2024; Kang et al.,2023). Moreover, a reward model can be trained using outputs from various LLMs(Cui et al.,2024; Zhu et al.,2024),\nenabling it to learn and capture the differences in capabilities between the LLMs. Zephyr(Tunstall et al.,2023)further collects responses from multiple LLMs and ranks them with GPT-4 to obtain preference data for training the policy using DPO. One advantage of IMF over EMF(Wan et al.,2024a;b; Shi et al.,2024)is that it eliminates the need for challenging alignment of vocabularies and fusion of distributions among different LLMs. Inspired by recent alignment techniques such as Direct Preference Optimization (DPO)(Rafailov et al.,2023)and Simple Preference Optimization (SimPO)(Meng et al.,2024), we propose a novel IMF method to transfer the capabilities of source LLMs to a target LLM through preference optimization. However, directly applying preference learning to outputs from heterogeneous LLMs presents challenges. Previous works have shown that DPO is highly sensitive to distribution shifts between the policy model and the preference data(Xu et al.,2024b; Tajwar et al.,2024; Zhou et al.,2024), and training a policy model on this preference data can lead to sub-optimal performance.\n\n\nTo demonstrate this, we conduct a preliminary experiment on the Ultrafeedback dataset(Cui et al.,2024), using LLaMA3-8B-Instruct(Dubey et al.,2024)as the target model and 10 strong open-source LLMs as source models.111Refer to Section4.1for more details.For each prompt, we first ask each source model to generate several responses and use the ArmoRM reward model(Wang et al.,2024a)to select the highest-reward response among all source LLMs as thepreferredresponse, with thedispreferredresponse coming from the target LLM’s completions. Figure1(a) visualizes the average log-probability distribution of the target LLMπθsubscript𝜋𝜃\\pi_{\\theta}italic_π start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTfor both response types, which reveals a significant deviation between the distributions of the source and target models. Although applying DPO directly on this deviated dataset marginally enhances the log-probabilities of source LLMs’ responses relative to those of the target LLM, as shown in Figure1(b), this results in sub-optimal performance compared to sampling both response types exclusively from the target LLM, as illustrated in Figure1(c).\n\n\nFigure 1:Distribution deviations between responses from heterogeneous source LLMs and the LLaMA3-8B-Instruct target LLM before (a) and after (b) DPO fine-tuning, with the prompts from Ultrafeedback(Cui et al.,2024)as input. Subfigure (c) shows the results (πDPO-offsubscript𝜋DPO-off\\pi_{\\text{DPO-off}}italic_π start_POSTSUBSCRIPT DPO-off end_POSTSUBSCRIPT) of preference optimization with this deviated preference dataset, compared to the results (πθsubscript𝜋𝜃\\pi_{\\theta}italic_π start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT) from directly applying the target model and those (πDPO-onsubscript𝜋DPO-on\\pi_{\\text{DPO-on}}italic_π start_POSTSUBSCRIPT DPO-on end_POSTSUBSCRIPT) from DPO fine-tuning on un-deviated preference data sampled from the target model.\n\n\nTo address the distributional deviations during implicit model fusion, we introduce a novel approach calledWeighted-RewardPreferenceOptimization (WRPO). Instead of directly relying on the source LLMs to provide preferred responses, we propose a progressive adaptation strategy that begins with the target LLM providing preferred responses and gradually shifts this responsibility to source LLMs. Specifically, this progressive adaptation is implemented in two stages. First, for each promptx𝑥xitalic_x, we construct a preference quadruple(x,yws,ywt,yl)𝑥subscript𝑦subscript𝑤𝑠subscript𝑦subscript𝑤𝑡subscript𝑦𝑙(x,y_{w_{s}},y_{w_{t}},y_{l})( italic_x , italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_POSTSUBSCRIPT , italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUBSCRIPT , italic_y start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT ), whereywssubscript𝑦subscript𝑤𝑠y_{w_{s}}italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_POSTSUBSCRIPTis a preferred response generated by the source LLMs, andywtsubscript𝑦subscript𝑤𝑡y_{w_{t}}italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUBSCRIPTandylsubscript𝑦𝑙y_{l}italic_y start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPTare preferred and dispreferred responses, respectively, from the target LLM. Second, we gradually decrease the weight of internal rewards222We use “internal reward” to refer to the reward generated during preference optimization for preferred or dispreferred responses, in contrast to the reward provided by an external reward model.forywtsubscript𝑦subscript𝑤𝑡y_{w_{t}}italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUBSCRIPTand increase the weight forywssubscript𝑦subscript𝑤𝑠y_{w_{s}}italic_y start_POSTSUBSCRIPT italic_w start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_POSTSUBSCRIPTduring preference optimization. This smoothing process facilitates the integration of strengths from the source models into the target model while mitigating the distributional discrepancies.\n\n\nTo assess the effectiveness of WRPO in implicit model fusion, we select 10 prominent open-source LLMs as the source models, with parameter sizes ranging from 9B to 236B. We chose LLaMA3-8B-Instruct(Dubey et al.,2024)as the target model due to its strong performance relative to its size. Our experiments are conducted on three widely-used instruction-following benchmarks, namely, MT-Bench(Zheng et al.,2023), AlpacaEval-2(Li et al.,2023), and Arena-Hard(Li et al.,2024). The results show that WRPO consistently outperforms existing fusion methods and various baselines. This highlights its ability to allow a model to implicitly learn from the diverse capabilities of heterogeneous LLMs while also addressing distributional shifts. Notably, the fused model, LLaMA3-8B-Instruct-WRPO, surpasses all source models on AlpacaEval-2 with a length-controlled win rate of 55.9%.\n\n",
    "related_work": "Collective LLMsGiven that LLMs are trained with various architectures and sizes on different datasets, it is reasonable to assume they possess unique capabilities and strengths. Therefore, leveraging the distinct advantages of different LLMs becomes a natural approach to developing more robust and high-capable models. Recent studies have increasingly emphasized the development of collective LLMs through the integration of diverse heterogeneous models.LLM-Blender(Jiang et al.,2023b)presents an ensemble framework that first employs a pairwise ranking mechanism to identify the top-K𝐾Kitalic_Koutputs generated by different LLMs. These selected outputs are then refined by a seq2seq model to produce enhanced results.\nMixture-of-Agents (MoA)(Wang et al.,2024b)utilizes a hierarchical structure where each layer consists of multiple LLM agents. The outputs from a previous layer are concatenated and refined by each agent in the subsequent layer. However, this approach significantly increases the number of LLMs needed during inference.\nIn addition to the sequence-level ensemble,Xu et al. (2024c)explored a token-level ensemble method that aggregates the distributions of LLMs at each decoding step through a global alignment matrix. Similarly, PackLLMs(Mavromatis et al.,2024)conducts distribution ensembling during inference utilizing sequence-level weights derived from the perplexity of each LLM on the input.FuseLLM(Wan et al.,2024a)and FuseChat(Wan et al.,2024b)aim to fuse LLMs of various architectures and sizes into a more robust model through multi-teacher knowledge distillation. They start by aligning the vocabularies and probabilistic distributions of the source LLMs, followed by merging their distributions and continuously fine-tuning the target LLM.\nProFuser(Shi et al.,2024)goes further by integrating both training mode (through cross-entropy loss) and inference mode (via model outputs), which provides a more comprehensive understanding of the capabilities of source LLMs. Although applicable to models with varying architectures and sizes, these methods face challenges such as vocabulary alignment and merging distribution matrices from different LLMs, which are complex and may also introduce noise and errors that affect the fusion results.\n\n\nDirect Preference OptimizationAligning LLMs with human preferences is crucial for their success. Reinforcement Learning from Human Feedback (RLHF)(Christiano et al.,2017; Schulman et al.,2017; Ziegler et al.,2019)is a widely used approach to achieve this alignment. However, RLHF depends on complex reinforcement learning techniques such as Proximal Policy Optimization (PPO), which are challenging to implement and often unstable during training. To address these challenges, approaches such as SLiC-HF(Zhao et al.,2023)and RRHF(Yuan et al.,2023)replace reinforcement learning with a ranking loss on preference pairs to better align LLMs with human preferences, while also incorporating a regularization term based on reference responses. Similarly, DPO(Rafailov et al.,2023)directly optimizes the policy model by training the reward model on human preference data. In addition to providing more stable training, lower computational costs, and easier implementation, this approach ensures high-quality alignment with human preferences.Subsequent research aims to address the potential limitations of DPO. For example, IPO(Azar et al.,2024)tackles the risk of overfitting by optimizing a nonlinear preference function, thus avoiding the transformation of pairwise preferences into pointwise rewards. KTO(Ethayarajh et al.,2024)is based on a new alignment objective of human-aware loss (HALO), which maximizes the utility of generated outputs directly from a binary signal indicating whether the output is desirable, rather than maximizing the likelihood of preferences. CPO(Xu et al.,2024a)and ORPO(Hong et al.,2024)aim to eliminate the need for a reference model by streamlining the optimization process, combining supervised fine-tuning (SFT) and preference alignment into a single step.\nR-DPO(Park et al.,2024)introduces a length-regularization term into the DPO objective to mitigate length biases that may be exploited by DPO.\nSimilarly, SimPO(Meng et al.,2024)revises the reward component in DPO to use the average log probability of positive or negative responses from the policy model. Another motivation for this method is that the training process aligns more closely with inference.However, none of the above works consider the hybrid scenario where one response is generated by the policy itself while the other comes from a different LLM. This situation may introduce serious distribution shifts relative to the policy, which in turn affects the policy’s optimization. The work closely related to our setup is WPO(Zhou et al.,2024), which assigns weights to off-policy preference pairs based on their likelihood under the policy model. These weights indicate the degree of deviation from the policy’s distribution and mitigate the influence of preference pairs with notable deviations.\n\n",
    "origin_citations": [
      "LLM-Blender: Ensembling large language models with pairwise ranking and generative fusion.",
      "Mixture-of-Agents enhances large language model capabilities.",
      "Bridging the gap between different vocabularies for llm ensemble.",
      "Pack of LLMs: Model fusion at test-time via perplexity optimization.",
      "Knowledge fusion of large language models.",
      "FuseChat: Knowledge fusion of chat models.",
      "ProFuser: Progressive fusion of large language models.",
      "Deep reinforcement learning from human preferences.",
      "Proximal policy optimization algorithms.",
      "Fine-tuning language models from human preferences.",
      "SLiC-HF: Sequence likelihood calibration with human feedback.",
      "RRHF: Rank responses to align language models with human feedback.",
      "Direct Preference Optimization: Your language model is secretly a reward model.",
      "A general theoretical paradigm to understand learning from human preferences.",
      "KTO: Model alignment as prospect theoretic optimization.",
      "Contrastive preference optimization: Pushing the boundaries of LLM performance in machine translation.",
      "ORPO: Monolithic preference optimization without reference model.",
      "Disentangling length from quality in direct preference optimization.",
      "SimPO: Simple preference optimization with a reference-free reward.",
      "WPO: Enhancing RLHF with weighted preference optimization."
    ],
    "citations": [
      307233,
      989021,
      1087690,
      136139,
      126096,
      464139,
      1091135,
      1354,
      2219,
      3568,
      261592,
      5930,
      315102,
      101430,
      126956,
      288579,
      423183,
      1055123,
      1090406
    ],
    "extract_rate": 0.95,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 85,
      "Final_Score": 87.2,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 92\nSynthesis_Analysis: 85\nFinal_Score: 87"
    },
    "overall_score": 86.92,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "fq24pEb8SL",
        "forum": "fq24pEb8SL",
        "number": 6241,
        "title": "Weighted-Reward Preference Optimization for Implicit Model Fusion",
        "normalized_title": "weighted reward preference optimization for implicit model fusion",
        "abstract": "While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9\\% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2\\% against GPT-4-0314 on Arena-Hard. Our code is available at https://github.com/SLIT-AI/WRPO.",
        "authors": [
          "Ziyi Yang",
          "Fanqi Wan",
          "Longguang Zhong",
          "Tianyuan Shi",
          "Xiaojun Quan"
        ],
        "decision": null,
        "pdate": 1737562454518,
        "cdate": 1727338196552
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "fq24pEb8SL",
          "forum": "fq24pEb8SL",
          "number": 6241,
          "title": "Weighted-Reward Preference Optimization for Implicit Model Fusion",
          "normalized_title": "weighted reward preference optimization for implicit model fusion",
          "abstract": "While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9\\% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2\\% against GPT-4-0314 on Arena-Hard. Our code is available at https://github.com/SLIT-AI/WRPO.",
          "authors": [
            "Ziyi Yang",
            "Fanqi Wan",
            "Longguang Zhong",
            "Tianyuan Shi",
            "Xiaojun Quan"
          ],
          "decision": null,
          "pdate": 1737562454518,
          "cdate": 1727338196552
        }
      ]
    }
  },
  {
    "paper_id": 27117,
    "title": "Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation",
    "abstract": "Recent works on open-vocabulary 3D instance segmentation show strong promise, but at the cost of slow inference speed and high computation requirements. This high computation cost is typically due to their heavy reliance on 3D clip features, which require computationally expensive 2D foundation models like Segment Anything (SAM) and CLIP for multi-view aggregation into 3D. As a consequence, this hampers their applicability in many real-world applications that require both fast and accurate predictions. To this end, we propose a fast yet accurate open-vocabulary 3D instance segmentation approach, named Open-YOLO 3D, that effectively leverages only 2D object detection from multi-view RGB images for open-vocabulary 3D instance segmentation.\nWe address this task by generating class-agnostic 3D masks for objects in the scene and associating them with text prompts.\nWe observe that the projection of class-agnostic 3D point cloud instances already holds instance information; thus, using SAM might only result in redundancy that unnecessarily increases the inference time.\nWe empirically find that a better performance of matching text prompts to 3D masks can be achieved in a faster fashion with a 2D object detector.\nWe validate our Open-YOLO 3D on two benchmarks, ScanNet200 and Replica,\nunder two scenarios:(i)with ground truth masks, where labels are required for given object proposals, and(ii)with class-agnostic 3D proposals generated from a 3D proposal network.\nOur Open-YOLO 3D achieves state-of-the-art performance on both datasets while obtaining up to∼similar-to\\sim∼16×\\times×speedup compared to the best existing method in literature.\nOn ScanNet200 val. set, our Open-YOLO 3D achieves mean average precision (mAP) of 24.7% while operating at 22 seconds per scene.\nCode and model are available atgithub.com/aminebdj/OpenYOLO3D",
    "introduction": "3D instance segmentation is a computer vision task that involves the prediction of masks for individual objects in a 3D point cloud scene. It holds significant importance in fields like robotics and augmented reality. Due to its diverse applications, this task has garnered increasing attention in recent years. Researchers have long focused on methods that typically operate within a closed-set framework, limiting their ability to recognize objects not present in the training data. This constraint poses challenges, particularly when novel objects must be identified or categorized in unfamiliar environments. Recent methods[34,42]address the problem of novel class segmentation, but they suffer from slow inference that ranges from 5 minutes for small scenes to 10 minutes for large scenes due to their reliance on computationally heavy foundation models like SAM[23]and CLIP[55]along with heavy computation for lifting 2D CLIP feature to 3D.\n\n\nOpen-vocabulary 3D instance segmentation is important for robotics tasks such as, material handling where the robot is expected to perform operations from text-based instructions like moving specific products, loading and unloading goods, and inventory management while being fast in the decision-making process. Although state-of-the-art open-vocabulary 3D instance segmentation methods show high promise in terms of generalizability to novel objects, they still operate in minutes of inference time due to their reliance on heavy foundation models such as SAM. Motivated by recent advances in 2D object detection[7], we look into an alternative approach that leverages fast object detectors instead of utilizing computationally expensive foundation models.\n\n\nThis paper proposes a novel open-vocabulary 3D instance segmentation method, named Open-YOLO 3D, that utilizes efficient, joint 2D-3D reasoning, using 2D bounding box predictions to replace computationally-heavy segmentation models. We employ an open-vocabulary 2D object detector to generate bounding boxes with their class labels for all frames corresponding to the 3D scene; on the other side, we utilize a 3D instance segmentation network to generate 3D class-agnostic instance masks for the point clouds, which proves to be much faster than 3D proposal generation methods from 2D instances[34,32].\nUnlike recent methods[42,34]which use SAM and CLIP to lift 2D clip features to 3D for prompting the 3D mask proposal, we propose an alternative approach that relies on the bounding box predictions from 2D object detectors which prove to be significantly faster than CLIP-based methods. We utilize the predicted bounding boxes in all RGB frames corresponding to the point cloud scene to construct a Low Granularity (LG) label map for every frame. One LG label map is a two-dimensional array with the same height and width as the RGB frame, with the bounding box areas replaced by their predicted class label. Next, we use intrinsic and extrinsic parameters to project the point cloud scene onto their respective LG label maps with top-k visibility for final class prediction.\nWe present an example output of our method in Figure1.\nOur contributions are following:•We introduce a 2D object detection-based approach for open-vocabulary labeling of 3D instances, which greatly improves the efficiency compared to 2D segmentation approaches.•We propose a novel approach to scoring 3D mask proposals using only bounding boxes from 2D object detectors.•Our Open-YOLO 3D achieves superior performance on two benchmarks, while being considerably faster than existing methods in the literature. On ScanNet200 val. set, our Open-YOLO 3D achieves an absolute gain of 2.3% at mAP50 while being∼similar-to\\sim∼16x faster compared to the recent Open3DIS[34].\n\n",
    "related_work": "Closed-vocabulary 3D segmentation:The 3D instance segmentation task aims at predicting masks for individual objects in a 3D scene, along with a class label belonging to the set of known classes. Some methods use a grouping-based approach in a bottom-up manner, by learning embeddings in the latent space to facilitate clustering of object points[4,14,15,21,26,29,46,54]. Conversely, proposal-based methods adopt a top-down strategy, initially detecting 3D bounding boxes and then segmenting the object region within each box[10,17,31,49,52]. Notably, inspired by advancements in 2D works[5,6], transformer designs[43]have been recently applied to 3D instance segmentation tasks[39,41,24,1,20]. Mask3D[39]introduces the first hybrid architecture that combines Convolutional Neural Networks (CNN) and transformers for this task. It uses a 3D CNN backbone to extract per-point features and a transformer-based instance mask decoder to refine a set of queries.\nBuilding on Mask3D, the authors of[1]show that using explicit spatial and semantic supervision at the level of the 3D backbone further improves the instance segmentation results.\nOneformer3D[24]follows a similar architecture and introduces learnable kernels in the transformer decoder for a unified semantic, instance, and panoptic segmentation.\nODIN[20]proposes an architecture that uses 2D-3D fusion to generate the masks and class labels.\nOther methods introduce weakly-supervised alternatives to dense annotation approaches, aiming to reduce the annotation cost associated with 3D data[8,18,47]. While these methodologies strive to enhance the quality of 3D instance segmentation, they typically rely on a predefined set of semantic labels. In contrast, our proposed approach aims at segmenting objects with both known and unknown class labels.\n\n\nOpen-vocabulary 2D recognition:This task aims at identifying both known and novel classes, where the labels of the known classes are available in the training set, while the novel classes are not encountered during training.\nIn the direction of open-vocabulary object detection (OVOD), several approaches have been proposed[58,36,30,53,45,22,51,7].\nAnother widely studied task is open-vocabulary segmentation (OVSS)[3,48,27,12,28].\nRecent open-vocabulary semantic segmentation methods[27,12,28]leverage pre-trained CLIP[55]to perform open-vocabulary segmentation, where the model is trained to output a pixel-wise feature that is aligned with the text embedding in the CLIP space.\nFurthermore, AttrSeg[33]proposes a decomposition-aggregation framework where vanilla class names are first decomposed into various attribute descriptions, and then different attribute representations are aggregated into a final class representation.\nOpen-vocabulary instance segmentation (OVIS) aims at predicting instance masks while preserving high zero-shot capabilities.\nOne approach[19]proposes a cross-modal pseudo-labeling framework, where a student model is supervised with pseudo-labels for the novel classes from a teacher model. Another approach[44]proposes an annotation-free method where a pre-trained vision-language model is used to produce annotations at both the box and pixel levels.\nAlthough these methods show high zero-shot performance and real-time speed, they are still limited to 2D applications only.\n\n\nOpen-vocabulary 3D segmentation:Several methods[35,13,16]have been proposed to address the challenges of open-vocabulary semantic segmentation where they use foundation models like clip for unknown class discovery, while the authors of[2]focus on weak supervision for unknown class discovery without relying on any 2D foundation model. OpenScene[35]makes use of 2D open-vocabulary semantic segmentation models to lift the pixel-wise 2D CLIP features into the 3D space, which allows the 3D model to perform 3D open-vocabulary point cloud semantic segmentation.\nOn the other hand, ConceptGraphs[13]relies on creating an open-vocabulary scene graph that captures object properties such as spatial location, enabling a wide range of downstream tasks including segmentation, object grounding, navigation, manipulation, localization, and remapping.\nIn the direction of 3D point cloud instance segmentation,\nOpenMask3D[42]uses a 3D instance segmentation network to generate class-agnostic mask proposals, along with SAM[23]and CLIP[55], to construct a 3D clip feature for each mask using RGB-D images associated with the 3D scene.\nUnlike OpenMask3D where a 3D proposal network is used, OVIR-3D[32]generates 3D proposals by fusing 2D masks obtained by a 2D instance segmentation model. Open3DIS[34]combines proposals from 2D and 3D with novel 2D masks fusion approaches via hierarchical agglomerative clustering, and also proposes to use point-wise 3D CLIP features instead of mask-wise features. The two most recent approaches in[34,42]show promising generalizability in terms of novel class discovery[42]and novel object geometries especially small objects[34]. However, they both suffer from slow inference speed, as they rely on SAM for 3D mask proposal clip feature aggregation in the case of OpenMask3D[42], and for novel 3D proposal masks generation from 2D masks[34].\n\n\nFigure 2:Proposed open-world 3D instance segmentation pipeline.We use a 3D instance segmentation network (3D Network) for generating class-agnostic proposals. For open-vocabulary prediction, a 2D Open-Vocabulary Object Detector (2D OVOD) generates bounding boxes with class labels. These predictions are used to construct label maps for all input frames. Next, we assign the top-k label maps to each 3D proposal based on visibility. Finally, we generate a Multi-View Prompt Distribution from the 2D projections of the proposals to match a text prompt to every 3D proposal.\n\n",
    "origin_citations": [
      "Hierarchical aggregation for 3d instance segmentation.",
      "Occuseg: Occupancy-aware 3d instance segmentation.",
      "Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution.",
      "Pointgroup: Dual-set point grouping for 3d instance segmentation.",
      "3d instance segmentation via multi-task metric learning.",
      "Instance segmentation in 3d scenes using semantic superpoint tree networks.",
      "Sgpn: Similarity group proposal network for 3d point cloud instance segmentation.",
      "Point cloud instance segmentation using probabilistic embeddings.",
      "3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation.",
      "3d-sis: 3d semantic instance segmentation of rgb-d scans.",
      "Learning gaussian instance segmentation in point clouds.",
      "Learning object bounding boxes for 3d instance segmentation on point clouds.",
      "Gspn: Generative shape proposal network for 3d instance segmentation in point cloud.",
      "Masked-attention mask transformer for universal image segmentation.",
      "Per-pixel classification is not all you need for semantic segmentation.",
      "Attention is all you need.",
      "Mask3D: Mask Transformer for 3D Semantic Instance Segmentation.",
      "Superpoint transformer for 3d scene instance segmentation.",
      "Oneformer3d: One transformer for unified point cloud segmentation.",
      "3d instance segmentation via enhanced spatial and semantic supervision.",
      "Odin: A single model for 2d and 3d perception.",
      "Box2mask: Weakly supervised 3d semantic instance segmentation using bounding boxes.",
      "Exploring data-efficient 3d scene understanding with contrastive scene contexts.",
      "Pointcontrast: Unsupervised pre-training for 3d point cloud understanding.",
      "Regionclip: Region-based language-image pretraining.",
      "Lp-ovod: Open-vocabulary object detection by linear probing.",
      "Grounding dino: Marrying dino with grounded pre-training for open-set object detection.",
      "Open-vocabulary detr with conditional matching.",
      "Object-aware distillation pyramid for open-vocabulary object detection.",
      "Multi-modal classifiers for open-vocabulary object detection.",
      "Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment.",
      "Yolo-world: Real-time open-vocabulary object detection.",
      "Zero-shot semantic segmentation.",
      "Groupvit: Semantic segmentation emerges from text supervision.",
      "Language-driven semantic segmentation.",
      "Scaling open-vocabulary image segmentation with image-level labels.",
      "Open-vocabulary semantic segmentation with mask-adapted clip.",
      "Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip.",
      "Open-vocabulary semantic segmentation via attribute decomposition-aggregation.",
      "Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling.",
      "Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations.",
      "Openscene: 3d scene understanding with open vocabularies.",
      "Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning.",
      "3d concept learning and reasoning from multi-view images.",
      "3d indoor instance segmentation in an open-world.",
      "Openmask3d: Open-vocabulary 3d instance segmentation.",
      "Segment anything.",
      "Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data.",
      "Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance."
    ],
    "citations": [
      95718,
      1546,
      431278,
      7446,
      7447,
      454735,
      398783,
      9114,
      1543,
      329227,
      521860,
      7441,
      359465,
      152903,
      1756,
      38,
      460654,
      242545,
      23765,
      413768,
      106799,
      521402,
      537669,
      312538,
      76830,
      501544,
      76650,
      305746,
      19076,
      122726,
      1781,
      174081,
      255086,
      536546,
      460116,
      77185,
      1090158,
      338145,
      492475,
      242581,
      477676,
      495078,
      478740,
      48698,
      2512,
      295758,
      349313
    ],
    "extract_rate": 0.9591836734693877,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 88,
      "Final_Score": 88.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 88\nFinal_Score: 89"
    },
    "overall_score": 86.86,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "CRmiX0v16e",
        "forum": "CRmiX0v16e",
        "number": 4987,
        "title": "Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation",
        "normalized_title": "open yolo 3d towards fast and accurate open vocabulary 3d instance segmentation",
        "abstract": "Recent works on open-vocabulary 3D instance segmentation show strong promise but at the cost of slow inference speed and high computation requirements. This high computation cost is typically due to their heavy reliance on aggregated clip features from multi-view, which require computationally expensive 2D foundation models like Segment Anything (SAM) and CLIP. Consequently, this hampers their applicability in many real-world applications that require both fast and accurate predictions. To this end, we propose a novel open-vocabulary 3D instance segmentation approach, named Open-YOLO 3D, that efficiently leverages only 2D object detection from multi-view RGB images for open-vocabulary 3D instance segmentation. \n We demonstrate that our proposed Multi-View Prompt Distribution (MVPDist) method makes use of multi-view information to account for misclassification from the object detector to predict a reliable label for 3D instance masks. Furthermore, since projections of 3D object instances are already contained within the 2D bounding boxes, we show that our proposed low granularity label maps, which require only a 2D object detector to construct, are sufficient and very fast to predict prompt IDs for 3D instance masks when used with our proposed MVPDist.\n We validate our Open-YOLO 3D on two benchmarks, ScanNet200 and Replica, \n under two scenarios: (i) with ground truth masks, where labels are required for given object proposals, and (ii) with class-agnostic 3D proposals generated from a 3D proposal network.\n Our Open-YOLO 3D achieves state-of-the-art performance on both datasets while obtaining up to $\\sim$16$\\times$ speedup compared to the best existing method in literature. On ScanNet200 val. set, our Open-YOLO 3D achieves mean average precision (mAP) of 24.7% while operating at 22 seconds per scene. github.com/aminebdj/OpenYOLO3D",
        "authors": [
          "Mohamed El Amine Boudjoghra",
          "Angela Dai",
          "Jean Lahoud",
          "Hisham Cholakkal",
          "Rao Muhammad Anwer",
          "Salman Khan",
          "Fahad Shahbaz Khan"
        ],
        "decision": null,
        "pdate": 1737562385098,
        "cdate": 1727291470028
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "CRmiX0v16e",
          "forum": "CRmiX0v16e",
          "number": 4987,
          "title": "Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation",
          "normalized_title": "open yolo 3d towards fast and accurate open vocabulary 3d instance segmentation",
          "abstract": "Recent works on open-vocabulary 3D instance segmentation show strong promise but at the cost of slow inference speed and high computation requirements. This high computation cost is typically due to their heavy reliance on aggregated clip features from multi-view, which require computationally expensive 2D foundation models like Segment Anything (SAM) and CLIP. Consequently, this hampers their applicability in many real-world applications that require both fast and accurate predictions. To this end, we propose a novel open-vocabulary 3D instance segmentation approach, named Open-YOLO 3D, that efficiently leverages only 2D object detection from multi-view RGB images for open-vocabulary 3D instance segmentation. \n We demonstrate that our proposed Multi-View Prompt Distribution (MVPDist) method makes use of multi-view information to account for misclassification from the object detector to predict a reliable label for 3D instance masks. Furthermore, since projections of 3D object instances are already contained within the 2D bounding boxes, we show that our proposed low granularity label maps, which require only a 2D object detector to construct, are sufficient and very fast to predict prompt IDs for 3D instance masks when used with our proposed MVPDist.\n We validate our Open-YOLO 3D on two benchmarks, ScanNet200 and Replica, \n under two scenarios: (i) with ground truth masks, where labels are required for given object proposals, and (ii) with class-agnostic 3D proposals generated from a 3D proposal network.\n Our Open-YOLO 3D achieves state-of-the-art performance on both datasets while obtaining up to $\\sim$16$\\times$ speedup compared to the best existing method in literature. On ScanNet200 val. set, our Open-YOLO 3D achieves mean average precision (mAP) of 24.7% while operating at 22 seconds per scene. github.com/aminebdj/OpenYOLO3D",
          "authors": [
            "Mohamed El Amine Boudjoghra",
            "Angela Dai",
            "Jean Lahoud",
            "Hisham Cholakkal",
            "Rao Muhammad Anwer",
            "Salman Khan",
            "Fahad Shahbaz Khan"
          ],
          "decision": null,
          "pdate": 1737562385098,
          "cdate": 1727291470028
        }
      ]
    }
  },
  {
    "paper_id": 25779,
    "title": "MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation",
    "abstract": "This study aims to construct an audio-video generative model with minimal computational cost by leveraging pre-trained single-modal generative models for audio and video.\nTo achieve this, we propose a novel method that guides single-modal models to cooperatively generate well-aligned samples across modalities.\nSpecifically, given two pre-trained base diffusion models, we train a lightweight joint guidance module to adjust scores separately estimated by the base models to match the score of joint distribution over audio and video.\nWe show that this guidance can be computed using the gradient of the optimal discriminator, which distinguishes real audio-video pairs from fake ones independently generated by the base models.\nBased on this analysis, we construct a joint guidance module by training this discriminator.\nAdditionally, we adopt a loss function to stabilize the discriminator’s gradient and make it work as a noise estimator, as in standard diffusion models.\nEmpirical evaluations on several benchmark datasets demonstrate that our method improves both single-modal fidelity and multimodal alignment with relatively few parameters.\nThe code is available at:https://github.com/SonyResearch/MMDisCo.",
    "introduction": "Deep generative modeling has progressed rapidly in the last few years.\nDiffusion models are one of the keys to this progress, and they can be applied to various tasks, including image, audio, and video generation(Yang et al.,2023).\nFollowing the success of single-modal data, several attempts have been made to apply diffusion models to multimodal data(Bao et al.,2023).\nHowever, since multimodal data is more complex and harder to collect than single-modal data, developing multimodal generative models by simply extending single-modal ones remains challenging.\nOne promising way to alleviate this problem is to integrate several pre-trained single-modal models to build a multimodal generative model(Tang et al.,2023; Xing et al.,2024).\nAs there are numerous publicly-available models that can generate high-quality single-modal data(Rombach et al.,2022; Liu et al.,2023; Guo et al.,2024), their effective integration would substantially reduce the computational cost to build multimodal generative models.\nIn this work, we focus on audio-video joint generation on top of two pre-trained diffusion models for audio and video.\n\n\nThere are two approaches for audio-video generation that integrate several single-modal models: training-free and training-based.\nThe training-free approach employs pre-trained single-modal base generative models with their parameters fixed.\nIt uses an off-the-shelf recognition model to guide them to generate well-aligned samples across modalities(Xing et al.,2024).\nWhile this can improve multimodal alignment without any training cost, it may degrade the fidelity of a single modality(Xing et al.,2024).\nIn contrast, the training-based approach extends single-modal generative models for multimodal data by designing a neural network tailored to it(Ruan et al.,2023; Tang et al.,2023).\nAlthough this approach can achieve better performance in terms of both multimodal alignment and the fidelity of each single modality, it tends to require a significant computational cost for training.\nMore importantly, their architectures for handling multimodal data heavily depend on those of base models (i.e., they are not model-agnostic).\nTherefore, when updating base models, we must manually redesign them, which requires a lot of trial and error.\nIn short, the existing two approaches involve a trade-off between the quality of generated samples and model dependency, which increases the computational cost.\n\n\nIn this paper, we propose a novel method that is training-based but model-agnostic.\nOur method does not require backpropagation through the base models for the optimization.\nSpecifically, we introduce a lightweight joint guidance module on top of audio and video base models that adjust their outputs for audio-video joint generation.\nWe assume that pre-trained base models are black box diffusion models (i.e., we can access only their outputs and do not depend on a specific architecture design like a cross-attention module to construct a joint generation model).\nWe formulate the joint generation process as an extension of the classifier guidance (C-guide) for single-modal data(Song et al.,2021; Dhariwal and Nichol,2021).\nWe show that this joint guidance can be computed through the gradient of the optimal discriminator that distinguishes real audio-video pairs from the fake ones independently generated by base models.\nWe only train the discriminator with proper regularization inspired by Denoising Likelihood Score Matching (DLSM)(Chao et al.,2022).\nExtensive experiments on several benchmark datasets demonstrate that our proposed method can efficiently integrate single-modal base models for audio and video into a joint generation model, maintaining the performance of each single-modal generation without incurring a significant computational cost (seeSectionA.9).\n\n",
    "related_work": "2.1Audio-video joint generation by diffusion modelsSince an audio-video pair is one of the most popular types of multimodal data, several works train diffusion models with such pairs to achieve a conditional single-modal generation: video-conditional audio generation(Luo et al.,2023; Mo et al.,2023; Su et al.,2023; Pascual et al.,2024)or audio-conditional video generation(Jeong et al.,2023; Lee et al.,2023; Yariv et al.,2024).\nHowever, these works mainly focus on a single modality as a generation target.\nExtending these works to the joint generation of audio and video is not trivial due to the high dimensionality and heterogeneous data structure of audio-video joint data.Joint generation of audio and video pairs has been addressed in only a few recent studies(Ruan et al.,2023; Tang et al.,2023; Xing et al.,2024).\nMM-Diffusion(Ruan et al.,2023)is a multimodal diffusion model specific for audio-video joint generation.\nWhile MM-Diffusion is trained from scratch on audio-video pairs, CoDi(Tang et al.,2023)integrates several pre-trained single-modal diffusion models by adopting environment encoders to share modality-specific information across modalities during the generation process.\nSince they adopt a novel architecture strongly tied to the main network of diffusion models, it is difficult to directly apply their method to other architectures, hindering its applicability.\nIn contrast, our method handles base models as black boxes and depends only on their outputs.\nTherefore, our method is widely applicable to any type of architecture used in base models.Xing et al. (2024)shares a similar motivation to ours in the sense that they achieve audio-video cooperative generation from pre-trained single-modal base models.\nGiven multimodal embedding models (e.g., ImageBind(Girdhar et al.,2023)), they utilize universal guidance(Bansal et al.,2023)to make the embeddings from two modalities close.\nAlthough their approach is model-agnostic and can be applied to any base model, their guidance roughly ensures semantic alignment in the space of embeddings learned by ImageBind.\nThus, it does not achieve sampling from the actual joint distribution of audio-video pairs.\nIn contrast, our method is theoretically grounded in adjusting the scores predicted from base models to the score of the joint distribution, explicitly achieving sampling from the joint distribution.\n\n\n2.2Guidance for pre-trained diffusion modelsThe guidance(Song et al.,2021; Dhariwal and Nichol,2021)provides a proper way to update intermediate representations at each generation step so that the generated samples satisfy a given condition, even when the model is not trained for that type of conditional generation.\nSince the guidance does not require additional training in diffusion models, it is widely used to control the generation process with additional conditional signals.\nC-guide(Song et al.,2021; Dhariwal and Nichol,2021)was proposed to guide an image generation model by the class label, such as\"cat\"or\"dog\", using an additionally trained classifier.\nSeveral works extend this to utilize off-the-shelf recognition models to achieve beyond class-conditional generation(Graikos et al.,2022; Bansal et al.,2023).\nOur proposed method is the first attempt to extend the C-guide for joint multimodal generation.\nWe derive our methodology from the theory of C-guide, enabling us to sample data from the joint distribution.The samples generated with C-guide may suffer from degraded quality without careful tuning of its scale(Dhariwal and Nichol,2021).Chao et al. (2022)denote this problem as the score mismatch issue, where the posterior scores estimated by a diffusion model and a classifier are unstable and deviate from the true ones.\nThey proposed DLSM loss to alleviate this problem.\nDLSM regularizes a classifier’s gradient to match the residual error of noise prediction by trained diffusion models, providing a stable gradient for the generation process.\nOur work can be seen as an extension of DLSM for multimodal generation.\nWe employ DLSM to stabilize the gradient of our discriminator.Kim et al. (2023)proposed Discriminator Guidance, which guides a Text-to-Image (T2I) model by a discriminator distinguishing real images from generated images to improve the quality of the generated samples by a pre-trained T2I model.\nAlthoughKim et al. (2023)share a similar concept to ours regarding using a discriminator to bridge the gap between the score predicted by a pre-trained model and the target score, our goal is to integrate single-modal models into a joint generation model.\nTheir method is particularly designed to handle the gap within a single modality, and it cannot be straightforwardly applied to the multimodal generation.\nIn contrast, our method is derived from directly bridging the gap between a single-modal distribution and a joint one.\nMoreover, we show a single discriminator can serve as a guidance module for both domains, resulting in minimal computational cost for developing a joint generation model.\n\n",
    "origin_citations": [
      "Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models.",
      "Diffava: Personalized text-to-audio generation with visual alignment.",
      "Physics-driven diffusion models for impact sound synthesis from videos.",
      "Masked generative video-to-audio transformers with enhanced synchronicity.",
      "The power of sound (tpos): Audio reactive video generation with stable diffusion.",
      "Aadiff: Audio-aligned video synthesis with text-to-image diffusion.",
      "Diverse and aligned audio-to-video generation via text-to-video model adaptation.",
      "Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation.",
      "Any-to-any generation via composable diffusion.",
      "Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners.",
      "Imagebind: One embedding space to bind them all.",
      "Universal guidance for diffusion models.",
      "Score-based generative modeling through stochastic differential equations.",
      "Diffusion models beat gans on image synthesis.",
      "Diffusion models as plug-and-play priors.",
      "Denoising likelihood score matching for conditional score-based data generation.",
      "Refining generative process with discriminator guidance in score-based diffusion models."
    ],
    "citations": [
      47254,
      260207,
      492472,
      907218,
      128910,
      116269,
      477779,
      140115,
      260820,
      463298,
      2236,
      488851,
      431228,
      165,
      81888,
      500479,
      242552
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 88,
      "Final_Score": 84.5,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 92,
      "Final_Score": 88.3,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 88\nSynthesis_Analysis: 92\nFinal_Score: 89"
    },
    "overall_score": 86.78,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "agbiPPuSeQ",
        "forum": "agbiPPuSeQ",
        "number": 5876,
        "title": "MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation",
        "normalized_title": "mmdisco multi modal discriminator guided cooperative diffusion for joint audio and video generation",
        "abstract": "This study aims to construct an audio-video generative model with minimal computational cost by leveraging pre-trained single-modal generative models for audio and video.\nTo achieve this, we propose a novel method that guides single-modal models to cooperatively generate well-aligned samples across modalities. \nSpecifically, given two pre-trained base diffusion models, we train a lightweight joint guidance module to adjust scores separately estimated by the base models to match the score of joint distribution over audio and video. \nWe show that this guidance can be computed using the gradient of the optimal discriminator, which distinguishes real audio-video pairs from fake ones independently generated by the base models. \nBased on this analysis, we construct a joint guidance module by training this discriminator.\nAdditionally, we adopt a loss function to stabilize the discriminator's gradient and make it work as a noise estimator, as in standard diffusion models. \nEmpirical evaluations on several benchmark datasets demonstrate that our method improves both single-modal fidelity and multimodal alignment with relatively few parameters.\nThe code is available at: [https://github.com/SonyResearch/MMDisCo](https://github.com/SonyResearch/MMDisCo).",
        "authors": [
          "Akio Hayakawa",
          "Masato Ishii",
          "Takashi Shibuya",
          "Yuki Mitsufuji"
        ],
        "decision": null,
        "pdate": 1737562431143,
        "cdate": 1727326205857
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "agbiPPuSeQ",
          "forum": "agbiPPuSeQ",
          "number": 5876,
          "title": "MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation",
          "normalized_title": "mmdisco multi modal discriminator guided cooperative diffusion for joint audio and video generation",
          "abstract": "This study aims to construct an audio-video generative model with minimal computational cost by leveraging pre-trained single-modal generative models for audio and video.\nTo achieve this, we propose a novel method that guides single-modal models to cooperatively generate well-aligned samples across modalities. \nSpecifically, given two pre-trained base diffusion models, we train a lightweight joint guidance module to adjust scores separately estimated by the base models to match the score of joint distribution over audio and video. \nWe show that this guidance can be computed using the gradient of the optimal discriminator, which distinguishes real audio-video pairs from fake ones independently generated by the base models. \nBased on this analysis, we construct a joint guidance module by training this discriminator.\nAdditionally, we adopt a loss function to stabilize the discriminator's gradient and make it work as a noise estimator, as in standard diffusion models. \nEmpirical evaluations on several benchmark datasets demonstrate that our method improves both single-modal fidelity and multimodal alignment with relatively few parameters.\nThe code is available at: [https://github.com/SonyResearch/MMDisCo](https://github.com/SonyResearch/MMDisCo).",
          "authors": [
            "Akio Hayakawa",
            "Masato Ishii",
            "Takashi Shibuya",
            "Yuki Mitsufuji"
          ],
          "decision": null,
          "pdate": 1737562431143,
          "cdate": 1727326205857
        }
      ]
    }
  },
  {
    "paper_id": 41446,
    "title": "UnSeg: One Universal Unlearnable Example Generator is Enough against All Image Segmentation",
    "abstract": "Image segmentation is a crucial vision task that groups pixels within an image into semantically meaningful segments, which is pivotal in obtaining a fine-grained understanding of real-world scenes. However, an increasing privacy concern exists regarding training large-scale image segmentation models on unauthorized private data. In this work, we exploit the concept of unlearnable examples to make images unusable to model training by generating and adding unlearnable noise into the original images. Particularly, we propose a novelUnlearnable Segmentation (UnSeg)framework to train a universal unlearnable noise generator that is capable of transforming any downstream images into their unlearnable version. The unlearnable noise generator is finetuned from the Segment Anything Model (SAM) via bilevel optimization on an interactive segmentation dataset towards minimizing the training error of a surrogate model that shares the same architecture with SAM but is trained from scratch. We empirically verify the effectiveness of UnSeg across 6 mainstream image segmentation tasks, 10 widely used datasets, and 7 different network architectures, and show that the unlearnable images can reduce the segmentation performance by a large margin. Our work provides useful insights into how to leverage foundation models in a data-efficient and computationally affordable manner to protect images against image segmentation models.†††Corresponding author.",
    "introduction": "With the growing popularity of large models, more and more data are being crawled and curated “freely\" into massive pre-training datasets to support large-scale pre-training. This has raised public concerns about the unauthorized usage of private data posed on the web for training large-scale deep learning models or even illegal purposes[1]. For example, it has been found that the startup company Clearview AI developed its commercial facial recognition models by illicitly scraping vast amounts of personal images from online social networks[23].\nThis has motivated researchers to develop proactive defense measures to prevent deep learning models from exploiting private data.\n\n\nOne promising technique is calledunlearnable examples(UEs)[25]which adds small unlearnable noise into images to make them unexploitable to deep neural networks (DNNs). In the context of image classification, the unlearnable noise was generated to reduce the error (or difficulty) of an image so as to trick the model into believing that there is nothing to learn from the image. When all the samples in a dataset are modified by unlearnable noise, they will become unexploitable to DNN training and thus are protected. Data protection techniques with a similar objective are also known asavailability attacks[58]orindiscriminate poisoning attacks[19].\nAlthough UEs have been extensively studied in image classification tasks, their effectiveness for more complex vision tasks such as image segmentation remains unclear.\n\n\nFigure 1:An illustration ofUnSegpipeline which transforms images into unlearnable examples with mask prompt to prevent the exploitation of segmentation models.\n\n\nIn this work, we aim to develop an effective UE generation method for image segmentation, a fine-grained vision task that segments the detailed elements in an image. Our work is largely motivated by the recent progress of the Segment Anything Model (SAM)[31]which demonstrates the possibility of large-scale object segmentation from daily images. Meanwhile, the recent advancement of vision-language models (VLMs) also alerts the risk of segmenting and interpreting the semantic content within the images we posted online[62,5,6,54,57]. These potential risks highlight the imperative to develop effective UEs against image segmentation models. Moreover, there is also an increasing need to protect sensitive objects such as faces, persons, buildings, or locations from being utilized to train commercial or even illegal segmentation models for malicious purposes.\n\n\nThere exist three key challenges for generating UEs for image segmentation: 1) data efficiency challenge, 2) generation efficiency challenge, and 3) transferability challenge. First, an effective UE generation method should learn to craft effective UEs based on a small number of images rather than existing large-scale image segmentation datasets, which refers to the data efficiency challenge. Second, when applied to protect private images, the method should be able to craft UEs directly without the need to optimize for each image, which is called the generation efficiency challenge. As for the transferability challenge, the UE generation method should stay effective when transferred to protect different downstream tasks and datasets. By examining existing UE generation methods designed for image classification, we find that none of them can address all three challenges. Specifically, gradient-based UE methods like UE[25], robust UE (RUE)[17], stable UE (SUE)[39], and transferable UE (TUE)[48]all fail to address the generation efficiency challenge. Generation-free UE methods like Synthetic Perturbations (SynPer)[58]are limited to classification UEs and thus cannot be directly applied to image segmentation. Furthermore, these methods all face transferability issues to different datasets, architectures, and training approaches, making them less suitable for image segmentation where the images and task scenarios are diverse and complex.\n\n\nIn this paper, we propose a novel UE generation framework calledUnlearnable Segmentation (UnSeg)to tackle the above three key challenges. UnSeg is a generative framework that finetunes the pre-trained SAM into a universal UE generator via bilevel min-min optimization. As shown in Figure1, UnSeg is the first interactive model capable of generating unlearnable noise for any object in an image. Furthermore, different from all previous methods, UnSeg requires no additional label information beyond the mask prompt for the object region to protect. Finetuned on a small-scale interactive segmentation dataset, the UE generator can be immediately and effectively applied to protect downstream image segmentation datasets.\n\n\nIn Summary, our main contributions are:•We propose a novel UE generation framework UnSeg for image segmentation to finetune a universal UE generator from pre-trained SAM. To the best of our knowledge, UnSeg is the first UE generation method developed to protect images from image segmentation models.•In UnSeg, we formulate the fine-tuning of UE generator as a novel interactive segmentation-based bilevel min-min optimization, which is defined on a small-scale interactive segmentation tasks and achieved by iteratively optimize a pre-trained SAM and a train-from-scratch SAM. We also propose an epsilon generalization technique to stabilize the finetuning using a smaller noise budgetϵitalic-ϵ\\epsilonitalic_ϵwhich can be directly scaled up to larger noise at inference time.•We conduct extensive experiments to evaluate the effectiveness, efficiency, and transferability of the noise generator finetuned by UnSeg across 6 image segmentation tasks, 10 datasets, and 7 network architectures. The image datasets protected by UnSeg can effectively evade the training of different image segmentation models, causing a significant performance drop, e.g., a 92% performance drop on the COCO instance segmentation task.\n\n",
    "related_work": "Image SegmentationThere exist different types of image segmentation tasks such as instance segmentation[21], semantic segmentation[9], and panoptic segmentation[30]. All these tasks group pixels within an image into multiple semantic segments or groups[35,44], but assign concepts of different granularity. The segmentation models adopted for different tasks may vary, mostly following a similar architecture. Particularly, existing image segmentation models can be categorized into two types: 1) pixel-based classification models such as DeepLab[9], U-Net[49], PSPNet[65], and SegFormer[55]; and 2) mask-based classification models such as Mask2Former[11]and its variants[61,26,34]. Recently, the Segment Anything Model (SAM)[31]has emerged as a foundation model for segmentation. Trained on the large-scale interactive dataset SA-1B[31], SAM demonstrates strong generalization capabilities in handling various types of visual segmentation tasks[42,4,33,7]. In this paper, we leverage the zero-shot capability of SAM to train a universal and transferable UE generator to protect images against mainstream image segmentation models.\n\n\nUnlearnable ExamplesThe concept of UEs was first introduced in[25], where small protective noise can be injected into the training dataset to prevent machine learning models from learning useful representations.\nThis was achieved by generating error-minimizing noise that can remove errors from the dataset such that the training model finds no error to minimize (learn).\nTargeted adversarial poisoning[16]has also been demonstrated to be an effective approach to creating such shortcuts to mislead model training.\nThe working mechanism of UEs was explained by later work as creating a “shortcut\" between the input and output using linearly separable features[58]. They further introduced the Synthetic Perturbations (SynPer) method to craft synthetic patterns as UEs. However, the linear separability has recently been shown by Pedro et al.[52,51]to be unnecessary for UEs.\nThere exist several remaining key challenges for the practical adoption of UEs for private data protection: 1) robustness to adversarial training[43], 2) transferability from supervised to unsupervised learning, 3) transferability to protect differently labeled data, and more importantly 4) the extension to broad vision tasks beyond image classification. The robustness of UEs to adversarial training has been effectively addressed by the RUE method introduced in[17]which directly minimizes the adversarial training loss. And, the transferability challenges across different learning paradigms and labeling granularities have been effectively addressed by the TUEs[48]and Unlearnable Clusters (UCs)[63]methods.\nDespite these advances, all existing UE methods are exclusively focused on image classification tasks, limiting their effectiveness to coarse-level vision tasks. Inspired by the potential of SAM and the increasing demand for protecting images against fine-grained probing and learning, in this work, we extend UEs to image segmentation tasks and propose a data-efficient and computationally affordable approach to turn SAM into a universal data protector.\n\n",
    "origin_citations": [
      "Mask r-cnn.",
      "Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs.",
      "Panoptic segmentation.",
      "Transformer-based visual segmentation: A survey.",
      "Image segmentation using deep learning: A survey.",
      "U-net: Convolutional networks for biomedical image segmentation.",
      "Pyramid scene parsing network.",
      "Segformer: Simple and efficient design for semantic segmentation with transformers.",
      "Masked-attention mask transformer for universal image segmentation.",
      "Mp-former: Mask-piloted transformer for image segmentation.",
      "Oneformer: One transformer to rule universal image segmentation.",
      "Mask dino: Towards a unified transformer-based framework for object detection and segmentation.",
      "Segment anything.",
      "Segment anything in medical images.",
      "Segment anything in 3d with nerfs.",
      "Lisa: Reasoning segmentation via large language model.",
      "Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model.",
      "Unlearnable examples: Making personal data unexploitable.",
      "Adversarial examples make strong poisons.",
      "Availability attacks create shortcuts.",
      "What can we learn from unlearnable datasets?",
      "Autoregressive perturbations for data poisoning.",
      "Towards deep learning models resistant to adversarial attacks.",
      "Robust unlearnable examples: Protecting data against adversarial learning.",
      "Transferable unlearnable examples.",
      "Unlearnable clusters: Towards label-agnostic unlearnable examples."
    ],
    "citations": [
      520,
      358502,
      1730,
      53948,
      8490,
      825,
      508,
      8842,
      152903,
      75925,
      32870,
      413005,
      2512,
      52794,
      1088897,
      182488,
      47510,
      38634,
      393968,
      278305,
      256564,
      412586,
      917,
      500316,
      336429
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 87,
      "Synthesis_Analysis": 92,
      "Final_Score": 87.3,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 87\nSynthesis_Analysis: 92\nFinal_Score: 88"
    },
    "overall_score": 86.74,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "c8HOQIMwKP",
        "forum": "c8HOQIMwKP",
        "number": 15614,
        "title": "UnSeg: One Universal Unlearnable Example Generator is Enough against All Image Segmentation",
        "normalized_title": "unseg one universal unlearnable example generator is enough against all image segmentation",
        "abstract": "Image segmentation is a crucial vision task that groups pixels within an image into semantically meaningful segments, which is pivotal in obtaining a fine-grained understanding of real-world scenes. However, an increasing privacy concern exists regarding training large-scale image segmentation models on unauthorized private data. In this work, we exploit the concept of unlearnable examples to make images unusable to model training by generating and adding unlearnable noise into the original images. Particularly, we propose a novel Unlearnable Segmentation (UnSeg) framework to train a universal unlearnable noise generator that is capable of transforming any downstream images into their unlearnable version. The unlearnable noise generator is finetuned from the Segment Anything Model (SAM) via bilevel optimization on an interactive segmentation dataset towards minimizing the training error of a surrogate model that shares the same architecture with SAM (but trains from scratch). We empirically verify the effectiveness of UnSeg across 6 mainstream image segmentation tasks, 10 widely used datasets, and 7 different network architectures, and show that the unlearnable images can reduce the segmentation performance by a large margin. Our work provides useful insights into how to leverage foundation models in a data-efficient and computationally affordable manner to protect images against image segmentation models.",
        "authors": [
          "Ye Sun",
          "Hao Zhang",
          "Tiehua Zhang",
          "Xingjun Ma",
          "Yu-Gang Jiang"
        ],
        "decision": null,
        "pdate": 1727288102298,
        "cdate": 1715762101019
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "c8HOQIMwKP",
          "forum": "c8HOQIMwKP",
          "number": 15614,
          "title": "UnSeg: One Universal Unlearnable Example Generator is Enough against All Image Segmentation",
          "normalized_title": "unseg one universal unlearnable example generator is enough against all image segmentation",
          "abstract": "Image segmentation is a crucial vision task that groups pixels within an image into semantically meaningful segments, which is pivotal in obtaining a fine-grained understanding of real-world scenes. However, an increasing privacy concern exists regarding training large-scale image segmentation models on unauthorized private data. In this work, we exploit the concept of unlearnable examples to make images unusable to model training by generating and adding unlearnable noise into the original images. Particularly, we propose a novel Unlearnable Segmentation (UnSeg) framework to train a universal unlearnable noise generator that is capable of transforming any downstream images into their unlearnable version. The unlearnable noise generator is finetuned from the Segment Anything Model (SAM) via bilevel optimization on an interactive segmentation dataset towards minimizing the training error of a surrogate model that shares the same architecture with SAM (but trains from scratch). We empirically verify the effectiveness of UnSeg across 6 mainstream image segmentation tasks, 10 widely used datasets, and 7 different network architectures, and show that the unlearnable images can reduce the segmentation performance by a large margin. Our work provides useful insights into how to leverage foundation models in a data-efficient and computationally affordable manner to protect images against image segmentation models.",
          "authors": [
            "Ye Sun",
            "Hao Zhang",
            "Tiehua Zhang",
            "Xingjun Ma",
            "Yu-Gang Jiang"
          ],
          "decision": null,
          "pdate": 1727288102298,
          "cdate": 1715762101019
        }
      ]
    }
  },
  {
    "paper_id": 12024,
    "title": "ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy",
    "abstract": "Modern computer vision offers a great variety of models to practitioners, and selecting a model from multiple options for specific applications can be challenging. Conventionally, competing model architectures and training protocols are compared by their classification accuracy on ImageNet. However, this single metric does not fully capture performance nuances critical for specialized tasks. In this work, we conduct an in-depth comparative analysis of model behaviors beyond ImageNet accuracy, for both ConvNet and Vision Transformer architectures, each across supervised and CLIP training paradigms. Although our selected models have similar ImageNet accuracies and compute requirements, we find that they differ in many other aspects: types of mistakes, output calibration, transferability, and feature invariance, among others. This diversity in model characteristics, not captured by traditional metrics, highlights the need for more nuanced analysis when choosing among different models. Code is available atgithub.com/kirill-vish/Beyond-INet.",
    "introduction": "The computer vision model landscape has become increasingly complex. From early ConvNets(LeCun et al.,1998)to advances in Vision Transformers(Dosovitskiy et al.,2020), the variety of models available has expanded significantly. Similarly, training paradigms have evolved from supervised training on ImageNet(Deng et al.,2009)to self-supervised learning(Chen et al.,2020; He et al.,2020)and image-text pair training like CLIP(Radford et al.,2021). While signaling progress, this explosion of choices poses a significant challenge for practitioners: how to select a model that suits their purposes?\n\n\nFigure 1:Models are often compared only by their ImageNet accuracy, without looking at many other important behaviors. In our work, we analyze models with similar ImageNet accuracies and find that they have vastly different properties.\n\n\nConventionally, ImageNet accuracy has served as the primary metric for evaluating model performance. It has driven remarkable progress since it ignited the deep learning revolution(Krizhevsky et al.,2012). However, this metric is becoming increasingly insufficient. While ImageNet is useful to measure a model’s general capability, it does not capture the nuanced differences arising from varying architectures, training paradigms, and data – models with different properties may appear similar if judged solely based on ImageNet accuracy (Fig.1). This limitation becomes more pronounced as models start to overfit the idiosyncrasies of ImageNet with saturated accuracies(Beyer et al.,2020).\n\n\nA particularly noteworthy example is CLIP. Despite having a similar ImageNet accuracy as a ResNet(He et al.,2016), CLIP’s vision encoder exhibits much better robustness and transferability. This has sparked research that explores and builds upon the unique strengths of CLIP(Ramesh et al.,2022; Luo et al.,2022; Wortsman et al.,2022; Vinker et al.,2023), which were not evident from the ImageNet metric alone. This demonstrates that analyzing alternative properties could help discover useful models.\n\n\nModelArchitecturePretrainingFinetuningParadigmFLOPs#ParamINet-1K val%ViT-supViT-B/16ImageNet-21KImageNet-1Ksupervised17.5G87M85.5ConvNeXt-supConvNeXt-BImageNet-21KImageNet-1Ksupervised15.4G89M85.5ViT-clipViT-B/16LAION-400M—CLIP17.5G87M67.0ConvNeXt-clipConvNeXt-BLAION-400M—CLIP15.4G89M66.3Table 1:Model summary in our analysis.We select ConvNeXt and ViT with similar ImageNet accuracies within each training paradigm.\n\n\nIn addition to fundamental research, the growing integration of vision models into production systems also calls for a deep understanding of their behaviors. Conventional metrics do not fully capture models’ ability to handle real-world vision challenges like varying camera poses, lighting conditions, or occlusions. For instance, models trained on datasets such as ImageNet often struggle(Yamada & Otani,2022)to transfer their performance to real-world applications where conditions and scenarios are more diverse.\n\n\nTo bridge this gap, we conduct an in-depth exploration focusing on model behaviors beyond ImageNet accuracy. We analyze four leading models in the computer vision: ConvNeXt(Liu et al.,2022), as a representative ConvNet, and Vision Transformer (ViT)(Dosovitskiy et al.,2020), each under supervised and CLIP training. The selected models are similar in parameter counts and show nearly identical accuracy on ImageNet-1K within each training paradigm, ensuring a fair comparison. Our study delves into a wide array of model characteristics, such as types of prediction errors, generalization capabilities, invariances of the learned representations, calibration, and many others. Importantly, our focus is on properties exhibited by the model without additional training or finetuning, providing insights for practitioners interested in using pretrained models directly.\n\n\nIn our analysis, we discover substantial variations in model behaviors among different architectures and training paradigms. For example, CLIP models make fewer classification errors relative to their ImageNet performance. However, supervised models are better calibrated and superior on ImageNet robustness benchmarks. ConvNeXt has an advantage on synthetic data but is more texture-biased than ViT. We also find that supervised ConvNeXt excels on many benchmarks and achieves transferability comparable to that of CLIP models. Based on these findings, it becomes evident that various models demonstrate their strengths in unique ways that are not captured by a single metric. Our research emphasizes the need for more detailed evaluation metrics for accurate, context-specific model selection and the creation of new benchmarks unrelated to ImageNet.\n\n",
    "related_work": "Architecture analysis.Several works compared ViTs and ConvNeXt from the perspective of internal representations(Raghu et al.,2021), synthetic data(Ruiz et al.,2022), transferability(Zhou et al.,2021), and robustness(Wang et al.,2022; Bai et al.,2021; Pinto et al.,2022; Djolonga et al.,2021). Other studies included analysis of Transformer properties(Naseer et al.,2021)and impact of neural network width and depth on learned representations(Nguyen et al.,2020). ViTs and ConvNets were also evaluated on ImageNet, showing that Transformers are more aligned with human error patterns(Tuli et al.,2021). A large variety of backbones, trained with various methods, were benchmarked in(Goldblum et al.,2024)across a diverse set of computer vision tasks, including classification, detection, and retrieval. In contrast to studies that analyze a single property, our work extensively compares models across many, maintaining a fair comparison by evaluating models with similar ImageNet accuracies.\n\n\nTraining objective analysis.A comprehensive analysis was conducted in(Walmer et al.,2023), comparing ViTs trained with supervised, self-supervised, and CLIP objectives. Analysis of the representations of models trained with supervised and self-supervised objectives was presented in(Grigg et al.,2021; Gwilliam & Shrivastava,2022). Two works(Park et al.,2023; Shekhar et al.,2023)focused on investigating the effect of training objective in self-supervised learning. Unlike studies emphasizing self-supervised models, our work compares supervised and CLIP models.\n\n\nLimitations of ImageNet.Recent research(Beyer et al.,2020; Recht et al.,2019; Tsipras et al.,2020; Yun et al.,2021)highlighted issues with the reliability and quality of ImageNet labels. Two studies(Kornblith et al.,2019; Miller et al.,2021)showed a strong relationship between performance on ImageNet and other datasets, although this can depend on the model’s architecture and training methods.\nOther works(Richards et al.,2023; Fang et al.,2023b)showed that high ImageNet accuracy does not guarantee good performance on diverse datasets. Current robustification training techniques were found to overfit(Yamada & Otani,2022)to ImageNet evaluations. In addition, ImageNet suffers from dichotomous data difficulty(Meding et al.,2021), obscuring differences between models. Our analysis does not directly address data-related problems of ImageNet but instead studies alternative properties.\n\n",
    "origin_citations": [
      "Do vision transformers see like convolutional neural networks?",
      "Finding differences between transformers and convnets using counterfactual simulation testing.",
      "Convnets vs. transformers: Whose visual representations are more transferable?",
      "Can cnns be more robust than transformers?",
      "Are transformers more robust than cnns?",
      "An impartial take to the cnn vs transformer robustness contest.",
      "On robustness and transferability of convolutional neural networks.",
      "Intriguing properties of vision transformers.",
      "Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth.",
      "Are convolutional neural networks or transformers more like human vision?",
      "Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks.",
      "Teaching matters: Investigating the role of supervision in vision transformers.",
      "Do self-supervised and supervised methods learn similar visual representations?",
      "Beyond supervised vs. unsupervised: Representative benchmarking and analysis of image representation learning.",
      "What do self-supervised vision transformers learn?",
      "Objectives matter: Understanding the impact of self-supervised objectives on vision transformer representations.",
      "Are we done with imagenet?",
      "Do imagenet classifiers generalize to imagenet?",
      "From ImageNet to image classification: Contextualizing progress on benchmarks.",
      "Re-labeling imagenet: from single to multi-labels, from global to localized labels.",
      "Do better imagenet models transfer better?",
      "Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization.",
      "Does progress on object recognition benchmarks improve real-world generalization?",
      "Does progress on imagenet transfer to real-world datasets?",
      "Does robustness on imagenet transfer to downstream tasks?",
      "Imagenet suffers from dichotomous data difficulty."
    ],
    "citations": [
      5243,
      242222,
      94463,
      412691,
      276485,
      420754,
      522469,
      5807,
      434842,
      9053,
      311198,
      222384,
      343875,
      82233,
      117788,
      52467,
      349998,
      1627,
      1623,
      38582,
      74484,
      371533,
      468701,
      302453,
      64576
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 85,
      "Final_Score": 83.6,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 85\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 88,
      "Final_Score": 88.4,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 95\nSynthesis_Analysis: 88\nFinal_Score: 89"
    },
    "overall_score": 86.47999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "9BGi9PEhNn",
        "forum": "9BGi9PEhNn",
        "number": 2345,
        "title": "ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy",
        "normalized_title": "convnet vs transformer supervised vs clip beyond imagenet accuracy",
        "abstract": "Modern computer vision offers a great variety of models to practitioners, and selecting a model from multiple options for specific applications can be challenging. Conventionally, competing model architectures and training protocols are compared by their classification accuracy on ImageNet. However, this single metric does not fully capture performance nuances critical for specialized tasks. In this work, we conduct an in-depth comparative analysis of model behaviors beyond ImageNet accuracy, for both ConvNet and Vision Transformer architectures, each across supervised and CLIP training paradigms. Although our selected models have similar ImageNet accuracies and compute requirements, we find that they differ in many other aspects: types of mistakes, output calibration, transferability, and feature invariance, among others. This diversity in model characteristics, not captured by traditional metrics, highlights the need for more nuanced analysis when choosing among different models.",
        "authors": [
          "Kirill Vishniakov",
          "Zhiqiang Shen",
          "Zhuang Liu"
        ],
        "decision": null,
        "pdate": 1714610310093,
        "cdate": 1706390225889
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "9BGi9PEhNn",
          "forum": "9BGi9PEhNn",
          "number": 2345,
          "title": "ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy",
          "normalized_title": "convnet vs transformer supervised vs clip beyond imagenet accuracy",
          "abstract": "Modern computer vision offers a great variety of models to practitioners, and selecting a model from multiple options for specific applications can be challenging. Conventionally, competing model architectures and training protocols are compared by their classification accuracy on ImageNet. However, this single metric does not fully capture performance nuances critical for specialized tasks. In this work, we conduct an in-depth comparative analysis of model behaviors beyond ImageNet accuracy, for both ConvNet and Vision Transformer architectures, each across supervised and CLIP training paradigms. Although our selected models have similar ImageNet accuracies and compute requirements, we find that they differ in many other aspects: types of mistakes, output calibration, transferability, and feature invariance, among others. This diversity in model characteristics, not captured by traditional metrics, highlights the need for more nuanced analysis when choosing among different models.",
          "authors": [
            "Kirill Vishniakov",
            "Zhiqiang Shen",
            "Zhuang Liu"
          ],
          "decision": null,
          "pdate": 1714610310093,
          "cdate": 1706390225889
        }
      ]
    }
  },
  {
    "paper_id": 18005,
    "title": "Evaluation of LLMs on Syntax-Aware Code Fill-in-the-Middle Tasks",
    "abstract": "We introduceSyntax-AwareFill-in-the-Middle (\\nolbreaksSAFIM), a new benchmark for evaluating Large Language Models (LLMs) on the code Fill-in-the-Middle (FIM) task. This benchmark focuses on syntax-aware completions of program structures such as code blocks and conditional expressions, and includes 17,720 examples from multiple programming languages, sourced from recent code submissions after April 2022 to minimize data contamination. SAFIM provides a robust framework with various prompt designs and novel syntax-aware post-processing techniques, facilitating accurate and fair comparisons across LLMs. Our comprehensive evaluation of 15 LLMs shows that FIM pretraining not only enhances FIM proficiency but also improves Left-to-Right (L2R) inference using LLMs. Our findings challenge conventional beliefs and suggest that pretraining methods and data quality have more impact than model size.\\nolbreaksSAFIM thus serves as a foundational platform for future research in effective pretraining strategies for code LLMs.\nThe evaluation toolkit and dataset are available athttps://github.com/gonglinyuan/safim, and the leaderboard is available athttps://safimbenchmark.com.",
    "introduction": "Recent advances in Large Language Models (LLMs) such as GPT-3.5(Ouyang et al.,2022), GPT-4(OpenAI,2023), and CodeLLaMa(Rozière et al.,2023)have revolutionized coding-related tasks. However, existing benchmarks like HumanEval(Chen et al.,2021a)and MBPP(Austin et al.,2021)focus on generating standalone functions or single-file code from natural language descriptions, and do not consider the more common practice of modifying and expanding existing code during development.\n\n\nRecognizing this gap, we introduce theSyntax-AwareFill-in-the-Middle (\\nolbreaksSAFIM) benchmark.\\nolbreaksSAFIM emphasizes syntax-aware completion within code’s Abstract Syntax Tree (AST), targeting algorithmic blocks, control-flow expressions, and API function calls, unlike existing Fill-in-the Middle (FIM) benchmarks such as HumanEval-Infilling(Bavarian et al.,2022), which are based on filling randomly masked lines or character spans.\\nolbreaksSAFIM is sourced from code on Codeforces and GitHub created after April 2022, deliberately aiming to avoid overlap with mainstream open-source pretraining corpora like The Stack(Kocetkov et al.,2022). This approach reduces the risks of data contamination caused by memoization of test cases, thereby bolstering the credibility of our results.\\nolbreaksSAFIM, with its 17,720 examples from 8,590 code files, not only surpasses the scale of HumanEval-Infilling, which draws from 164 short code files, but also expands the scope to include multiple programming languages.\\nolbreaksSAFIM primarily relies on execution-based evaluation, and uses syntactical match evaluation only when execution is not feasible due to external API calls.\n\n\nOur comprehensive evaluation of 15 LLMs on\\nolbreaksSAFIM reveals its effectiveness in providing a fair comparison of models. We implement five distinct prompt designs to accommodate various model types and introduce a syntax-aware truncation algorithm for post-processing the outputs. Our approach unveils the true capabilities of non-FIM-trained models, allowing for a fair comparison with FIM-trained models.\n\n\nMoreover,\\nolbreaksSAFIM sheds light on the strengths of various pretraining paradigms and challenges some prevalent beliefs in the field. Specifically, our findings suggest that FIM pretraining not only improves LLMs’ performance in FIM inference but also enhances their performance in classical Left-to-Right (L2R) inference scenarios. This supports the growing trend of using FIM as the primary pretraining objective in code LLM development.\nWe also observe that pretraining methods and data quality often outweigh the sheer model size—smaller models with sophisticated pretraining paradigms often outperform larger models. This is particularly evident in task-specific performances on\\nolbreaksSAFIM, where models pretrained with additional repo-level information excel in API function call completion, while those trained with code execution feedback perform better in control-flow expression generation. However, it is crucial to note that these comparisons across different model families are not controlled experiments and could be influenced by differences in pretraining environments. This suggests future work in pretraining such models under the same environment to validate these observations further. That said, our benchmark,\\nolbreaksSAFIM, provides a solid foundation for such future research, and opens up new opportunities in designing effective pretraining and fine-tuning paradigms for code LLMs.\n\n",
    "related_work": "Large Language Models for Code.The emergence of Large Language Models (LLMs) like GPT-3(Brown et al.,2020)in natural language processing has led to the understanding that merely increasing the number of parameters in pretrained language models will ensure superior performance on unseen tasks. This has led to the application of LLMs to code-related tasks, particularly in code generation. For such tasks, decoder-only models are typically used. Initially, these models, such as Codex(Chen et al.,2021a), PaLM(Chowdhery et al.,2022), PolyCoder(Xu et al.,2022), and CodeGen(Nijkamp et al.,2023), primarily focused on Left-to-Right (L2R) pretraining, a.k.a. “Next Token Prediction.” However, the Fill-in-the-Middle (FIM) objective, a.k.a. “Infilling,” has become increasingly popular, with models like InCoder(Fried et al.,2023), StarCoder(Li et al.,2023), SantaCoder(Allal et al.,2023), DeepSeek-Coder(Guo et al.,2024), and CodeLLaMa(Rozière et al.,2023)showing their effectiveness. Additionally, proprietary models such as GPT-3.5(Ouyang et al.,2022), GPT-4(OpenAI,2023), and Gemini(Team et al.,2023), which use undisclosed pretraining methods, also contribute to this domain. While GLM-like models(Du et al.,2022)or encoder-decoder models, including CodeGeeX(Zheng et al.,2023), PLBART(Ahmad et al.,2021), AlphaCode(Li et al.,2022), CodeT5(Wang et al.,2021,2023a), and AST-T5(Gong et al.,2024a)exist, they are outside of our paper’s scope. Our paper evaluates a select group of these LMs using the\\nolbreaksSAFIM benchmark. We develop insights into their performance in code FIM tasks, explore the strengths and weaknesses of various pretraining paradigms, and challenge the prevailing belief that a larger number of parameters automatically leads to better performance.\n\n\nBenchmarking Generative Code LLMs.Existing benchmarks for code generation in LLMs have a gap in effectively evaluating code generation capability for real-world development. Widely-used benchmarks like HumanEval(Chen et al.,2021a)and MBPP(Austin et al.,2021)are limited to single Python functions and also subject to data contamination(Yang et al.,2023). Extensions like HumanEval-X(Zheng et al.,2023), MultiPLe(Cassano et al.,2022), and MBXP(Athiwaratkun et al.,2023)expand these benchmarks to other programming languages. Competition-style coding benchmarks like APPS(Hendrycks et al.,2021)and CodeContests(Li et al.,2022), broaden the scope to file-level code generation. However, they still do not reflect typical development, which often involves iterative codebase expansion and invoking external API libraries. On the other hand, contextually richer benchmarks, such as JuICe(Agashe et al.,2019), DS-1000(Lai et al.,2022), ARCADE(Yin et al.,2022), NumpyEval(Zhang et al.,2023b), and PandasEval(Jain et al.,2021), PlotCoder(Chen et al.,2021b), ADELT(Gong et al.,2024b)in data science, and APIBench(Patil et al.,2023), RepoBench(Liu et al.,2023), ODEX(Wang et al.,2023b), SWE-Bench(Jimenez et al.,2023), GoogleCodeRepo(Shrivastava et al.,2023), RepoEval(Zhang et al.,2023a), and CoCoMIC-Data(Ding et al.,2023)in software engineering, are often very small, heavily reliant on imperfect match-based evaluation metrics, or lacking in execution-based evaluation. Our\\nolbreaksSAFIM benchmark, based on Fill-in-the-Middle (FIM) tasks, bridges this gap by providing a comprehensive evaluation framework.\n\n\nFill-in-the-Middle in Training and Evaluating Code LLMs.Fill-in-the-Middle (FIM) originates from masked language modeling (MLM) for training encoder-only models(Devlin et al.,2019)and T5-style span corruption for training encoder-decoder models(Raffel et al.,2020), with span lengths usually limited to 1 to 5 tokens, with the goal of targeting representation learning rather than generation. For coding tasks, InCoder(Fried et al.,2023)shows the effectiveness of FIM as a pretraining objective for decoder-only models.Fried et al. (2023)further establishes the HumanEval-Infilling benchmark, further explored byBavarian et al. (2022)in evaluating GPT-3/Codex variants, showing that a pretraining mix with a 90% FIM ratio does not harm Left-to-Right (L2R) generation performance. CodeLLaMa’s evaluations on HumanEval-Infilling support these findings, underscoring the value of FIM in pretraining code-focused LLMs(Rozière et al.,2023). However, this benchmark, limited to the 164 tiny Python snippets of HumanEval, emphasize the need for a more robust benchmark.\\nolbreaksSAFIM addresses this need by introducing a comprehensive, syntax-aware FIM benchmark for more detailed evaluations.\n\n",
    "origin_citations": [
      "Language models are few-shot learners.",
      "Evaluating large language models trained on code.",
      "PaLM: Scaling language modeling with pathways.",
      "A systematic evaluation of large language models of code.",
      "Codegen: An open large language model for code with multi-turn program synthesis.",
      "InCoder: A generative model for code infilling and synthesis.",
      "StarCoder: may the source be with you!",
      "SantaCoder: don’t reach for the stars!",
      "DeepSeek-Coder: When the large language model meets programming – the rise of code intelligence.",
      "Code Llama: Open foundation models for code.",
      "Training language models to follow instructions with human feedback.",
      "GPT-4 technical report.",
      "Gemini: A family of highly capable multimodal models.",
      "GLM: General language model pretraining with autoregressive blank infilling.",
      "CodeGeeX: A pre-trained model for code generation with multilingual evaluations on humaneval-x.",
      "Unified pre-training for program understanding and generation.",
      "Competition-level code generation with AlphaCode.",
      "CodeT5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation.",
      "CodeT5+: Open code large language models for code understanding and generation.",
      "AST-T5: Structure-aware pretraining for code generation and understanding.",
      "Program synthesis with large language models.",
      "Rethinking benchmark and contamination for language models with rephrased samples.",
      "MultiPL-E: A scalable and extensible approach to benchmarking neural code generation.",
      "Multi-lingual evaluation of code generation models.",
      "Measuring coding challenge competence with APPS.",
      "JuICe: A large scale distantly supervised dataset for open domain context-based code generation.",
      "DS-1000: A natural and reliable benchmark for data science code generation.",
      "Natural language to code generation in interactive data science notebooks.",
      "ToolCoder: Teach code generation models to use api search tools.",
      "Jigsaw: Large language models meet program synthesis.",
      "PlotCoder: Hierarchical decoding for synthesizing visualization code in programmatic context.",
      "ADELT: Transpilation between deep learning frameworks.",
      "Gorilla: Large language model connected with massive apis.",
      "RepoBench: Benchmarking repository-level code auto-completion systems.",
      "Execution-based evaluation for open-domain code generation.",
      "SWE-Bench: Can language models resolve real-world Github issues?",
      "Repository-level prompt generation for large language models of code.",
      "RepoCoder: Repository-level code completion through iterative retrieval and generation.",
      "CoCoMIC: Code completion by jointly modeling in-file and cross-file context.",
      "BERT: Pre-training of deep bidirectional transformers for language understanding.",
      "Exploring the limits of transfer learning with a unified text-to-text transformer.",
      "Efficient training of language models to fill in the middle."
    ],
    "citations": [
      679,
      7465,
      1554,
      2980,
      2560,
      63708,
      7561,
      302951,
      124306,
      404859,
      364,
      9115,
      8028,
      2212,
      492180,
      216280,
      89441,
      172678,
      114356,
      485221,
      1573,
      294812,
      68987,
      334070,
      3525,
      415753,
      245054,
      140217,
      116253,
      152232,
      77716,
      258749,
      306930,
      139714,
      321181,
      80197,
      494518,
      139925,
      7,
      9,
      419563
    ],
    "extract_rate": 0.9761904761904762,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 82,
      "Final_Score": 82.69999999999999,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 82\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 88,
      "Final_Score": 89.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 95\nSynthesis_Analysis: 88\nFinal_Score: 89"
    },
    "overall_score": 86.47999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "jKYyFbH8ap",
        "forum": "jKYyFbH8ap",
        "number": 8983,
        "title": "Evaluation of LLMs on Syntax-Aware Code Fill-in-the-Middle Tasks",
        "normalized_title": "evaluation of llms on syntax aware code fill in the middle tasks",
        "abstract": "We introduce **S**yntax-**A**ware **F**ill-**i**n-the-**M**iddle (SAFIM), a new benchmark for evaluating Large Language Models (LLMs) on the code Fill-in-the-Middle (FIM) task. This benchmark focuses on syntax-aware completions of program structures such as code blocks and conditional expressions, and includes 17,720 examples from multiple programming languages, sourced from recent code submissions after April 2022 to minimize data contamination. SAFIM provides a robust framework with various prompt designs and novel syntax-aware post-processing techniques, facilitating accurate and fair comparisons across LLMs. Our comprehensive evaluation of 15 LLMs shows that FIM pretraining not only enhances FIM proficiency but also improves Left-to-Right (L2R) inference using LLMs. Our findings challenge conventional beliefs and suggest that pretraining methods and data quality have more impact than model size. SAFIM thus serves as a foundational platform for future research in effective pretraining strategies for code LLMs. The evaluation toolkit and dataset are available at https://github.com/gonglinyuan/safim, and the leaderboard is available at https://safimbenchmark.com.",
        "authors": [
          "Linyuan Gong",
          "Sida Wang",
          "Mostafa Elhoushi",
          "Alvin Cheung"
        ],
        "decision": null,
        "pdate": 1714610482546,
        "cdate": 1706867715263
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "jKYyFbH8ap",
          "forum": "jKYyFbH8ap",
          "number": 8983,
          "title": "Evaluation of LLMs on Syntax-Aware Code Fill-in-the-Middle Tasks",
          "normalized_title": "evaluation of llms on syntax aware code fill in the middle tasks",
          "abstract": "We introduce **S**yntax-**A**ware **F**ill-**i**n-the-**M**iddle (SAFIM), a new benchmark for evaluating Large Language Models (LLMs) on the code Fill-in-the-Middle (FIM) task. This benchmark focuses on syntax-aware completions of program structures such as code blocks and conditional expressions, and includes 17,720 examples from multiple programming languages, sourced from recent code submissions after April 2022 to minimize data contamination. SAFIM provides a robust framework with various prompt designs and novel syntax-aware post-processing techniques, facilitating accurate and fair comparisons across LLMs. Our comprehensive evaluation of 15 LLMs shows that FIM pretraining not only enhances FIM proficiency but also improves Left-to-Right (L2R) inference using LLMs. Our findings challenge conventional beliefs and suggest that pretraining methods and data quality have more impact than model size. SAFIM thus serves as a foundational platform for future research in effective pretraining strategies for code LLMs. The evaluation toolkit and dataset are available at https://github.com/gonglinyuan/safim, and the leaderboard is available at https://safimbenchmark.com.",
          "authors": [
            "Linyuan Gong",
            "Sida Wang",
            "Mostafa Elhoushi",
            "Alvin Cheung"
          ],
          "decision": null,
          "pdate": 1714610482546,
          "cdate": 1706867715263
        }
      ]
    }
  },
  {
    "paper_id": 26334,
    "title": "Disentangling and mitigating the impact of task similarity for continual learning",
    "abstract": "Continual learning of partially similar tasks poses a challenge for artificial neural networks, as task similarity presents both an opportunity for knowledge transfer and a risk of interference and catastrophic forgetting.\nHowever, it remains unclear how task similarity in input features and readout patterns influences knowledge transfer and forgetting, as well as how they interact with common algorithms for continual learning.\nHere, we develop a linear teacher-student model with latent structure and show analytically that high input feature similarity coupled with low readout similarity is catastrophic for both knowledge transfer and retention.\nConversely, the opposite scenario is relatively benign.\nOur analysis further reveals that task-dependent activity gating improves knowledge retention at the expense of transfer, while task-dependent plasticity gating does not affect either retention or transfer performance at the over-parameterized limit.\nIn contrast, weight regularization based on the Fisher information metric significantly improves retention, regardless of task similarity, without compromising transfer performance. Nevertheless, its diagonal approximation and regularization in the Euclidean space are much less robust against task similarity.\nWe demonstrate consistent results in a permuted MNIST task with latent variables. Overall, this work provides insights into when continual learning is difficult and how to mitigate it.",
    "introduction": "Artificial neural networks surpass human capabilities in various domains, yet struggle with continual learning. These networks tend to forget previously learned tasks when trained sequentially—a problem known as catastrophic forgetting[42,15,22,31].\nThis phenomenon affects not only supervised training of feedforward networks but also extends to recurrent neural networks[10], reinforcement learning tasks[29], and fine-tuning of large language models[39].\nMany algorithms for mitigating catastrophic forgetting have been developed previously, including rehearsal techniques[46,47,57], weight regularization[29,35,65], and activity-gating methods[14,52,40,54], among others[55,48,61,20].\nHowever, these methods often hinder forward and backward knowledge transfer[27,38,28], and thus it remains unclear how to achieve knowledge transfer and retention simultaneously.\n\n\nA key factor for continual learning is task similarity.\nIf two subsequent tasks are similar, there is a potential for a knowledge transfer from one task to another, but the risk of interference also becomes high[45,27,34,11,38].\nThe impact of task similarity on transfer and retention performance is particularly complicated because two tasks can be similar in different manners[63,31].\nSometimes familiar input features need to be associated with novel output patterns, but at other times, novel input features need to be associated with familiar output patterns. Previous works observed that these two scenarios influence continual learning differently[34], yet the impact of the input and output similarity on knowledge transfer and retention has not been well understood.\nMoreover, it remains unknown how the task similarity interacts with algorithms for continual learning such as activity gating or weight regularization.\n\n\nTo gain insight into these questions, in this work, we investigate how transfer and retention performance depend on task similarity, task-dependent gating, and weight regularization in analytically tractable teacher-student models.\nTeacher-student models are simple, typically linear, neural networks in which the generative model of data is specified explicitly by the teacher network[17,64,3].\nThese models have provided tremendous insights into generalization property[53,43,18,1], convergence rate[59,56,37], and learning dynamics[49,50,4,24]of neural networks, due to their analytical tractability.\nSeveral works also studied continual learning using teacher-student settings[2,34,26,11,19,36,12](see Related works section for details).\n\n\nWe develop a linear teacher-student model with a low-dimensional latent structure and analyze how the similarity of input features and readout patterns between tasks affect continual learning.\nWe show analytically that a combination of low feature similarity and high readout similarity is relatively benign for continual learning, as the retention performance remains high and the transfer performance remains non-negative.\nHowever, the opposite, a combination of high feature similarity and low readout similarity is harmful. In this regime, both transfer and retention performance become below the chance level even when the two subsequent tasks are positively correlated.\nFurthermore, transfer performance depends on the feature similarity non-monotonically, such that, beyond a critical point, the higher the feature similarity is, the lower the transfer performance becomes.\n\n\nWe further analyze how common algorithms for continual learning, activity and plasticity gating[14,40,44], activity sparsification[55], and weight regularization[29,35,65], interact with task similarity in our problem setting, deriving several non-trivial conclusions.\nActivity gating improves retention at the cost of transfer when the gating highly sparsifies the activity, but helps both transfer and retention on average if the activity is kept relatively dense.\nPlasticity gating and activity sparsification, by contrast, do not influence either transfer or retention performance at the over-parameterized limit.\nLastly, weight regularization in the Fisher information metric helps retention without affecting knowledge transfer and achieves perfect retention regardless of task similarity in the presence of low-dimensional latent.\nHowever, its diagonal approximation and the regularization in the Euclidean metric are much less robust against both task similarity and regularizer amplitude.\n\n\nFurthermore, we test our key predictions numerically in a permuted MNIST task with a latent structure.\nWhen the input pixels are permuted from one task to the next, the retention performance remains high.\nHowever, when the mapping from the latent to the target output is changed, both the retention and transfer performance go below the chance level, as predicted.\nRandom task-dependent gating of input and hidden layer activity improves retention at the cost of knowledge transfer, but adaptive gating mitigates this tradeoff.\nWe also show that in a fully-connected feedforward network, there exists an efficient layer-wise approximation of the weight regularization in the Fisher information metric, which outperforms its diagonal approximation and the regularization in the Euclidean metric.\nNevertheless, the performance of the diagonal approximation is much closer to the layer-wise approximation of the Fisher information metric than to the Euclidean weight regularization.\n\n\nOur theory thus reveals when continual learning is difficult, and how different algorithms mitigate these challenging situations, providing a basic framework for analyzing continual learning in artificial and biological neural networks.\n\n",
    "related_work": "Previous works on continual learning in linear teacher-student models found that forgetting is most prominent at the intermediate task similarity[11,19,12]as observed empirically[45].\nHowever, these works did not address the tradeoff between forgetting and knowledge transfer, and these simple settings did not disentangle the effect of the similarity in input feature and readout pattern.\nForward and backward transfer performance in continual learning were also analyzed in linear and deep-linear networks[32,8], yet its relationship with catastrophic forgetting has not been well characterized.Lee et al.[34]analyzed dynamics of both forgetting and forward transfer in one-hidden layer nonlinear network under a multi-head continual learning setting.\nHowever, their analysis of readout similarity and its comparison to feature similarity were conducted numerically and it did not address the effect of common heuristics, such as gating or weight regularization, either.\n\n\nBy comparison, we introduce a low-dimensional latent structure into a linear teacher-student model, which enables us to decouple the influence of feature and readout similarity on knowledge transfer and retention.\nMoreover, this low-dimensionality assumption on the latent enables us to evaluate the transfer and retention performance analytically even in the presence of gating or weight regularization in the Fisher information metric.\n\n\nContinual learning has been studied from many theoretical frameworks beyond teacher-student modeling, including neural tangent kernel[5,9,26], PAC learning[6,58], and computational complexity[30].\nLearning of low-dimensional latent representation has also been studied in the context of multi-task learning[41,60].\nIn addition, several works investigated the effect of weight regularization on continual learning in analytically tractable settings[29,11,23].\n\n",
    "origin_citations": [
      "How catastrophic can catastrophic forgetting be in linear regression?",
      "Analysis of catastrophic forgetting for random orthogonal transformation tasks in the overparameterized regime.",
      "The joint effect of task similarity and overparameterization on catastrophic forgetting–an analytical model.",
      "Anatomy of catastrophic forgetting: Hidden representations and task semantics.",
      "An analytic theory of generalization dynamics and transfer learning in deep linear networks.",
      "Phase transitions in transfer learning for high-dimensional perceptrons.",
      "Continual learning in the teacher-student setup: Impact of task similarity.",
      "Generalisation guarantees for continual learning with orthogonal gradient descent.",
      "A theoretical analysis of catastrophic forgetting through the ntk overlap matrix.",
      "Learning curves for continual learning in neural networks: Self-knowledge transfer and forgetting.",
      "Memory bounds for continual learning.",
      "Incorporating neuro-inspired adaptability for continual learning in artificial intelligence.",
      "Optimal continual learning has perfect memory and is np-hard.",
      "The benefit of multitask representation learning.",
      "On the statistical benefits of curriculum learning.",
      "Overcoming catastrophic forgetting in neural networks.",
      "Provable continual learning via sketched jacobian approximations."
    ],
    "citations": [
      148160,
      414008,
      125046,
      199896,
      390564,
      39679,
      371623,
      168763,
      356160,
      152846,
      369164,
      403878,
      350857,
      344420,
      275808,
      1562,
      151405
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 87,
      "Final_Score": 83.5,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 87\nFinal_Score: 83.5"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 90,
      "Final_Score": 88.4,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 92\nSynthesis_Analysis: 90\nFinal_Score: 89"
    },
    "overall_score": 86.44,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "bE7GWLQzkM",
        "forum": "bE7GWLQzkM",
        "number": 5670,
        "title": "Disentangling and mitigating the impact of task similarity for continual learning",
        "normalized_title": "disentangling and mitigating the impact of task similarity for continual learning",
        "abstract": "Continual learning of partially similar tasks poses a challenge for artificial neural networks, as task similarity presents both an opportunity for knowledge transfer and a risk of interference and catastrophic forgetting.\nHowever, it remains unclear how task similarity in input features and readout patterns influences knowledge transfer and forgetting, as well as how they interact with common algorithms for continual learning.\nHere, we develop a linear teacher-student model with latent structure and show analytically that high input feature similarity coupled with low readout similarity is catastrophic for both knowledge transfer and retention. \nConversely, the opposite scenario is relatively benign. \nOur analysis further reveals that task-dependent activity gating improves knowledge retention at the expense of transfer, while task-dependent plasticity gating does not affect either retention or transfer performance at the over-parameterized limit. \nIn contrast, weight regularization based on the Fisher information metric significantly improves retention, regardless of task similarity, without compromising transfer performance. Nevertheless, its diagonal approximation and regularization in the Euclidean space are much less robust against task similarity. \nWe demonstrate consistent results in a permuted MNIST task with latent variables. Overall, this work provides insights into when continual learning is difficult and how to mitigate it.",
        "authors": [
          "Naoki Hiratani"
        ],
        "decision": null,
        "pdate": 1727287791477,
        "cdate": 1715563529972
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "bE7GWLQzkM",
          "forum": "bE7GWLQzkM",
          "number": 5670,
          "title": "Disentangling and mitigating the impact of task similarity for continual learning",
          "normalized_title": "disentangling and mitigating the impact of task similarity for continual learning",
          "abstract": "Continual learning of partially similar tasks poses a challenge for artificial neural networks, as task similarity presents both an opportunity for knowledge transfer and a risk of interference and catastrophic forgetting.\nHowever, it remains unclear how task similarity in input features and readout patterns influences knowledge transfer and forgetting, as well as how they interact with common algorithms for continual learning.\nHere, we develop a linear teacher-student model with latent structure and show analytically that high input feature similarity coupled with low readout similarity is catastrophic for both knowledge transfer and retention. \nConversely, the opposite scenario is relatively benign. \nOur analysis further reveals that task-dependent activity gating improves knowledge retention at the expense of transfer, while task-dependent plasticity gating does not affect either retention or transfer performance at the over-parameterized limit. \nIn contrast, weight regularization based on the Fisher information metric significantly improves retention, regardless of task similarity, without compromising transfer performance. Nevertheless, its diagonal approximation and regularization in the Euclidean space are much less robust against task similarity. \nWe demonstrate consistent results in a permuted MNIST task with latent variables. Overall, this work provides insights into when continual learning is difficult and how to mitigate it.",
          "authors": [
            "Naoki Hiratani"
          ],
          "decision": null,
          "pdate": 1727287791477,
          "cdate": 1715563529972
        }
      ]
    }
  },
  {
    "paper_id": 49700,
    "title": "Data Unlearning in Diffusion Models",
    "abstract": "Recent work has shown that diffusion models memorize and reproduce training data examples. At the same time, large copyright lawsuits and legislation such as GDPR have highlighted the need for erasing datapoints from diffusion models. However, retraining from scratch is often too expensive. This motivates the setting of data unlearning, i.e., the study of efficient techniques for unlearning specific datapoints from the training set. Existing concept unlearning techniques require an anchor prompt/class/distribution to guide unlearning, which is not available in the data unlearning setting. General-purpose machine unlearning techniques were found to be either unstable or failed to unlearn data. We therefore propose a family of new loss functions called Subtracted Importance Sampled Scores (SISS) that utilize importance sampling and are the first method to unlearn data with theoretical guarantees. SISS is constructed as a weighted combination between simpler objectives that are responsible for preserving model quality and unlearning the targeted datapoints. When evaluated on CelebA-HQ and MNIST, SISS achieved Pareto optimality along the quality and unlearning strength dimensions. On Stable Diffusion, SISS successfully mitigated memorization on nearly90%percent9090\\%90 %of the prompts we tested. We release our code online.111https://github.com/claserken/SISS",
    "introduction": "The recent advent of diffusion models has revolutionized high-quality image generation, with large text-to-image models such as Stable DiffusionRombach et al. (2022)demonstrating impressive stylistic capabilities. However, these models have been shown to memorize and reproduce specific training images, raising significant concerns around data privacy, copyright legality, and the generation of inappropriate contentCarlini et al. (2023); Cilloni et al. (2023). Incidents such as the discovery of child sexual abuse material in LAION(Thiel,2023; Schuhmann et al.,2022)as well as the need to comply with regulations like the General Data Protection Regulation and California Consumer Privacy Act that establish a “right to be forgotten”Hong et al. (2024); Wu et al. (2024), underscore the urgency of developing effective methods to remove memorized data from diffusion models.\n\n\nRetraining on a new dataset is often prohibitively expensive, and the bulk of traditional machine unlearning techniques have been built for classical supervised machine learningCao & Yang (2015); Ginart et al. (2019); Izzo et al. (2021); Bourtoule et al. (2021). Recently, a new wave of research on unlearning in diffusion models has emerged, but it has focused almost exclusively onconcept unlearningin text-conditional modelsGandikota et al. (2023); Kumari et al. (2023); Zhang et al. (2023); Gandikota et al. (2024); Schramowski et al. (2023); Heng & Soh (2023); Fan et al. (2024). These approaches aim to remove higher-level concepts, e.g., the styles of painters or nudity, rather than specific datapoints from the training data needed to battle unwanted memorization. This paper focuses on the problem of efficient machine unlearning in diffusion models with the objective of removing specific datapoints, a problem that we refer to asdata unlearning.\n\n\nUnlike concept unlearning, the data unlearning setting has a concrete gold standard: retraining without the data to be unlearned. The goal of data unlearning is to achieve unlearning performance as close as possible to retraining while using less computational resources. To quantify unlearning performance, we focus on three separate areas: the degree of unlearning, the amount of quality degradation after unlearning, and the amount of compute needed. Examples of these areas are highlighted in Figures1and2.\n\n\nPretrainedSISS (λ=0.5𝜆0.5\\lambda=0.5italic_λ = 0.5)EraseDiffNegGradCelebA-HQMNIST T-ShirtStable Diffusion(LAION-5B)Figure 1:Examples of quality degradation across unlearning methods. On all3333datasets, we find that our SISS method is the only method capable of unlearning specific training datapoints while maintaining the original model quality. See Tables1,2and Figure6(a)for complete quantitative results on quality preservation.\n\n\nWhen applied to data unlearning, general-purpose machine unlearning techniques face certain limitations: naive deletion (fine-tuning on data to be kept) tends to be slow in unlearning, while NegGrad (gradient ascent on data to be unlearned)(Golatkar et al.,2020)forgets the data to be kept, leading to rapid quality degradation. State-of-the-art class and concept unlearning techniques do not apply to our setting because they require an anchor prompt/class/distribution to guide the unlearning towards which we do not assume access to. For example,Heng & Soh (2023)select a uniform distribution over pixel values to replace the00class on MNIST; however, the desired unlearning behavior would be to instead generate digits1−9191-91 - 9when conditioned on00. Even if one selects the target distribution to be a random example/label from all other classes asFan et al. (2024)do, it is not always clear what “all other classes” are in the text-conditional case. Furthermore, these prior experiments are targeting prompts/classes instead of datapoints and do not apply in the case of unconditional diffusion models. A notable exception is EraseDiff(Wu et al.,2024)which can unlearn data by fitting the predicted noise to random noise targets that are not associated with a prompt/class.\n\n\nIn this work, we derive an unlearning objective that combines the objectives of naive deletion and NegGrad. For further computational efficiency, we unify the objective through importance sampling, cutting the number of forward passes needed to compute it by half. We term our objective Subtracted Importance Sampled Scores (SISS). As seen in Figure1, SISS allows for the computationally efficient unlearning of training data subsets while preserving model quality. It does so because the naive deletion component preserves the data to be kept, while the NegGrad component targets the data to be unlearned. The addition of importance sampling balances between the two components through a parameterλ𝜆\\lambdaitalic_λ, whereλ=0𝜆0\\lambda=0italic_λ = 0andλ=1𝜆1\\lambda=1italic_λ = 1behave like naive deletion and NegGrad, respectively. We find thatλ=0.5𝜆0.5\\lambda=0.5italic_λ = 0.5behaves as a mixture of the two, giving the desirable combination of quality preservation and strong unlearning.\n\n\nWe demonstrate the effectiveness of SISS on CelebA-HQ(Karras et al.,2018), MNIST with T-Shirt, and Stable Diffusion. On all3333sets of experiments, SISS preserved the original model quality as shown in Figure1. On CelebA-HQ with the objective of unlearning a celebrity face, SISS was Pareto-optimal with respect to the FID and SSCD similarity metric in Figure2, cutting the latter by over half. The base model for MNIST with T-Shirt was trained on MNIST(Deng,2012)augmented with a specific T-shirt from Fashion-MNIST(Xiao et al.,2017). The objective was to unlearn the T-shirts, and SISS was again found to be Pareto-optimal with respect to the Inception Score and exact likelihood, increasing the latter by a factor of8888. Finally, on Stable Diffusion, we found SISS to successfully mitigate memorization on almost90%percent9090\\%90 %of the prompts we tested.\n\n",
    "related_work": "Machine Unlearning.Machine unlearning is the notoriously difficult problem of removing the influence of datapoints that models were previously trained onCao & Yang (2015); Bourtoule et al. (2021); Shaik et al. (2024). Over the past years, it has received increased attention and relevance due to privacy regulation such as the EU’s Right to be ForgottenGinart et al. (2019); Izzo et al. (2021); Golatkar et al. (2020); Tarun et al. (2023). The first wave of methods mostly approached classical machine learning methods like linear and logistic regressionIzzo et al. (2021),k𝑘kitalic_k-means clusteringGinart et al. (2019), statistical query learning methodsCao & Yang (2015), Bayesian methodsNguyen et al. (2020)or various types of supervised deep learning methodsGolatkar et al. (2020); Tarun et al. (2023). Some of these methods require modifications to the training procedure, e.g., training multiple models on distinct dataset shardsBourtoule et al. (2021); Golatkar et al. (2024), whereas others can be applied purely in post-training such as NegGradGolatkar et al. (2020)and BlindSpotTarun et al. (2023). Recently, generative models have become a popular paradigm. While unlearning in this paradigm is less explored, there are early approaches looking at Generative Adversarial Networks (GANs)Kong & Alfeld (2023), language modelsLiu et al. (2024); Yao et al. (2024)and on diffusion models via shardingGolatkar et al. (2024).\n\n\nMemorization in Diffusion Models.Large-scale diffusion models trained on image generation have recently attracted the attention of copyright lawsuits since they are prone to memorizing training examplesSomepalli et al. (2023a); Carlini et al. (2023); Somepalli et al. (2023b); Webster (2023).Somepalli et al. (2023a)showed that Stable DiffusionRombach et al. (2022)exhibits verbatim memorization for heavily duplicated training data examples.Webster (2023)classified different types of memorization, introducing types of partial memorization.Carlini et al. (2023)discusses various black-box extraction and membership inference attacks and demonstrates them successfully on Stable Diffusion. Most recently, mitigation strategies have been introduced, e.g., by manually modifying the text promptsSomepalli et al. (2023b)or taking gradients steps in prompt space to minimize the magnitude of text-conditional noise predictionsWen et al. (2024).\n\n\nConcept Unlearning in Diffusion Models.Recently, the problem of unlearning has become popular in the context of diffusion models, though almost exclusively in the form of concept unlearning: while the classical setting of machine unlearning deals with forgetting specific datapoints from the training set – which we calldata unlearningfor clarity – the setting ofconcept unlearningdeals with forgetting higher level concepts in text-conditional models, e.g., nudity or painting stylesShaik et al. (2024); Gandikota et al. (2023); Kong & Chaudhuri (2024); Kumari et al. (2023); Zhang et al. (2024); Heng & Soh (2023).Zhang et al. (2023)introduces Forget-Me-Not, a concept unlearning technique that minimizes the cross-attention map for an undesired prompt and also introduces ConceptBench as a benchmark.Gandikota et al. (2023)find an alternate approach to concept unlearning fine-tuning by fitting to noise targets that are biased away from the predicted noise with respect to an undesirable prompt. Similarly,Schramowski et al. (2023)also bias the noise away from an undesired prompt but do so only at inference time. UnlearnCanvas is a benchmark introduced byZhang et al. (2024)to measure concept unlearning for artistic styles and objects. EraseDiffWu et al. (2024)discusses the data unlearning setting but only studies the settings of unlearning classes or concepts. Lastly,Li et al. (2024)indeed studies data unlearning but solely in the case of image-to-image models. To the authors’ knowledge, the data unlearning setting remains a gap in the diffusion model literature.\n\n",
    "origin_citations": [
      "Towards making systems forget with machine unlearning.",
      "Machine unlearning.",
      "Exploring the landscape of machine unlearning: A comprehensive survey and taxonomy, 2024.",
      "Making AI forget you: Data deletion in machine learning.",
      "Approximate data deletion from machine learning models.",
      "Eternal sunshine of the spotless net: Selective forgetting in deep networks.",
      "Deep regression unlearning.",
      "Variational bayesian unlearning.",
      "Training data protection with compositional diffusion models, 2024.",
      "Approximate data deletion in generative models.",
      "Rethinking machine unlearning for large language models, 2024.",
      "Large language model unlearning, 2024.",
      "Diffusion art or digital forgery? investigating data replication in diffusion models.",
      "Extracting training data from diffusion models.",
      "Understanding and mitigating copying in diffusion models.",
      "A reproducible extraction of training images from diffusion models, 2023.",
      "High-resolution image synthesis with latent diffusion models.",
      "Detecting, explaining, and mitigating memorization in diffusion models.",
      "Erasing concepts from diffusion models.",
      "Data redaction from conditional generative models.",
      "Ablating concepts in text-to-image diffusion models.",
      "Unlearncanvas: Stylized image dataset for enhanced machine unlearning evaluation in diffusion models, 2024.",
      "Selective amnesia: A continual learning approach to forgetting in deep generative models.",
      "Forget-me-not: Learning to forget in text-to-image diffusion models, 2023.",
      "Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models.",
      "Erasediff: Erasing data influence in diffusion models, 2024.",
      "Machine unlearning for image-to-image generative models."
    ],
    "citations": [
      203788,
      115214,
      158244,
      165544,
      450444,
      872738,
      435859,
      182186,
      79530,
      97529,
      319902,
      222386,
      3026,
      548398,
      1075960,
      2240,
      550167,
      75921,
      261105,
      494095,
      769082,
      261749,
      492172,
      33136,
      483864,
      101949
    ],
    "extract_rate": 0.9629629629629629,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 84,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 86,
      "Final_Score": 87.8,
      "raw_text": "Content_Coherence: 84\nCitation_Quality: 92\nSynthesis_Analysis: 86\nFinal_Score: 88"
    },
    "overall_score": 86.32,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "SuHScQv5gP",
        "forum": "SuHScQv5gP",
        "number": 9831,
        "title": "Data Unlearning in Diffusion Models",
        "normalized_title": "data unlearning in diffusion models",
        "abstract": "Recent work has shown that diffusion models memorize and reproduce training data examples. At the same time, large copyright lawsuits and legislation such as GDPR have highlighted the need for erasing datapoints from diffusion models. However, retraining from scratch is often too expensive. This motivates the setting of data unlearning, i.e., the study of efficient techniques for unlearning specific datapoints from the training set. Existing concept unlearning techniques require an anchor prompt/class/distribution to guide unlearning, which is not available in the data unlearning setting. General-purpose machine unlearning techniques were found to be either unstable or failed to unlearn data. We therefore propose a family of new loss functions called Subtracted Importance Sampled Scores (SISS) that utilize importance sampling and are the first method to unlearn data with theoretical guarantees. SISS is constructed as a weighted combination between simpler objectives that are responsible for preserving model quality and unlearning the targeted datapoints. When evaluated on CelebA-HQ and MNIST, SISS achieved Pareto optimality along the quality and unlearning strength dimensions. On Stable Diffusion, SISS successfully mitigated memorization on nearly 90% of the prompts we tested. We release our code online.",
        "authors": [
          "Silas Alberti",
          "Kenan Hasanaliyev",
          "Manav Shah",
          "Stefano Ermon"
        ],
        "decision": null,
        "pdate": 1737562681506,
        "cdate": 1727431135015
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "SuHScQv5gP",
          "forum": "SuHScQv5gP",
          "number": 9831,
          "title": "Data Unlearning in Diffusion Models",
          "normalized_title": "data unlearning in diffusion models",
          "abstract": "Recent work has shown that diffusion models memorize and reproduce training data examples. At the same time, large copyright lawsuits and legislation such as GDPR have highlighted the need for erasing datapoints from diffusion models. However, retraining from scratch is often too expensive. This motivates the setting of data unlearning, i.e., the study of efficient techniques for unlearning specific datapoints from the training set. Existing concept unlearning techniques require an anchor prompt/class/distribution to guide unlearning, which is not available in the data unlearning setting. General-purpose machine unlearning techniques were found to be either unstable or failed to unlearn data. We therefore propose a family of new loss functions called Subtracted Importance Sampled Scores (SISS) that utilize importance sampling and are the first method to unlearn data with theoretical guarantees. SISS is constructed as a weighted combination between simpler objectives that are responsible for preserving model quality and unlearning the targeted datapoints. When evaluated on CelebA-HQ and MNIST, SISS achieved Pareto optimality along the quality and unlearning strength dimensions. On Stable Diffusion, SISS successfully mitigated memorization on nearly 90% of the prompts we tested. We release our code online.",
          "authors": [
            "Silas Alberti",
            "Kenan Hasanaliyev",
            "Manav Shah",
            "Stefano Ermon"
          ],
          "decision": null,
          "pdate": 1737562681506,
          "cdate": 1727431135015
        }
      ]
    }
  },
  {
    "paper_id": 40755,
    "title": "DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation",
    "abstract": "Recent advancements in 2D/3D generative techniques have facilitated the generation of dynamic 3D objects from monocular videos. Previous methods mainly rely on the implicit neural radiance fields (NeRF) or explicit Gaussian Splatting as the underlying representation, and struggle to achieve satisfactory spatial-temporal consistency and surface appearance. Drawing inspiration from modern 3D animation pipelines, we introduce DreamMesh4D, a novel framework combining mesh representation with geometric skinning technique to generate high-quality 4D object from a monocular video. Instead of utilizing classical texture map for appearance, we bind Gaussian splats to triangle face of mesh for differentiable optimization of both the texture and mesh vertices. In particular, DreamMesh4D begins with a coarse mesh obtained through an image-to-3D generation procedure. Sparse points are then uniformly sampled across the mesh surface, and are used to build a deformation graph to drive the motion of the 3D object for the sake of computational efficiency and providing additional constraint. For each step, transformations of sparse control points are predicted using a deformation network, and the mesh vertices as well as the surface Gaussians are deformed via a novel geometric skinning algorithm. The skinning algorithm is a hybrid approach combining LBS (linear blending skinning) and DQS (dual-quaternion skinning), mitigating drawbacks associated with both approaches. The static surface Gaussians and mesh vertices as well as the dynamic deformation network are learned via reference view photometric loss, score distillation loss as well as other regularization losses in a two-stage manner. Extensive experiments demonstrate superior performance of our method in terms of both rendering quality and spatial-temporal consistency. Furthermore, our method is compatible with modern graphic pipelines, showcasing its potential in the 3D gaming and film industry.\nThe source code is available at our website:https://lizhiqi49.github.io/DreamMesh4D.",
    "introduction": "The emergence and development of Generative Artificial Intelligence (GenAI) have significantly revolutionized 3D generation techniques in recent years[20]. The technology has effectively allowed the creation of static objects, including their shape, texture, and even an entire scene from a simple text prompt or a single image. Recently, the wave of advancement has been propelled to the filed of dynamic (4D) content generation[45], which offers immense potential in fields including, but not limited to, AR/VR, filming, gaming and animation. However, it’s still quite challenging to efficiently generate high-quality 4D content due to its increased spatial-temporal complexity and higher demand on algorithm design.\n\n\nThe promising strides in 3D generation are largely attributed to the pre-trained large 2D diffusion models[39,41,31]. In particular, score distillation sampling (SDS)[33]enables the 3D generation[30,55,18,42,9]from scratch by distilling 3D knowledge from a pre-trained 2D diffusion model[39]. Following works on text-to-3D[24,57,44,15]and image-to-3D[26,35,50,48]have further improved the performance of 3D generation tasks both in quality and stability.\nInspired by the successes of SDS-based 3D generation, recent works[69,1,38,12,65,66]explore generating 4D assets by distilling prior knowledge from pre-trained video diffusion models[54,3]or novel-view synthesis models[26,43].\nBoth text-to-4D[45,1,25]and image-to-4D methods[69]mainly rely on pre-trained video diffusion models, which are not yet capable of generating high-quality video, thus usually struggle to generate high-quality 4D content. On the contrary, video-to-4D methods[38,12,66,65]directly generate 4D assets from off-the-shelf monocular videos, making the results more appealing and with better spatial-temporal consistency.\nExisting video-to-4D methods either rely on the implicit dynamic NeRFs[12]or explicit dynamic Gaussian splatting[38,65,66]as the underlying representation. Nevertheless, both of them do not have tight constrains for surface, leading to redundant optimization space and impeding the learning of deformation.\n\n\nInspired by modern graphic pipelines for 3D animation, we proposeDreamMesh4D, which exploits 3D triangular mesh representation and sparse-controlled geometric skinning methods[47,16]for video-to-4D generation. To better supervise the generation with 2D signals, instead of using classic mesh with UV texture maps, we choose a hybrid representation, SuGaR[9], which marries 3D Gaussians to mesh surface for more elaborate appearance modeling. Flat Gaussians are bound to mesh faces based on barycentric coordinates hence the rendering process of 2D images is differentiable with respect to both the position of mesh vertices and the attributes of Gaussians. For high-quality object modeling and efficient motion driving of the object, our method is designed in a static-to-dynamic optimization manner. In particular, during the static stage, an initial coarse mesh is generated utilizing existing image-to-3D generation methods[26,68,61]. Then we refine its both geometry and texture by jointly optimizing the mesh vertices as well as the attributes of bound surface Gaussians under the hybrid representation via both the reference image photometric loss and the SDS loss. For dynamic learning, we uniformly sample sparse control nodes from the surface of the refined mesh, to build a deformation graph. Then at each timestamp, transformation associated with each control node is predicted by a deformation network. The deformation of all mesh vertices and surface Gaussians are obtained from those predicted transformations via a novel geometric skinning algorithm, which benefits from both the LBS (linear blending skinning) and DQS (dual-quaternion skinning) methods. The deformation network is optimized under the supervision of photometric loss from reference video frames, novel-view SDS loss and geometric regularization terms.\n\n\nExtensive experiments are conducted and demonstrate that our method can generate high-fidelity dynamic textured mesh from monocular video, and significantly outperforms previous works both quantitatively and qualitatively, establishing new benchmark in the field of video-to-4D generation. As shown in Fig.1, our generated assets can be directly simulated in modern 3D engines, showcasing its potential in the 3D gaming and film industries.\n\n",
    "related_work": "3D GenerationSince the introduction of score distillation sampling (SDS) by DreamFusion[33], subsequent works[24,5,57,35,48,49,23]have significantly improved the performance of optimization-based 3D generation algorithms.\nMany works adopt a multi-stage optimization strategy[24,5,57,48]to enhance generated appearance. Another line of research[26,44,27,19]focuses on training multi-view diffusion models to inject multi-view supervision into SDS loss for addressing the Janus problem.\nDreamGaussian[49]and GaussianDreamer[64]pioneer the usage of 3D Gaussian[18]as the underlying representation and achieve 3D content generation in minutes.\nAlthough these methods demonstrate the potential of 3D Gaussian in 3D content generation, obtaining an object with high-quality geometry is still quite challenging.\nIn this work, we explore to employ a Gaussian-mesh hybrid representation[9]in our 4D generation tasks for better modeling of both object surface geometry and texture.\n\n\n4D RepresentationsDynamic 3D representations form the foundation of 4D reconstruction and generation tasks. Most current methods extend static NeRF[30]to dynamic scenarios. These approaches, such as deformable[32,34,51,59]and time-varying[4,7,8,6]NeRF-based methods, are prevalent.\nThere are also some works trying to model dynamic with 4D neural implicit surface[56,14].\nHowever, these implicit representations suffer from long optimization time and low reconstruction quality due to its computationally expensive volume rendering and implicit representation.\nRecent interest in 4D tasks has focused on 3D Gaussian representations due to their fast rendering speed and explicit nature. Some works[63,58,22]train networks to predict Gaussian kernel deformations, while others model kernel motion via polynomial representation or per-frame optimization[28,21].\nBesides, both SC-GS[11]and HiFi4G[13]employ sparse control points for Gaussian kernel deformation, with SC-GS using LBS and HiFi4G using DQS to drive Gaussians motion, thus ensuring spatial-temporal consistency.\nIn this work, we propose to deform the object through a novel geometric skinning method, handling the artifacts associated with both LBS and DQS.\n\n\n4D GenerationAlthough great progress has been achieved in 3D generation tasks, 4D generation is still challenging due to its requirement on additional temporal supervision. Since current pre-trained video diffusion models[3,54]still struggle to generate high-quality video contents, it is challenging to distill useful motion knowledge via SDS optimization. Hence, the performance of existing text-to-4D[45,1,25]and image-to-4D methods[69]usually struggle with low appearance quality.\nAnother line of works focus on video-to-4D generation[12,38,65]. These methods leverage current multiview diffusion models[26,43,27]to calculate the SDS loss[38,12,60]or generate per-frame multi-view images[66,62]as supervision signal. Among them, a concurrent work of our method, SC4D[60], optimizes a set of sparse-controlled dynamic 3D Gaussians by per-frame SDS loss from Zero123[26]with a coarse-to-fine strategy. However, the issue of unsatisfying surface quality caused by 3D Gaussian-based representation still exists. In contrast, our method is grounding on a Gaussian-mesh hybrid representation[9], enhancing the reconstruction quality both in texture and geometry.\n\n",
    "origin_citations": [
      "Dreamfusion: Text-to-3d using 2d diffusion.",
      "Magic3d: High-resolution text-to-3d content creation.",
      "Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation.",
      "Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation.",
      "Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors.",
      "Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior.",
      "Dreamgaussian: Generative gaussian splatting for efficient 3d content creation.",
      "Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Mvdream: Multi-view diffusion for 3d generation.",
      "Syncdreamer: Generating multiview-consistent images from a single-view image.",
      "Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d.",
      "Gaussiandreamer: Fast generation from text to 3d gaussian splatting with point cloud priors.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering.",
      "Nerf: Representing scenes as neural radiance fields for view synthesis.",
      "Nerfies: Deformable neural radiance fields.",
      "D-nerf: Neural radiance fields for dynamic scenes.",
      "Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video.",
      "D2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video.",
      "Hexplane: A fast representation for dynamic scenes.",
      "K-planes: Explicit radiance fields in space, time, and appearance.",
      "Dynamic view synthesis from dynamic monocular video.",
      "Fast dynamic radiance fields with time-aware neural voxels.",
      "Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction.",
      "Unbiased 4d: Monocular 4d reconstruction with a neural deformation model.",
      "Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction.",
      "4d gaussian splatting for real-time dynamic scene rendering.",
      "Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis.",
      "Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis.",
      "Spacetime gaussian feature splatting for real-time dynamic view synthesis.",
      "Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes.",
      "Hifi4g: High-fidelity human performance rendering via compact gaussian splatting.",
      "Stable video diffusion: Scaling latent video diffusion models to large datasets.",
      "Modelscope text-to-video technical report.",
      "Text-to-4d dynamic scene generation.",
      "4d-fy: Text-to-4d generation using hybrid score distillation sampling.",
      "Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models.",
      "Animate124: Animating one image to 4d dynamic scene.",
      "Consistent4d: Consistent 360 {{\\{{\\\\\\backslash\\deg}}\\}} dynamic object generation from monocular video.",
      "Dreamgaussian4d: Generative 4d gaussian splatting.",
      "4dgen: Grounded 4d content generation with spatial-temporal consistency.",
      "Zero123++: a single image to consistent multi-view diffusion base model.",
      "Sc4d: Sparse-controlled video-to-4d generation and motion transfer.",
      "Stag4d: Spatial-temporal anchored generative 4d gaussians.",
      "Diffusion2: Dynamic 3d content generation via score composition of orthogonal diffusion models."
    ],
    "citations": [
      176110,
      245032,
      493936,
      258271,
      1009434,
      312653,
      477674,
      25226,
      495077,
      1087300,
      129306,
      323293,
      180969,
      24495,
      253579,
      431438,
      431037,
      105722,
      996331,
      213655,
      213434,
      249143,
      145318,
      221790,
      82218,
      479493,
      320337,
      348887,
      406453,
      346462,
      273928,
      273166,
      23455,
      179813,
      212984,
      22093,
      347883,
      23669,
      346294,
      346260,
      313538,
      448503,
      425405,
      449280
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 85,
      "Final_Score": 85.7,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 92\nSynthesis_Analysis: 85\nFinal_Score: 86"
    },
    "overall_score": 86.02000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "6ZwJSk2kvU",
        "forum": "6ZwJSk2kvU",
        "number": 6283,
        "title": "DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation",
        "normalized_title": "dreammesh4d video to 4d generation with sparse controlled gaussian mesh hybrid representation",
        "abstract": "Recent advancements in 2D/3D generative techniques have facilitated the generation of dynamic 3D objects from monocular videos. Previous methods mainly rely on the implicit neural radiance fields (NeRF) or explicit Gaussian Splatting as the underlying representation, and struggle to achieve satisfactory spatial-temporal consistency and surface appearance. Drawing inspiration from modern 3D animation pipelines, we introduce DreamMesh4D, a novel framework combining mesh representation with geometric skinning technique to generate high-quality 4D object from a monocular video. Instead of utilizing classical texture map for appearance, we bind Gaussian splats to triangle face of mesh for differentiable optimization of both the texture and mesh vertices. In particular, DreamMesh4D begins with a coarse mesh obtained through an image-to-3D generation procedure. Sparse points are then uniformly sampled across the mesh surface, and are used to build a deformation graph to drive the motion of the 3D object for the sake of computational efficiency and providing additional constraint. For each step, transformations of sparse control points are predicted using a deformation network, and the mesh vertices as well as the surface Gaussians are deformed via a novel geometric skinning algorithm. The skinning algorithm is a hybrid approach combining LBS (linear blending skinning) and DQS (dual-quaternion skinning), mitigating drawbacks associated with both approaches. The static surface Gaussians and mesh vertices as well as the dynamic deformation network are learned via reference view photometric loss, score distillation loss as well as other regularization losses in a two-stage manner. Extensive experiments demonstrate superior performance of our method in terms of both rendering quality and spatial-temporal consistency. Furthermore, our method is compatible with modern graphic pipelines, showcasing its potential in the 3D gaming and film industry.",
        "authors": [
          "Zhiqi Li",
          "Yiming Chen",
          "Peidong Liu"
        ],
        "decision": null,
        "pdate": 1727287809890,
        "cdate": 1715589046048
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "6ZwJSk2kvU",
          "forum": "6ZwJSk2kvU",
          "number": 6283,
          "title": "DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation",
          "normalized_title": "dreammesh4d video to 4d generation with sparse controlled gaussian mesh hybrid representation",
          "abstract": "Recent advancements in 2D/3D generative techniques have facilitated the generation of dynamic 3D objects from monocular videos. Previous methods mainly rely on the implicit neural radiance fields (NeRF) or explicit Gaussian Splatting as the underlying representation, and struggle to achieve satisfactory spatial-temporal consistency and surface appearance. Drawing inspiration from modern 3D animation pipelines, we introduce DreamMesh4D, a novel framework combining mesh representation with geometric skinning technique to generate high-quality 4D object from a monocular video. Instead of utilizing classical texture map for appearance, we bind Gaussian splats to triangle face of mesh for differentiable optimization of both the texture and mesh vertices. In particular, DreamMesh4D begins with a coarse mesh obtained through an image-to-3D generation procedure. Sparse points are then uniformly sampled across the mesh surface, and are used to build a deformation graph to drive the motion of the 3D object for the sake of computational efficiency and providing additional constraint. For each step, transformations of sparse control points are predicted using a deformation network, and the mesh vertices as well as the surface Gaussians are deformed via a novel geometric skinning algorithm. The skinning algorithm is a hybrid approach combining LBS (linear blending skinning) and DQS (dual-quaternion skinning), mitigating drawbacks associated with both approaches. The static surface Gaussians and mesh vertices as well as the dynamic deformation network are learned via reference view photometric loss, score distillation loss as well as other regularization losses in a two-stage manner. Extensive experiments demonstrate superior performance of our method in terms of both rendering quality and spatial-temporal consistency. Furthermore, our method is compatible with modern graphic pipelines, showcasing its potential in the 3D gaming and film industry.",
          "authors": [
            "Zhiqi Li",
            "Yiming Chen",
            "Peidong Liu"
          ],
          "decision": null,
          "pdate": 1727287809890,
          "cdate": 1715589046048
        }
      ]
    }
  },
  {
    "paper_id": 52260,
    "title": "RaSA: Rank-Sharing Low-Rank Adaptation",
    "abstract": "Low-rank adaptation (LoRA) has been prominently employed for parameter-efficient fine-tuning of large language models (LLMs). However, the limited expressive capacity of LoRA, stemming from the low-rank constraint, has been recognized as a bottleneck, particularly in rigorous tasks like code generation and mathematical reasoning. To address this limitation, we introduce Rank-Sharing Low-Rank Adaptation (RaSA), an innovative extension that enhances the expressive capacity of LoRA by leveraging partial rank sharing across layers. By forming a shared rank pool and applying layer-specific weighting, RaSA effectively increases the number of ranks without augmenting parameter overhead. Our theoretically grounded and empirically validated approach demonstrates that RaSA not only maintains the core advantages of LoRA but also significantly boosts performance in challenging code and math tasks. Code, data and scripts are available at:https://github.com/zwhe99/RaSA.",
    "introduction": "Low-rank adaptation (LoRA,Hu et al. (2022)) has become a de facto parameter-efficient fine-tuning (PEFT) method for adapting large language models (LLMs) to specific downstream tasks.\nIts core idea is to constrain the parameter updates to be low-rank, which significantly reduces the number of trainable parameters and allows them to be merged back into the original model, thereby avoiding additional inference latency.\nDespite its advantages, recent studies have shown that LoRA still lags behind full fine-tuning (FFT), particularly in scenarios involving large training datasets and complex tasks such as mathematical reasoning and code generation(Jiang et al.,2024; Biderman et al.,2024).\nA plausible explanation for this performance gap is that the low-rank constraint limits the expressive capacity of LoRA.\nFor instance,Biderman et al. (2024)empirically found that the effective rank required for FFT is 10-100× higher than typical LoRA configuration, andZeng & Lee (2024)theoretically demonstrated that a Transformer network(Vaswani et al.,2017)requires a rank at least half the size of the model dimension to approximate another model of similar size.\n\n\nAlthough the limited number of trainable parameters results in limited expressive capacity, recent studies still indicate redundancy in LoRA’s parameters.\nFor example,Kopiczko et al. (2024); Song et al. (2024); Renduchintala et al. (2024); Li et al. (2024)further reduced the number of LoRA’s parameters by sharing them across layers and modules with only slight performance loss.Brüel-Gabrielsson et al. (2024)compressed 1,000 LoRAs trained from different tasks by sharing their parameter spaces.\nThis contradiction suggests that LoRA’s parameters are still not being fully utilized.\n\n\nCombining the above two observations, we proposeRank-Sharing Low-RankAdaptation (RaSA), an approach that boosts the expressive capacity of LoRA by enabling partial rank sharing across layers.\nSpecifically, given an LLM withL𝐿Litalic_Llayers, RaSA extractsk𝑘kitalic_kranks from each layer’s LoRA update to form a rank pool ofL×k𝐿𝑘L\\times kitalic_L × italic_kranks, which is shared across all layers with layer-specific weighting.\nRaSA retains the core advantages of LoRA – keeping the same parameter overhead and allowing for easy merging back into the model.\nMoreover, since modern LLMs typically have deep architectures (i.e., largeL𝐿Litalic_L), RaSA greatly increase the effective rank of the parameter update by(L−1)×k𝐿1𝑘(L-1)\\times k( italic_L - 1 ) × italic_k.\n\n\nHowever, a higher rank does not necessarily lead to better expressive capacity.\nTo rigorously assess the benefits of RaSA, we analyze its capacity to reconstruct high-rank matrices compared to LoRA.\nTheoretically, we prove that RaSA’s minimum reconstruction error is bounded by that of LoRA.\nEmpirically, we show that whenk𝑘kitalic_kis relatively small, RaSA can be easily optimized to achieve a significantly lower reconstruction error than LoRA.\nFinally, we conducted experiments on mathematical reasoning and code generation, demonstrating that the lower reconstruction error translates to improved downstream task performance.\n\n\nOur contributions are summarized as follows:•We propose RaSA, a novel extension of LoRA by by allowing partial rank sharing across layers, which significantly improves the method’s efficiency and expressiveness (§2).•We provide a comprehensive analysis – both theoretical and empirical – showcasing RaSA’s superior capacity for matrix reconstruction (§3) and its resultant improved performance on rigorous downstream tasks (e.g. code and math) (§4).\n\n",
    "related_work": "Parameter-Efficient Fine-Tuning (PEFT)PEFT methods aim to minimize the number of trainable parameters needed for fine-tuning large models, thus reducing memory and computational requirements.\nPioneering methods include adapter-based(Houlsby et al.,2019)and prompt-based(Lester et al.,2021; Li & Liang,2021)that introduce additional tunable adapter or prefix tokens to enable efficient fine-tuning while keeping the original model parameters fixed.\nHowever, these approaches can slow down inference speed due to the extra components introduced.\nLoRA overcomes this drawback by introducing low-rank matrices directly into the weight update process during fine-tuning, effectively reducing trainable parameters without increasing inference latency.\nDue to its robust performance, LoRA and its variants have been widely used to adapt LLMs for specific tasks(Yu et al.,2024; Xu et al.,2023; Biderman et al.,2024; Chen et al.,2024; Meng et al.,2024; yang Liu et al.,2024).Benedek & Wolf (2024)andZhang et al. (2023)show that the number of ranks required for each parameter matrix across the model’s layers is not uniform.\nTherefore, they propose dynamically assigning ranks based on the importance of parameters during training.\nThese rank-allocating approaches typically involve real-time estimation of parameter importance and pruning during the training process.\nIn contrast, RaSA uses a shared rank pool combined with layer-specific weighting, eliminating the need for complex importance estimation or pruning.Biderman et al. (2024)conduct a comprehensive empirical study on LoRA, and reveal that while LoRA still lags behind FFT, it exhibits less catastrophic forgetting. We show that our proposed RaSA forgets even less than LoRA, and learns more and faster.\n\n\nParameter Redundancy of LoRAAlthough LoRA has significantly reduced the number of trainable parameters, recent research suggest that it is possible to further minimize these parameters without compromising performance.Kopiczko et al. (2024)achieve a 99% reduction in LoRA parameters by fully sharing a pair of low-rank, frozen random matrices across all layers, adjusted with learnable scaling vectors.Koohpayegani et al. (2024)propose learning linear combinations of a set of random matrix bases, whileLi et al. (2024)push this further by replacing the matrix bases with a vector bank.Song et al. (2024)andRenduchintala et al. (2024)explore the effects of different sharing and selective fine-tuning strategies.\nBy sharing parameter spaces,Brüel-Gabrielsson et al. (2024)compress 1,000 LoRAs trained from different task, enabling more efficient serving.\nThese findings collectively suggest that LoRA’s parameter has not been fully utilized and that different LoRAs exhibit similarities across layers, modules, and even different tasks.\nRather than focusing on extreme parameter reduction, this work aims to maintain the same parameter count while exploring how inter-layer sharing can enhance parameter utilization.\nWe theoretically and empirically demonstrate that sharing ranks across layers leads to lower reconstruction error and thus better expressive capacity.\n\n",
    "origin_citations": [
      "Parameter-efficient transfer learning for NLP.",
      "The power of scale for parameter-efficient prompt tuning.",
      "Prefix-tuning: Optimizing continuous prompts for generation.",
      "Metamath: Bootstrap your own mathematical questions for large language models.",
      "A paradigm shift in machine translation: Boosting translation performance of large language models, 2023.",
      "Lora learns less and forgets less.",
      "LongloRA: Efficient fine-tuning of long-context large language models.",
      "Pissa: Principal singular values and singular vectors adaptation of large language models.",
      "DoRA: Weight-decomposed low-rank adaptation.",
      "PRILoRA: Pruned and rank-increasing low-rank adaptation.",
      "Adaptive budget allocation for parameter-efficient fine-tuning.",
      "VeRA: Vector-based random matrix adaptation.",
      "NOLA: Compressing lora using linear combination of random basis.",
      "Vb-lora: Extreme parameter efficient fine-tuning with vector banks.",
      "Sharelora: Parameter efficient and robust large language model fine-tuning via shared low-rank adaptation.",
      "Tied-LoRA: Enhancing parameter efficiency of LoRA with weight tying.",
      "Compress then serve: Serving thousands of lora adapters with little overhead, 2024."
    ],
    "citations": [
      7567,
      8536,
      1591,
      3513,
      480122,
      890118,
      479786,
      448949,
      97191,
      125693,
      1090069,
      315382,
      889685,
      789796,
      808467,
      26089,
      967255
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 75,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 85,
      "Final_Score": 86.0,
      "raw_text": "Content_Coherence: 75\nCitation_Quality: 95\nSynthesis_Analysis: 85\nFinal_Score: 86"
    },
    "overall_score": 85.96000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "GdXI5zCoAt",
        "forum": "GdXI5zCoAt",
        "number": 14067,
        "title": "RaSA: Rank-Sharing Low-Rank Adaptation",
        "normalized_title": "rasa rank sharing low rank adaptation",
        "abstract": "Low-rank adaptation (LoRA) has been prominently employed for parameter-efficient fine-tuning of large language models (LLMs). However, the limited expressive capacity of LoRA, stemming from the low-rank constraint, has been recognized as a bottleneck, particularly in rigorous tasks like code generation and mathematical reasoning. To address this limitation, we introduce Rank-Sharing Low-Rank Adaptation (RaSA), an innovative extension that enhances the expressive capacity of LoRA by leveraging partial rank sharing across layers. By forming a shared rank pool and applying layer-specific weighting, RaSA effectively increases the number of ranks without augmenting parameter overhead. Our theoretically grounded and empirically validated approach demonstrates that RaSA not only maintains the core advantages of LoRA but also significantly boosts performance in challenging code and math tasks. Code, data and scripts are available at: https://github.com/zwhe99/RaSA.",
        "authors": [
          "Zhiwei He",
          "Zhaopeng Tu",
          "Xing Wang",
          "Xingyu Chen",
          "Zhijie Wang",
          "Jiahao Xu",
          "Tian Liang",
          "Wenxiang Jiao",
          "Zhuosheng Zhang",
          "Rui Wang"
        ],
        "decision": null,
        "pdate": 1737563049448,
        "cdate": 1727518907525
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "GdXI5zCoAt",
          "forum": "GdXI5zCoAt",
          "number": 14067,
          "title": "RaSA: Rank-Sharing Low-Rank Adaptation",
          "normalized_title": "rasa rank sharing low rank adaptation",
          "abstract": "Low-rank adaptation (LoRA) has been prominently employed for parameter-efficient fine-tuning of large language models (LLMs). However, the limited expressive capacity of LoRA, stemming from the low-rank constraint, has been recognized as a bottleneck, particularly in rigorous tasks like code generation and mathematical reasoning. To address this limitation, we introduce Rank-Sharing Low-Rank Adaptation (RaSA), an innovative extension that enhances the expressive capacity of LoRA by leveraging partial rank sharing across layers. By forming a shared rank pool and applying layer-specific weighting, RaSA effectively increases the number of ranks without augmenting parameter overhead. Our theoretically grounded and empirically validated approach demonstrates that RaSA not only maintains the core advantages of LoRA but also significantly boosts performance in challenging code and math tasks. Code, data and scripts are available at: https://github.com/zwhe99/RaSA.",
          "authors": [
            "Zhiwei He",
            "Zhaopeng Tu",
            "Xing Wang",
            "Xingyu Chen",
            "Zhijie Wang",
            "Jiahao Xu",
            "Tian Liang",
            "Wenxiang Jiao",
            "Zhuosheng Zhang",
            "Rui Wang"
          ],
          "decision": null,
          "pdate": 1737563049448,
          "cdate": 1727518907525
        }
      ]
    }
  },
  {
    "paper_id": 49519,
    "title": "SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models",
    "abstract": "Traditional autonomous driving systems often struggle to integrate high-level reasoning with low-level control, resulting in suboptimal and sometimes unsafe driving behaviors. The emergence of multimodal large language models (MLLMs), which can process both visual and textual data, presents an opportunity to unify perception and reasoning tasks within a single framework. However, effectively embedding precise safety knowledge into MLLMs for autonomous driving remains a significant challenge.\nTo address this, we proposeSafeAuto, a novel framework that enhances MLLM-based autonomous driving systems by incorporating both unstructured and structured knowledge. Specifically, we first introduce the Position-Dependent Cross-Entropy (PDCE) loss function, designed to improve the accuracy of low-level control signal predictions when numerical values are represented as text. Second, to ensure safe autonomous driving by explicitly integrating precise safety knowledge into the MLLM, we develop a reasoning component forSafeAuto. This component translates driving safety regulations into first-order logic rules (e.g., “red light⟹\\implies⟹stop”) and incorporates these rules into a probabilistic graphical model, such as a Markov Logic Network (MLN). The MLN is trained to verify the predicted next actions using environmental attributes identified by attribute recognition models (e.g., detecting a red light) to form the predicates.\nAdditionally, we construct a Multimodal Retrieval-Augmented Generation (Multimodal RAG) model that leverages video data, control signals, and environmental attributes to learn more effectively from past similar driving experiences.By integrating PDCE, MLN, and Multimodal RAG,SafeAutosignificantly outperforms existing baselines across multiple datasets. This advancement paves the way for more accurate, reliable, and safer autonomous driving systems that effectively learn from experience, adhere to traffic regulations, and execute precise control actions.",
    "introduction": "Autonomous Driving (AD) systems(Kim et al.,2018; Jin et al.,2023; Hu et al.,2023; Jiang et al.,2023; Zhang et al.,2024b)have made significant strides in recent years, yet they often rely on separate modules for high-level decision-making (e.g., “the car should slow to a stop”) and low-level control signal prediction (e.g., providing the specific speed or steering angle for the next few moments). However, these two aspects are inherently correlated, as high-level actions directly guide low-level control signals. This modular design often overlooks this correlation, leading to inefficiencies and less cohesive driving behaviors.\nRecent advancements inMultimodal Large Language Models(MLLMs)(Liu et al.,2023b,a; Lin et al.,2023)offer a promising avenue to bridge high-level reasoning and low-level control in autonomous driving (AD), which provide a unified framework capable of processing and reasoning over multiple data modalities, such as images, videos, and text. Some recent works(Wang et al.,2023; Xu et al.,2024; Wang et al.,2024)have begun to leverage MLLMs to generate both high-level action descriptions and low-level control signals in an end-to-end manner. However, they are predominantly data-driven and often fail to perform at human levels due to several limitations.\n\n\nFirstly, for low-level action prediction, current approaches in adapting MLLMs generally follow two fashions. The first fashion treats the prediction of float numbers as text generation(Gruver et al.,2024; Xu et al.,2024), directly training the MLLM using cross-entropy (CE) loss for token prediction. Some variations(Brohan et al.,2023; Sima et al.,2023)of this method involve tokenizing the prediction range into several bins and adding new tokens for each bin into the LLM’s vocabulary, allowing the model to predict the corresponding bin token ID. However, these methods remain somewhat coarse compared to traditional regression techniques(Hu et al.,2023)using Mean Squared Error (MSE) loss. Alternatively, another fashion(Jin et al.,2024)employs a linear layer to decode the float number from the output hidden embeddings of the MLLM, enabling the use of MSE loss to train the model. While this approach may improve numerical accuracy, it compromises the autoregressive capability of the LLM, as the model can then only be purely used for numerical prediction and cannot perform any further QA-for example, handling high-level question-answering.\nAdditionally, regarding high-level action prediction, a significant limitation of current methods is their inability to effectively utilize both structured and unstructured knowledge when making decisions. Specifically, existing approaches often focus solely on data-driven techniques, inadequately incorporating structured knowledge such as traffic rules and safety constraints. Although some methods(Sima et al.,2023; Mao et al.,2023; Wang et al.,2024)attempt to include traffic regulations by embedding them into the model’s context, this implicit approach is insufficient. Due to the inherent tendency of MLLMs to hallucinate, they may still generate unsafe or illegal actions. Meanwhile, while RAG(Lewis et al.,2020)has been employed in language models(Semnani et al.,2023; Zhang et al.,2024a)to mitigate issues like hallucination by incorporating relevant information from external sources,Yuan et al. (2024)first propose to combine the rich multimodal data inherent in autonomous driving contexts—such as videos, images, and control signals—to learn from past driving experiences as unstructured knowledge.\n\n\nFigure 1:Overview of ourSafeAutopipeline for end-to-end high-level and low-level prediction in autonomous driving, featuring: (1) the Position-Dependent Cross-Entropy Loss (Section3.1) for improved low-level numerical predictions using soft, weighted digit probability distributions; (2) Knowledge-Enhanced Post-Safety Verification (Section3.2) with Markov Logic Networks to verify high-level actions against traffic rules; and (3) a Multimodal RAG (Section3.3) training method that incorporates similar driving experiences via text-based rankings for better context-aware decision-making.\n\n\nTo address these challenges, we propose a novel frameworkSafeAutothat enhances MLLMs for autonomous driving through three key contributions as shown inFigure1: (1)Position-Dependent Cross-Entropy (PDCE) Loss:We propose a PDCE loss that retains the autoregressive nature of the MLLM while behaving like an MSE loss during training. This loss function improves numerical prediction accuracy without compromising the model’s language generation abilities.\n(2)Knowledge-Enhanced Post-Safety Verification:We employMarkov Logic Networks(MLNs)(Richardson & Domingos,2006)to explicitly encode domain knowledge and structured traffic rules into the decision-making process of the MLLM. This knowledge-enabled reasoning allows us to verify and correct the high-level actions suggested by the MLLM, ensuring they comply with traffic regulations and safety constraints.\n(3)Multimodal RAG for Autonomous Driving:We introduce a method that utilizes video data, control signals, and the environmental predicates used in the MLN to retrieve similar driving experiences. By learning a joint embedding across these modalities based on the ranking derived from text description of the current scenario—which contain rich semantic information—we can effectively leverage past experiences to inform current decision-making.\nBy integrating these components,SafeAutoprovides a comprehensive solution to the challenges faced by current MLLMs in autonomous driving. We evaluate our approach on two benchmark datasets:BDD-X(Kim et al.,2018)andDriveLM(Sima et al.,2023), both featuring low-level control signals and high-level action descriptions. Our experimental results demonstrate significant improvements in both low-level control accuracy and high-level action prediction.\nFirst, for low-level prediction on the BDD-X dataset, it reduces the Root Mean Square Error (RMSE) for speed and course predictions by an additional 5.8% and 14.1% over the state-of-the-art (SOTA) baselines, respectively. Furthermore, on the DriveLM dataset, it decreases the Average Displacement Error (ADE) for motion prediction by 44.4%. Second, for high-level prediction on the BDD-X dataset, our method boosts high-level action performance beyond the SOTA by 28.0% under the CIDEr metric. Meanwhile, on the DriveLM dataset, it improves the high-level behavior prediction accuracy by an additional 13.0% over the SOTA.\n\n",
    "related_work": "Advancements in autonomous driving have produced comprehensive frameworks like UniAD(Hu et al.,2023), which integrates modules for tracking, mapping, motion prediction, and occupancy estimation for low-level planning. However, UniAD lacks high-level action descriptions and textual justifications. To address high-level explanations,Kim et al. (2018)proposed an attention-based video-to-text model generating explanations of current driving actions. Similarly, ADAPT(Jin et al.,2023)employs a video Swin Transformer(Liu et al.,2022)to extract video tokens for separate high-level and low-level action predictions.\n\n\nAutonomous Driving with MLLM.The emergence of MLLMs enables unified end-to-end generation of both high-level and low-level outputs. Most of these works often treat numerical control signals as text, training models using token prediction with cross-entropy loss. For example, DriveGPT4(Xu et al.,2024)just treats low-level control signals as text, fine-tuning an MLLM to sequentially predict high-level and low-level actions in a conversational manner using the BDD-X dataset. DriveLM-Agent(Sima et al.,2023), influenced by RT-2(Brohan et al.,2023), discretizes waypoints into bins, expanding the tokenizer vocabulary accordingly and fine-tuning the BLIP-2(Li et al.,2023). While this facilitates end-to-end training, it remains coarse compared to UniAD(Hu et al.,2023), which uses MSE loss. Time-LLM(Jin et al.,2024)decodes numerical predictions directly from output embeddings using a linear layer with MSE loss but diminishes the language model’s autoregressive capabilities, limiting high-level question-answering abilities. Additionally,Tan et al. (2024)suggest that employing the LLM backbone in this way does not enhance regression performance. In contrast, we propose a novel PDCE loss that adapts the cross-entropy loss for numerical training to behave more like MSE loss while preserving the model’s ability to perform high-level question-answering.\n\n\nSafety Guarantee.Further advancements involve integrating perception and planning tools into the MLLM context. Agent-Driver(Mao et al.,2023)incorporates modules from UniAD into an MLLM framework, serving as a language agent for autonomous driving. OmniDrive(Wang et al.,2024)introduces a framework combining 3D perception, reasoning, and planning. However, these methods remain purely data-driven and lack explicit safety verification for generated actions. Given the safety-critical nature of autonomous driving, ensuring that output actions are safe and compliant with traffic rules is essential. To address this, we incorporate extracted knowledge—specifically structured traffic rules—into a probabilistic graphical model like a Markov Logic Network (MLN) for explicit post-safety verification, which has been widely used in previous work(Yang et al.,2022; Zhang et al.,2023). Besides, RAGDriver(Yuan et al.,2024)further enhances reasoning by retrieving similar driving experiences through triplet loss-based metric learning.\nWe extend this approach by developing a more flexible and efficient retrieval system, directly training a joint embedding based on multimodal inputs to learn relative rankings from text similarity. Most importantly, we find that the incorporation of binary structured environmental predicates (e.g., the presence of a stop sign) from the previous reasoning components, namely MLNs, significantly improves retrieval performance.\n\n",
    "origin_citations": [
      "Planning-oriented autonomous driving.",
      "Textual explanations for self-driving vehicles.",
      "Adapt: Action-aware driving caption transformer.",
      "Video swin transformer.",
      "Drivegpt4: Interpretable end-to-end autonomous driving via large language model.",
      "Drivelm: Driving with graph visual question answering.",
      "Rt-2: Vision-language-action models transfer web knowledge to robotic control.",
      "Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.",
      "Time-llm: Time series forecasting by reprogramming large language models.",
      "Are language models actually useful for time series forecasting?",
      "A language agent for autonomous driving.",
      "Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning.",
      "Improving certified robustness via statistical learning with logical reasoning.",
      "Care: Certifiably robust learning with reasoning via variational inference.",
      "Rag-driver: Generalisable driving explanations with retrieval-augmented in-context learning in multi-modal large language model."
    ],
    "citations": [
      139854,
      1809,
      491865,
      392942,
      323895,
      347718,
      7672,
      212324,
      323767,
      872533,
      25497,
      859854,
      164450,
      229003,
      467011
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 85,
      "Final_Score": 87.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 85\nFinal_Score: 87"
    },
    "overall_score": 85.84,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2025/Conference",
        "venue": "ICML",
        "year": "2025",
        "openreview_id": "nKJGjovmZz",
        "forum": "nKJGjovmZz",
        "number": 15475,
        "title": "SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models",
        "normalized_title": "safeauto knowledge enhanced safe autonomous driving with multimodal foundation models",
        "abstract": "Traditional autonomous driving systems often struggle to connect high-level reasoning with low-level control, leading to suboptimal and sometimes unsafe behaviors. Recent advances in multimodal large language models (MLLMs), which process both visual and textual data, offer an opportunity to unify perception and reasoning. However, effectively embedding precise safety knowledge into MLLMs for autonomous driving remains a significant challenge.\nTo address this, we propose SafeAuto, a framework that enhances MLLM-based autonomous driving by incorporating both unstructured and structured knowledge. First, we introduce a Position-Dependent Cross-Entropy (PDCE) loss to improve low-level control signal predictions when values are represented as text. Second, to explicitly integrate safety knowledge, we develop a reasoning component that translates traffic rules into first-order logic (e.g., \"red light => stop\") and embeds them into a probabilistic graphical model (e.g., Markov Logic Network) to verify predicted actions using recognized environmental attributes.\nAdditionally, our Multimodal Retrieval-Augmented Generation (RAG) model leverages video, control signals, and environmental attributes to learn from past driving experiences. Integrating PDCE, MLN, and Multimodal RAG, SafeAuto outperforms existing baselines across multiple datasets, enabling more accurate, reliable, and safer autonomous driving. The code is available at https://github.com/AI-secure/SafeAuto.",
        "authors": [
          "Jiawei Zhang",
          "Xuan Yang",
          "Taiqi Wang",
          "Yu Yao",
          "Aleksandr Petiushko",
          "Bo Li"
        ],
        "decision": null,
        "pdate": 1746105707191,
        "cdate": 1737704408049
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2025/Conference",
          "venue": "ICML",
          "year": "2025",
          "openreview_id": "nKJGjovmZz",
          "forum": "nKJGjovmZz",
          "number": 15475,
          "title": "SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models",
          "normalized_title": "safeauto knowledge enhanced safe autonomous driving with multimodal foundation models",
          "abstract": "Traditional autonomous driving systems often struggle to connect high-level reasoning with low-level control, leading to suboptimal and sometimes unsafe behaviors. Recent advances in multimodal large language models (MLLMs), which process both visual and textual data, offer an opportunity to unify perception and reasoning. However, effectively embedding precise safety knowledge into MLLMs for autonomous driving remains a significant challenge.\nTo address this, we propose SafeAuto, a framework that enhances MLLM-based autonomous driving by incorporating both unstructured and structured knowledge. First, we introduce a Position-Dependent Cross-Entropy (PDCE) loss to improve low-level control signal predictions when values are represented as text. Second, to explicitly integrate safety knowledge, we develop a reasoning component that translates traffic rules into first-order logic (e.g., \"red light => stop\") and embeds them into a probabilistic graphical model (e.g., Markov Logic Network) to verify predicted actions using recognized environmental attributes.\nAdditionally, our Multimodal Retrieval-Augmented Generation (RAG) model leverages video, control signals, and environmental attributes to learn from past driving experiences. Integrating PDCE, MLN, and Multimodal RAG, SafeAuto outperforms existing baselines across multiple datasets, enabling more accurate, reliable, and safer autonomous driving. The code is available at https://github.com/AI-secure/SafeAuto.",
          "authors": [
            "Jiawei Zhang",
            "Xuan Yang",
            "Taiqi Wang",
            "Yu Yao",
            "Aleksandr Petiushko",
            "Bo Li"
          ],
          "decision": null,
          "pdate": 1746105707191,
          "cdate": 1737704408049
        }
      ]
    }
  },
  {
    "paper_id": 7835,
    "title": "OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models",
    "abstract": "Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements.\nAlthough recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization.\nTo tackle this issue, we introduce an Omnidirectionally calibrated Quantization (OmniQuant) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters.\nOmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights.\nOperating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization.\nFor instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples.\nExtensive experiments validate OmniQuant’s superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices.\nCodes are available athttps://github.com/OpenGVLab/OmniQuant.",
    "introduction": "Large language models (LLMs) such as GPT-4(Bubeck et al.,2023)and LLaMA(Touvron et al.,2023a), have demonstrated impressive performance across various natural language benchmarks(Hendrycks et al.,2020; Zellers et al.,2019).\nFurthermore, the language understanding capabilities inherent in LLMs can be successfully transferred into multimodal models(Mu et al.,2023; Xu et al.,2023; Zhang et al.,2023a; Huang et al.,2024;2023). Thereby, LLMs can be regarded as precursors to artificial general intelligence(Bubeck et al.,2023).\nHowever, the considerable computational and memory requirements of LLMs pose substantial challenges(Zhang et al.,2023b; Hu et al.,2023). For instance, the GPT-3 model(Brown et al.,2020)requires 350G of memory to load its parameters in FP16 format, which corresponds to the requirement of at least five A100-80G GPUs for inference.\nThis significant demand for computational resources and associated communication overheads impedes the practical deployment of LLMs in real-world applications.\n\n\nQuantization has shown to be promising to mitigate both computational and memory overhead in LLMs. In general, it comes in two types including post-training quantization (PTQ) and quantization-aware training (QAT).\nAlthough QAT can lead to more competitive accuracy than PTQ, it is not practical due to the high training cost because the whole model is trained with the awareness of the quantization process.\nAs a result, PTQ is commonly utilized in existing quantization methods on LLMs.\nFor example, lots of PTQ methods(Frantar et al.,2022; Lin et al.,2023; Dettmers et al.,2023b)reduce memory consumption by weight-only quantization which quantizes the weights while maintaining full-precision activation. To further reduce the computational overhead, another line of work(Xiao et al.,2023; Wei et al.,2022; Yuan et al.,2023; Wei et al.,2023; Liu et al.,2023a)employs weight-activation quantization which quantizes both weight and activation into low-bit values for the execution of low-bit matrix multiplication.\n\n\nFigure 1:(a) provides an overview of LLaMA-7B with W4A4 quantization, highlighting OmniQuant’s ability to achieve quantization-aware training (QAT) performance with post-training quantization (PTQ) time and data efficiency. (b) and (c) showcase the perplexity (low is better) of quantized LLaMA-13B across different bit-widths on WikiText2.\n\n\nExisting quantization methods have demonstrated significant achievements in various scenarios, including W4A16 (i.e.4-bit weight and 16-bit activation) weight-only quantization such as(Lin et al.,2023; Dettmers et al.,2023b; Lee et al.,2023), as well as W8A8 weight-activation quantization(Wei et al.,2023).\nHowever, they usually exhibit significant performance degradation when confronted with low-bit quantization, such as W2A16 and W4A4, as illustrated in Figure1(b & c).\nThis performance shortfall in low-bit quantization can be attributed to the fact that these methods(Frantar et al.,2022; Lin et al.,2023; Wei et al.,2023)primarily rely on handcrafted quantization parameters such as migration strength(Xiao et al.,2023)and scaling parameters(Wei et al.,2023), which often leads to lower performance.\nAlthough Quantization-Aware Training (QAT)(Liu et al.,2023b)is effective in determining the optimal quantization configurations, it introduces substantial training overhead in both training and data efficiency. It is thus hard to quantize LLMs with QAT-based techniques efficiently such as LLMQAT(Liu et al.,2023b). For instance, GPTQ(Frantar et al.,2022), a PTQ approach, can complete the quantization of LLaMA-13B in an hour using 128 samples on a single A100 GPU, while LLM-QAT(Liu et al.,2023b)requires 100k samples and hundreds of GPU hours.\nThis leads us to a central question:can we attain the performance of QAT, while maintaining the time and data efficiency of PTQ?\n\n\nThis paper introduces a novel quantization technique, OmniQuant, which effectively addresses the above question. OmniQuant achieves state-of-the-art performance across various quantization scenarios, particularly in low-bit settings, while preserving the time and data efficiency of PTQ, as illustrated in Figure1.\nUnlike Quantization-Aware Training (QAT)(Liu et al.,2023b)which involves cumbersome weight optimization, OmniQuant freezes the original full-precision weight and only incorporates a few learnable quantization parameters.\nAs shown in Figure2,\nOmniQuant consists of two key components that incorporate different types of learnable quantization parameters, including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). Specifically, LWC modulates the extreme values of weights by optimizing the clipping threshold. In the meanwhile, LET tackles activation outliers by learning mathematically equivalent transformations in a transformer encoder.\n\n\nInstead of jointly optimizing all parameters across the LLM, OmniQuant sequentially quantizes the parameters of one layer before moving on to the next under a block-wise quantization error minimization framework.\nIn this way,\nOminiQuant can be optimized efficiently using a simple Stochastic Gradient Descent (SGD) algorithm.\nThanks to the differentiable optimization, LWC and LET can be seamlessly integrated into the quantization.\nWe find that LWC can mitigate the difficulty in quantizing weights and LET further shifts the challenge of quantization from activations to weights, facilitating OmniQuant a versatile quantization framework for both weight-only and weight-activation quantization.\nNotably, OmniQuant introduces no extra computation or parameters for the quantized model because the clipping threshold in LWC and equivalent factors in LET can be fused into quantized weights.\n\n\nFigure 2:Characteristics of OmniQuant on LLaMA family.\n\n\nAs depicted in Figure2, OmniQuant is easy to implement even with limited resources. Especially, taking the LLaMA-2 model family (7B-70B) as an example, all models can be quantized on a single A100-40G GPU utilizing only 128 training samples. The training time ranges from 1 to 16 hours, depending on the size of the quantized model, which ranges from 7B to 70B.\nOwing to the seamless integration of LWC and LET achieved by differentiable optimization, OmniQuant exhibits superior performance compared to prior PTQ-based methods in various quantization settings. For example, when LLaMA-13B is quantized into W2A16, OmniQuant achieves a perplexity of13.2113.2113.2113.21, while GPTQ incurs a significant increase in perplexity to3832383238323832, as demonstrated in Figure1. A similar performance advancement is also observed in the W4A4 quantization.\n\n\nThe contributions ofOmniQuantare summarized as follows. 1) We formulate a novel quantization pipeline for LLM, OmniQuant, which freezes original full-precision weights while incorporating a restrained set of learnable parameters. OmniQuant imbues quantization with gradient updates while preserving the time and data efficiency of PTQ methods. 2) OmniQuant consists of Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). These strategies make full-precision weights and activations more amenable to quantization.\n3) Through extensive experiments, we demonstrate that OmniQuant outperforms previous methods across a spectrum of quantization settings (W416, W3A16, W2A16, W6A6, W4A4), various model families (OPT, LLaMA, LLaMA-2, LLaMA-2-chat, Falcon), and a range of model sizes (125M-180B). The computation speedup and memory reduction of OmniQuant are also demonstrated on real devices.\n\n",
    "related_work": "2.1Quantization Methods.Quantization reduces neural network bit-precision, leading to smaller models and faster inference. Current methods are largely divided into Quantization Aware Training (QAT)(Liu et al.,2023b)and Post-training Quantization (PTQ)(Xiao et al.,2023; Frantar et al.,2022). While QAT maintains performance by simulating quantization during training, its training cost makes it unsuitable for LLM. PTQ techniques like AdaRound(Nagel et al.,2020)and BRECQ(Li et al.,2021)use gradient optimization to determine optimal rounding, but tuning all weights is time-intensive for larger models. Thus, most LLM quantization methods(Xiao et al.,2023; Frantar et al.,2022; Dettmers et al.,2023b; Lee et al.,2023; Wei et al.,2023)prioritize training-free PTQ, which limit performance in lower-bit situations. Our goal is to integrate gradient updates in LLM quantization, mirroring QAT’s approach, while retaining PTQ’s efficiency.\n\n\n2.2Quantization of LLM.Considering the quantized object, exiting LLM quantization can be classified into two fields: weight-only quantization and weight-activation quantization.Weight-only quantization.Weight-only quantization focuses on converting weights to low-bit values. For instance, GPTQ(Frantar et al.,2022)uses block-wise reconstruction for 3/4-bit quantization. SpQR(Dettmers et al.,2023b), OWQ(Lee et al.,2023), and AWQ(Lin et al.,2023)emphasize the significance of weights tied to higher-magnitude activations. Therefore, SpQR and OWQ employ mixed-precision quantization to safeguard vital weights, while AWQ opts for channel-wise scaling to avoid mixed-precision’s hardware inefficiency.\nQlora(Dettmers et al.,2023a)and INT2.1(Chee et al.,2023)restore the capabilities of the quantized model through parameter-efficient fine-tuning. Our method, in contrast, enhances the quantization process directly, making OmniQuant complementary to Qlora and INT2.1.Weight-activation quantization.Weight-activation quantization compresses both weights and activations. SmoothQuant(Xiao et al.,2023), LLM.int8()(Dettmers et al.,2022), and Outlier Suppression(Wei et al.,2022)achieve W8A8 quantization by managing activation outliers. LLM.int8() uses mixed-precision decomposition, while the other two employ channel-wise scaling. Furthermore, Outlier Suppression+(Wei et al.,2023)adds channel-wise shifting to drive W6A6 quantization. Unlike previous heuristic designs, we use gradient optimization and expand equivalent transformations to attention mechanisms, further boosting the K/V cache quantization.\nRecently, RPTQ(Yuan et al.,2023)and LLM-QAT(Liu et al.,2023b)have achieved W4A4 quantization. However, RPTQ adopts deployment-unfriendly group-wise activation quantization, and LLM-QAT employs time-consuming QAT. In distinction from RPTQ and LLM-QAT, we achieve W4A4 quantization through deployment-friendly per-token quantization and maintain the PTQ efficiency.Figure 3:Details of OmniQuantin a transformer block. Note that all learnable parameters can be eliminated after quantization.\n\n",
    "origin_citations": [
      "Llm-qat: Data-free quantization aware training for large language models.",
      "Smoothquant: Accurate and efficient post-training quantization for large language models.",
      "Gptq: Accurate post-training quantization for generative pre-trained transformers.",
      "Up or down? adaptive rounding for post-training quantization.",
      "Brecq: Pushing the limit of post-training quantization by block reconstruction.",
      "Spqr: A sparse-quantized representation for near-lossless llm weight compression.",
      "Owq: Lessons learned from activation outliers for weight quantization in large language models.",
      "Outlier suppression+: Accurate quantization of large language models by equivalent and optimal shifting and scaling.",
      "Awq: Activation-aware weight quantization for llm compression and acceleration.",
      "Qlora: Efficient finetuning of quantized llms.",
      "Quip: 2-bit quantization of large language models with guarantees.",
      "Llm. int8 (): 8-bit matrix multiplication for transformers at scale.",
      "Outlier suppression: Pushing the limit of low-bit transformer language models.",
      "Rptq: Reorder-based post-training quantization for large language models."
    ],
    "citations": [
      257203,
      245034,
      945903,
      191955,
      118706,
      306939,
      1088952,
      54257,
      308051,
      2253,
      468609,
      888003,
      176804,
      20700
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "8Wuvhh0LYW",
        "forum": "8Wuvhh0LYW",
        "number": 1585,
        "title": "OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models",
        "normalized_title": "omniquant omnidirectionally calibrated quantization for large language models",
        "abstract": "Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization ($\\textbf{OmniQuant}$) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \n\\url{https://github.com/OpenGVLab/OmniQuant}.",
        "authors": [
          "Wenqi Shao",
          "Mengzhao Chen",
          "Zhaoyang Zhang",
          "Peng Xu",
          "Lirui Zhao",
          "Zhiqian Li",
          "Kaipeng Zhang",
          "Peng Gao",
          "Yu Qiao",
          "Ping Luo"
        ],
        "decision": null,
        "pdate": 1705410824878,
        "cdate": 1695091545031
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "8Wuvhh0LYW",
          "forum": "8Wuvhh0LYW",
          "number": 1585,
          "title": "OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models",
          "normalized_title": "omniquant omnidirectionally calibrated quantization for large language models",
          "abstract": "Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization ($\\textbf{OmniQuant}$) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \n\\url{https://github.com/OpenGVLab/OmniQuant}.",
          "authors": [
            "Wenqi Shao",
            "Mengzhao Chen",
            "Zhaoyang Zhang",
            "Peng Xu",
            "Lirui Zhao",
            "Zhiqian Li",
            "Kaipeng Zhang",
            "Peng Gao",
            "Yu Qiao",
            "Ping Luo"
          ],
          "decision": null,
          "pdate": 1705410824878,
          "cdate": 1695091545031
        }
      ]
    }
  },
  {
    "paper_id": 43521,
    "title": "Hollowed Net for On-Device Personalization of Text-to-Image Diffusion Models",
    "abstract": "Recent advancements in text-to-image diffusion models have enabled the personalization of these models to generate custom images from textual prompts. This paper presents an efficient LoRA-based personalization approach for on-device subject-driven generation, where pre-trained diffusion models are fine-tuned with user-specific data on resource-constrained devices. Our method, termed Hollowed Net, enhances memory efficiency during fine-tuning by modifying the architecture of a diffusion U-Net to temporarily remove a fraction of its deep layers, creating ahollowedstructure. This approach directly addresses on-device memory constraints and substantially reduces GPU memory requirements for training, in contrast to previous methods that primarily focus on minimizing training steps and reducing the number of parameters to update. Additionally, the personalized Hollowed Net can be transferred back into the original U-Net, enabling inference without additional memory overhead. Quantitative and qualitative analyses demonstrate that our approach not only reduces training memory to levels as low as those required for inference but also maintains or improves personalization performance compared to existing methods.",
    "introduction": "Recent research on text-to-image (T2I) diffusion models[1,2], which generate high-resolution images from text prompts, has increasingly focused on personalizing and customizing these generative models effectively[3,4,5,6,7]. A primary approach, termed subject-driven generation[5], involves fine-tuning pre-trained diffusion models with a few user-specific images to generate varied representations of a subject using simple text prompts. This allows users to create personalized images of specific subjects, such as family, friends, pets, or personal items, with preferred appearances, backgrounds, and styles. Such capabilities enable creative applications including art renditions, property modifications, and accessorization.\n\n\nFrom a practical standpoint, implementing subject-driven generation on-device offers significant benefits in efficiency and privacy. By operating independently of congested cloud servers or networks, users can generate personalized images anywhere at no additional cost and do not need to compromise their privacy as all data and personal information remain on the device.\n\n\nDespite extensive research aimed at efficiently personalizing diffusion models, limited attention has been paid to memory I/O, a critical bottleneck in on-device learning. Recent studies have mainly explored two strategies: (1) decreasing the number of training steps and (2) reducing the number of updating parameters. The first methods[8,9,10,11,12]utilize additional large pre-trained models to generate a set of personalized Low-Rank Adaptation (LoRA) parameters[13], text embeddings, or image prompts from a user-specific image. This strategy provides a better initial setup for personalizing the diffusion models, effectively reducing required training steps. Some models[10,11,12]even support zero-shot personalization, although they underline that further fine-tuning can enhance personalization quality and address failure cases. Nonetheless, these methods are not viable for environments with severely limited computational resources, as they necessitate additional inference using large pre-trained models (e.g., 2.7B parameters for BLIP-2 in BLIP-Diffusion[12]and 2.5B for apprentice models in SuTI[11]), which are substantially larger than standard diffusion models (e.g., 1B for Stable Diffusion v2[2]), making their application challenging in on-device settings.\n\n\nThe second approach[7,14], often involving LoRA, aims to reduce the number of updating parameters by limiting updates to specific layers or decomposing weight matrices. However, even with fewer parameters to update, these parameters reside within large pre-trained models, and thus the backward pass through the large models is required to compute gradients. Given limited computational resources, where even simple inference tasks with diffusion models can strain GPU memory, performing backpropagation while keeping the entire diffusion model in GPU memory remains a significant limitation.\n\n\nA promising approach to address these challenges is side-tuning[15,16,17,18], which fine-tunes a smaller auxiliary network rather than directly updating the parameters of a large pre-trained network. This method significantly reduces the heavy memory costs associated with computing backpropagation on the larger network. Particularly for Natural Language Processing (NLP) tasks, Ladder Side Tuning (LST)[18]has proven effective, reducing the memory costs required for fine-tuning large language models (LLMs) by 69 percent. However, applying LST directly to diffusion U-Nets presents significant challenges. Unlike transformer layers in LLMs, which maintain consistent input and output dimensions, diffusion U-Nets have varying spatial dimensions and channels, as well as skip-connections across different blocks. Additionally, the requirements for structural pruning and weight initialization to build side-tuning networks further complicate the rapid adaptability of LST to personalization tasks across different subjects and domains.\n\n\nTo this end, we introduce a novel personalization technique calledHollowed Net, which is illustrated in Fig.1. Based on our observation that deep layers in the middle of diffusion U-Nets play significantly less important roles than the rest of the layers, we propose to fine-tune LoRA parameters for the personalization using Hollowed Net, a layer-pruned U-Net featuring a central hollow, which is constructed by temporarily removing the\nmiddle deep layers from the pre-trained diffusion U-Net. By utilizing the symmetrical \"U-shape\" architecture of the diffusion U-Net, we avoid complicated processes of applying structural pruning and weight initialization to build a side network, and neither additional models nor extensive pre-training with large datasets are required.\n\n\nBy fine-tuning LoRA parameters using Hollowed Net, we can significantly reduce the memory needed for storing model weights in GPU. Once the LoRA parameters are fine-tuned with Hollowed Net, they can be seamlessly transferred back to the original Diffusion U-Net for inference, without requiring any additional memory beyond the small set of transferred parameters. Our experiments demonstrate that Hollowed Net enables achieving performance that is comparable to or better than the direct fine-tuning with LoRA, while using 26 percent less GPU memory, which is only 11 percent increased GPU memory relative to an inference.\n\n\nTo the best of our knowledge, Hollowed Net is the first technique that addresses subject-driven generation in terms of memory efficiency. Our method shows how T2I diffusion models can be fine-tuned under extremely limited computational resources with as low GPU memory as required for inference. Furthermore, it is important to note that our method does not preclude the use of previously described strategies for efficient personalization. Both enhanced parameter-efficient strategies and improved initializations with additional pre-trained models can be integrated with our approach to further increase efficiency according to given resource constraints.\n\n\nOur contributions can be summarized as follows:•We introduceHollowed Net, a novel personalization technique for T2I diffusion models under limited computational resources. Our method significantly reduces the memory demands on GPU to levels as low as those required for inference, while maintaining a high-fidelity personalization capacity. This demonstrates its potential as a feasible on-device learning solution for resource-constrained devices.•Our method provides a scalable and controllable solution for on-device learning. As this method does not require any additional models or pre-training with large datasets, it is easily scalable to other architectures such as SDXL and Transformers. Moreover, we can simply adjust the fraction of hollowed layers to control the trade-offs between performance and memory requirements, depending on the target application and resources.•Unlike previous side-tuning methods, Hollowed Net does not need to be retained for inference. The LoRA parameters fine-tuned with Hollowed Net can be seamlessly transferred back to its original network, enabling inference with no additional memory cost.\n\n\nFigure 1:The LoRA personalization with Hollowed Net for resource-constrained environments. The input image is from the DreamBooth dataset[5].\n\n",
    "related_work": "2.1Efficient Personalization of T2I Diffusion ModelsRecent research on the personalization of T2I diffusion models has introduced various methods to fine-tune the models for generating diverse images of user-specific subjects from a few given images. Two foundational works in this area are Textual Inversion and DreamBooth[3,5]. Textual Inversion[3]aims to learn new text embeddings to represent a given subject, while DreamBooth[5]proposes fine-tuning an entire diffusion model to align the subject with a unique token.Building on these foundational works, recent research has focused on enhancing the efficiency of this personalization process, primarily through two approaches. The first approach involves decreasing the number of training steps, mostly by utilizing an additional large pre-trained model. A popular method is to use a pre-trained image/multi-modal encoder to generate personalized text embeddings or image prompts from a user-specific image[9,12,10]. Other recent works[8,11]propose utilizing a set of pre-optimized LoRA parameters or millions of fine-tuned expert models to pre-initialize for efficient fine-tuning or enable zero-shot generation with in-context learning. While these models demonstrate significant reductions in the number of training steps, the requirement for additional large pre-trained models limits their application to on-device settings. Moreover, models with zero-shot personalization capacities[11,12,10]cannot be a one-size-fits-all solution for addressing different types of user-subject prompts. These models often struggle with flexibility in editing subjects or maintaining subject fidelity, and in these cases, additional fine-tuning with specific subjects is needed to further enhance their personalization capacity[11,12].On the other hand, another stream of work adapts parameter-efficient fine-tuning (PEFT) approaches. These methods demonstrate significant reductions in the number of training parameters by limiting updates to a small subset of model weights in cross-attention layers[7]or further reducing the updating parameters by applying singular vector decomposition to weight matrices[14]. However, these methods are still limited in environments with extremely low computational resources, as they require backpropagation over large diffusion models and do not reduce memory usage from the model weights. Therefore, it is crucial to explore new approaches for personalizing T2I diffusion models in resource-limited settings, as we propose with our novel method, Hollowed Net. Notably, our method can be integrated with previously discussed techniques to further improve efficiency based on specific resource constraints.\n\n\n2.2Fine-Tuning with Side NetworksThe idea of of side-tuning has been introduced by Zhang et al.[15], proposing the training of a lightweight \"side\" network instead of directly fine-tuning a pre-trained network for adaptation. In terms of efficiency, Cai et al.[16]has demonstrated an additional lightweight residual module can reduce memory overhead associated with the activations of the original network. Similarly, AuxAdapt[17]has shown that a small auxiliary network can be fine-tuned to adjust the main network’s decisions, enabling efficient test-time adaptation for video semantic segmentation tasks.In the context of generative models, LST[18]has demonstrated the effectiveness of side networks for different NLP tasks with LLMs by introducing a small side network that takes intermediate activations of the main network as input via shortcut connections. However, directly applying LST to diffusion U-Nets poses challenges due to varying spatial dimensions, channel sizes, and skip-connections across blocks, unlike the consistent dimensions in transformer layers of LLMs. Furthermore, the structural pruning and specific weight initialization required to construct side-tuning networks complicate LST’s adaptability for personalized tasks across a range of subjects and domains.\n\n\n2.3Layer Pruning of Large Generative ModelsSeveral concurrent works demonstrate that layer-pruning methods can be applied to generative models, particularly for NLP tasks. Gromov et al.[19]suggest that for fine-tuning LLM models, up to 40% of deep layers can be removed, while still achieving comparable results. The authors propose that the optimal block of layers to prune can be selected based on similarity across layers. Similarly, Kim et al.[20]also propose a depth-pruning approach by evaluating block-level importance.These approaches differ from ours due to the distinct characteristics of LLMs versus diffusion U-Nets. The aforementioned approaches involve the complete removal of deep layers for both fine-tuning and inference, considering that those layers store less critical knowledge. However, our study finds that the deep layers of diffusion U-Nets may be less involved with personalization but still contain crucial high-level image features for generating high-fidelity images. Thus, their removal can lead to severe performance degradation, even with additional pre-training[21], as shown in AppendixA. This highlights the importance of our two-stage fine-tuning strategy, which excludes layers during fine-tuning to reduce memory overhead while preserving the knowledge from these excluded layers throughout both training and inference stages.\n\n",
    "origin_citations": [
      "An image is worth one word: Personalizing text-to-image generation using textual inversion.",
      "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation.",
      "Encoder-based domain tuning for fast personalization of text-to-image models.",
      "Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing.",
      "Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation.",
      "Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models.",
      "Subject-driven text-to-image generation via apprenticeship learning.",
      "Multi-concept customization of text-to-image diffusion.",
      "Svdiff: Compact parameter space for diffusion fine-tuning.",
      "Side-tuning: a baseline for network adaptation via additive side networks.",
      "Tinytl: Reduce activations, not trainable parameters for efficient on-device learning.",
      "Auxadapt: Stable and efficient test-time adaptation for temporally consistent video semantic segmentation.",
      "Lst: Ladder side-tuning for parameter and memory efficient transfer learning.",
      "The unreasonable ineffectiveness of the deeper layers.",
      "Shortened llama: A simple depth pruning for large language models.",
      "Bk-sdm: A lightweight, fast, and cheap version of stable diffusion."
    ],
    "citations": [
      236013,
      67241,
      226314,
      259157,
      225547,
      184967,
      21143,
      222136,
      495091,
      979844,
      521196,
      512904,
      411359,
      423818,
      100564,
      258505
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "Pa8jsrdOnU",
        "forum": "Pa8jsrdOnU",
        "number": 14149,
        "title": "Hollowed Net for On-Device Personalization of Text-to-Image Diffusion Models",
        "normalized_title": "hollowed net for on device personalization of text to image diffusion models",
        "abstract": "Recent advancements in text-to-image diffusion models have enabled the personalization of these models to generate custom images from textual prompts. This paper presents an efficient LoRA-based personalization approach for on-device subject-driven generation, where pre-trained diffusion models are fine-tuned with user-specific data on resource-constrained devices. Our method, termed Hollowed Net, enhances memory efficiency during fine-tuning by modifying the architecture of a diffusion U-Net to temporarily remove a fraction of its deep layers, creating a hollowed structure. This approach directly addresses on-device memory constraints and substantially reduces GPU memory requirements for training, in contrast to previous methods that primarily focus on minimizing training steps and reducing the number of parameters to update. Additionally, the personalized Hollowed Net can be transferred back into the original U-Net, enabling inference without additional memory overhead. Quantitative and qualitative analyses demonstrate that our approach not only reduces training memory to levels as low as those required for inference but also maintains or improves personalization performance compared to existing methods.",
        "authors": [
          "Wonguk Cho",
          "Seokeon Choi",
          "Debasmit Das",
          "Matthias Reisser",
          "Taesup Kim",
          "Sungrack Yun",
          "Fatih Porikli"
        ],
        "decision": null,
        "pdate": 1727288061452,
        "cdate": 1715746741726
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "Pa8jsrdOnU",
          "forum": "Pa8jsrdOnU",
          "number": 14149,
          "title": "Hollowed Net for On-Device Personalization of Text-to-Image Diffusion Models",
          "normalized_title": "hollowed net for on device personalization of text to image diffusion models",
          "abstract": "Recent advancements in text-to-image diffusion models have enabled the personalization of these models to generate custom images from textual prompts. This paper presents an efficient LoRA-based personalization approach for on-device subject-driven generation, where pre-trained diffusion models are fine-tuned with user-specific data on resource-constrained devices. Our method, termed Hollowed Net, enhances memory efficiency during fine-tuning by modifying the architecture of a diffusion U-Net to temporarily remove a fraction of its deep layers, creating a hollowed structure. This approach directly addresses on-device memory constraints and substantially reduces GPU memory requirements for training, in contrast to previous methods that primarily focus on minimizing training steps and reducing the number of parameters to update. Additionally, the personalized Hollowed Net can be transferred back into the original U-Net, enabling inference without additional memory overhead. Quantitative and qualitative analyses demonstrate that our approach not only reduces training memory to levels as low as those required for inference but also maintains or improves personalization performance compared to existing methods.",
          "authors": [
            "Wonguk Cho",
            "Seokeon Choi",
            "Debasmit Das",
            "Matthias Reisser",
            "Taesup Kim",
            "Sungrack Yun",
            "Fatih Porikli"
          ],
          "decision": null,
          "pdate": 1727288061452,
          "cdate": 1715746741726
        }
      ]
    }
  },
  {
    "paper_id": 45153,
    "title": "Bridging the Divide: Reconsidering Softmax and Linear Attention",
    "abstract": "Widely adopted in modern Vision Transformer designs, Softmax attention can effectively capture long-range visual information; however, it incurs excessive computational cost when dealing with high-resolution inputs. In contrast, linear attention naturally enjoys linear complexity and has great potential to scale up to higher-resolution images. Nonetheless, the unsatisfactory performance of linear attention greatly limits its practical application in various scenarios.\nIn this paper, we take a step forward to close the gap between the linear and Softmax attention with novel theoretical analyses, which demystify the core factors behind the performance deviations.\nSpecifically, we present two key perspectives to understand and alleviate the limitations of linear attention: the injective property and the local modeling ability. Firstly, we prove that linear attention is not injective,\nwhich is prone to assign identical attention weights to different query vectors, thus adding to severe semantic confusion since different queries correspond to the same outputs.\nSecondly, we confirm that effective local modeling is essential for the success of Softmax attention,\nin which linear attention falls short.\nThe aforementioned two fundamental differences significantly contribute to the disparities between these two attention paradigms,\nwhich is demonstrated by our substantial empirical validation in the paper.\nIn addition, more experiment results indicate that linear attention, as long as endowed with these two properties, can outperform Softmax attention across various tasks while maintaining lower computation complexity.\nCode is available athttps://github.com/LeapLabTHU/InLine.",
    "introduction": "Recent years have witnessed the\nunprecedented success\nof Transformer and attention[32]in the field of computer vision[7]. Softmax attention, also known as dot-product attention, has demonstrated remarkable expressive power, leading to state-of-the-art performance across various vision tasks[30,2,23,35]. However, applying Softmax attention in vision also faces challenges. The quadratic complexity of Softmax attention results in prohibitively high computational cost when applied with a global receptive field. Previous works[33,19,36,6,45]have strived to reduce the computation complexity by restricting receptive fields or introducing sparsity. Although effective, these approaches inevitably compromise Softmax attention’s ability for long-range modeling and scalability.\n\n\nThe nature of Softmax attention forces to compute the dot-products between queries and keysQ⁢K⊤∈ℝN×N𝑄superscript𝐾topsuperscriptℝ𝑁𝑁QK^{\\top}\\!\\in\\!\\mathbb{R}^{N\\times{}N}italic_Q italic_K start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_N × italic_N end_POSTSUPERSCRIPTat first, and then aggregates valuesV∈ℝN×C𝑉superscriptℝ𝑁𝐶V\\!\\in\\!\\mathbb{R}^{N\\times{}C}italic_V ∈ blackboard_R start_POSTSUPERSCRIPT italic_N × italic_C end_POSTSUPERSCRIPTby the normalized scoreSoftmax⁢(Q⁢K⊤/d)Softmax𝑄superscript𝐾top𝑑\\mathrm{Softmax}(QK^{\\top}/\\sqrt{d})roman_Softmax ( italic_Q italic_K start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT / square-root start_ARG italic_d end_ARG ), which accounts for the quadratic𝒪⁢(N2)𝒪superscript𝑁2\\mathcal{O}(N^{2})caligraphic_O ( italic_N start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT )complexityw.r.t.the sequence lengthN𝑁Nitalic_N.\nOn the contrary, linear attention relaxes the similarity score betweenQ𝑄Qitalic_QandK𝐾Kitalic_Kfrom Softmax to other functions which can be decomposed into kernels,i.e., the linear attention replaces the original score functionSim⁢(Q,K)=Softmax⁢(Q⁢K⊤/d)Sim𝑄𝐾Softmax𝑄superscript𝐾top𝑑\\mathrm{Sim}(Q,K)\\!=\\!\\mathrm{Softmax}(QK^{\\top}/\\sqrt{d})roman_Sim ( italic_Q , italic_K ) = roman_Softmax ( italic_Q italic_K start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT / square-root start_ARG italic_d end_ARG )withSim⁢(Q,K)=ϕ⁢(Q)⁢ϕ⁢(K)⊤Sim𝑄𝐾italic-ϕ𝑄italic-ϕsuperscript𝐾top\\mathrm{Sim}(Q,K)\\!=\\!\\phi(Q)\\phi(K)^{\\top}roman_Sim ( italic_Q , italic_K ) = italic_ϕ ( italic_Q ) italic_ϕ ( italic_K ) start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT, whereϕ⁢(⋅)italic-ϕ⋅\\phi(\\cdot)italic_ϕ ( ⋅ )is the kernel function.\nThis substitution enables a change in the computation order from(ϕ⁢(Q)⁢ϕ⁢(K)⊤)⁢Vitalic-ϕ𝑄italic-ϕsuperscript𝐾top𝑉\\left(\\phi(Q)\\phi(K)^{\\top}\\right)V( italic_ϕ ( italic_Q ) italic_ϕ ( italic_K ) start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT ) italic_Vtoϕ⁢(Q)⁢(ϕ⁢(K)⊤⁢V)italic-ϕ𝑄italic-ϕsuperscript𝐾top𝑉\\phi(Q)\\left(\\phi(K)^{\\top}{}V\\right)italic_ϕ ( italic_Q ) ( italic_ϕ ( italic_K ) start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT italic_V )based on the associative law of matrix multiplication, reducing the complexity from𝒪⁢(N2)𝒪superscript𝑁2\\mathcal{O}(N^{2})caligraphic_O ( italic_N start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT )to𝒪⁢(N)𝒪𝑁\\mathcal{O}(N)caligraphic_O ( italic_N )w.r.t.the sequence lengthN𝑁Nitalic_N.\nNevertheless, every coin has two sides. Linear attention proves to be less effective than Softmax attention[3,27,40], whose poor expressive power limits its practical application. Although many pieces of research[27,28,22,9,40]have attempted to alleviate this issue in different ways, we still do not have a complete understanding of the key factors that contribute to the gap between linear and Softmax attention.\n\n\nIn this paper, we delve into the fundamental differences between linear and Softmax attention, offering two insightful perspectives to demystify the topic: the injective property and local modeling capability. Firstly, we consider attention as a function that maps a query to an attention score. We find that the injectivity of this attention function greatly affects the performance of the model. Specifically, if the attention function is not injective, different queries will induce identical attention distributions, leading to severe semantic confusion within the feature space.\nOur rigorous analysis has demonstrated that the Softmax attention function is an injective function, whereas the linear attention function is not.\nTherefore, linear attention is vulnerable to the semantic confusion problem, which largely leads to its insufficient expressiveness.\nSecondly, our analysis of the attention weight distribution has confirmed that the success of Softmax attention is not solely dependent on its strong long-range modeling capabilities. Effective local modeling is also crucial to achieving optimal outcomes.\n\n\nTo validate our analyses, we present two simple yet effective methods to endow linear attention with the injective property and the local modeling ability, respectively. The widely employed Swin Transformer[19]architecture is used to validate our findings. The results highlight the importance of both properties in the gap between linear and Softmax attention. Moreover, comprehensive experiments demonstrate that linear attention, endowed with these two properties, outperforms the widely used Softmax attention across diverse tasks.\n\n\nOur main contributions and takeaways are summarized as follows:(1) Injectivityis a key disparity between linear and Softmax attention. While Softmax attention is injective, the non-injective nature of linear attention causes semantic confusion and severely impairs model performance. To the best of our knowledge, our work is the first to conceptualize attention as a mapping function and prove the vital importance of its injective property.(2) Local modelingis still essential to the effectiveness of the attention mechanism, even though it is renowned for its large receptive field and outstanding long-range modeling ability.(3)We challenge the viewpoint that linear attention is inferior to Softmax attention and demonstrate that with the above two properties, linear attention can outperform Softmax attention while maintaining lower computation complexity.\n\n",
    "related_work": "Vision Transformer and Softmax Attention.Vision Transformer[7]is the pioneer work that introduces self-attention to vision. Since then, attention has found success in various vision tasks[7,2,24]. The widely used attention mechanism is Softmax attention[32], also known as dot-product attention, which computes the similarity between all query-key pairs. Although effective, its quadratic computation complexity leads to unmanageable cost when processing global feature maps. Therefore, various approaches[19,34,36,14,45,12]have been proposed to reduce the computational overhead of Softmax attention. PVT[33]employs downsampling of keys and values to reduce computational complexity. Swin Transformer[19]restricts the receptive field by introducing window attention pattern. NAT[14]mimics convolution and calculates attention within the neighborhood of each feature, and DAT[36]presents an input-dependent sparse attention pattern.\n\n\nLinear Attention.As opposed to Softmax attention, Linear attention is another attention paradigm with a natural linear complexity of𝒪⁢(N)𝒪𝑁\\mathcal{O}(N)caligraphic_O ( italic_N ). Linear attention replaces Softmax with kernel functions, thereby reducing computational complexity to linear through a change in computation order. Nonetheless, prior studies[27,3,29,40,11]have demonstrated that linear attention performs markedly worse than Softmax attention. CosFormer attributed this discrepancy to the efficient re-weighting mechanism of Softmax attention and proposed cosine re-weighting to enhance linear attention. Nyströmformer[38]and SOFT[22]use matrix decomposition to further approximate Softmax operation. Efficient Attention[28]applies Softmax function to queries and keys. TransNormer[26]identifies that unbounded gradients and attention dilution harm linear attention. FLatten Transformer[9]introduces a focused function to address the over-smoothing issue. MLLA[10]draws inspiration from Mamba[8]to improve linear attention.\n\n\nDespite their elegant outcomes, the fundamental reason for the disparity between linear attention and Softmax attention remains unclear. In this work, we perform an in-depth analysis of the disparities between linear and Softmax attention, identifying two crucial properties of high-performance Softmax attention: injectivity and local modeling ability. We present both theoretical proofs and experimental verification to validate our findings.\n\n\nFigure 1:An illustration of injective property and confusion problem. Non-injectivity leads to various semantic confusions in linear attention when different kernel functions are employed. (a) Withϕ⁢(⋅)=ReLU⁢(⋅)italic-ϕ⋅ReLU⋅\\phi(\\cdot)=\\mathrm{ReLU}(\\cdot)italic_ϕ ( ⋅ ) = roman_ReLU ( ⋅ ), linear attention assigns the same attention values to collinear queries of varying lengths. (b) Usingϕ(⋅)=ReLU(A⋅+b)\\phi(\\cdot)=\\mathrm{ReLU}(A\\cdot+b)italic_ϕ ( ⋅ ) = roman_ReLU ( italic_A ⋅ + italic_b ), linear attention faces severe confusion problem, producing identical attention distribution for certain queries with different directions and lengths.\n\n",
    "origin_citations": [
      "An image is worth 16x16 words: Transformers for image recognition at scale.",
      "End-to-end object detection with transformers.",
      "Adanat: Exploring adaptive policy for token-based image generation.",
      "Attention is all you need.",
      "Swin transformer: Hierarchical vision transformer using shifted windows.",
      "Pvt v2: Improved baselines with pyramid vision transformer.",
      "Vision transformer with deformable attention.",
      "Neighborhood attention transformer.",
      "Biformer: Vision transformer with bi-level routing attention.",
      "Dynamic perceiver for efficient visual recognition.",
      "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions.",
      "cosformer: Rethinking softmax in attention.",
      "Rethinking attention with performers.",
      "Vicinity vision transformer.",
      "Castling-vit: Compressing self-attention via switching towards linear-angular attention at vision transformer inference.",
      "Agent attention: On the integration of softmax and linear attention.",
      "Nyströmformer: A nyström-based algorithm for approximating self-attention.",
      "Soft: Softmax-free transformer with linear complexity.",
      "Efficient attention: Attention with linear complexities.",
      "The devil in linear transformer.",
      "Flatten transformer: Vision transformer using focused linear attention.",
      "Demystify mamba in vision: A linear attention perspective.",
      "Mamba: Linear-time sequence modeling with selective state spaces."
    ],
    "citations": [
      732,
      7360,
      907541,
      38,
      1501,
      5774,
      256286,
      63187,
      704886,
      49752,
      2577,
      175109,
      8384,
      81164,
      245008,
      270362,
      119364,
      513079,
      7827,
      336304,
      182609,
      1029595,
      274507
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 90,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 85\nSynthesis_Analysis: 90\nFinal_Score: 85"
    },
    "overall_score": 85.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "RSiGFzQapl",
        "forum": "RSiGFzQapl",
        "number": 2854,
        "title": "Bridging the Divide: Reconsidering Softmax and Linear Attention",
        "normalized_title": "bridging the divide reconsidering softmax and linear attention",
        "abstract": "Widely adopted in modern Vision Transformer designs, Softmax attention can effectively capture long-range visual information; however, it incurs excessive computational cost when dealing with high-resolution inputs. In contrast, linear attention naturally enjoys linear complexity and has great potential to scale up to higher-resolution images. Nonetheless, the unsatisfactory performance of linear attention greatly limits its practical application in various scenarios. In this paper, we take a step forward to close the gap between the linear and Softmax attention with novel theoretical analyses, which demystify the core factors behind the performance deviations. Specifically, we present two key perspectives to understand and alleviate the limitations of linear attention: the injective property and the local modeling ability. Firstly, we prove that linear attention is not injective, which is prone to assign identical attention weights to different query vectors, thus adding to severe semantic confusion since different queries correspond to the same outputs. Secondly, we confirm that effective local modeling is essential for the success of Softmax attention, in which linear attention falls short. The aforementioned two fundamental differences significantly contribute to the disparities between these two attention paradigms, which is demonstrated by our substantial empirical validation in the paper. In addition, more experiment results indicate that linear attention, as long as endowed with these two properties, can outperform Softmax attention across various tasks while maintaining lower computation complexity. Code is available at https://github.com/LeapLabTHU/InLine.",
        "authors": [
          "Dongchen Han",
          "Yifan Pu",
          "Zhuofan Xia",
          "Yizeng Han",
          "Xuran Pan",
          "Xiu Li",
          "Jiwen Lu",
          "Shiji Song",
          "Gao Huang"
        ],
        "decision": null,
        "pdate": 1727287702885,
        "cdate": 1715173511687
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "RSiGFzQapl",
          "forum": "RSiGFzQapl",
          "number": 2854,
          "title": "Bridging the Divide: Reconsidering Softmax and Linear Attention",
          "normalized_title": "bridging the divide reconsidering softmax and linear attention",
          "abstract": "Widely adopted in modern Vision Transformer designs, Softmax attention can effectively capture long-range visual information; however, it incurs excessive computational cost when dealing with high-resolution inputs. In contrast, linear attention naturally enjoys linear complexity and has great potential to scale up to higher-resolution images. Nonetheless, the unsatisfactory performance of linear attention greatly limits its practical application in various scenarios. In this paper, we take a step forward to close the gap between the linear and Softmax attention with novel theoretical analyses, which demystify the core factors behind the performance deviations. Specifically, we present two key perspectives to understand and alleviate the limitations of linear attention: the injective property and the local modeling ability. Firstly, we prove that linear attention is not injective, which is prone to assign identical attention weights to different query vectors, thus adding to severe semantic confusion since different queries correspond to the same outputs. Secondly, we confirm that effective local modeling is essential for the success of Softmax attention, in which linear attention falls short. The aforementioned two fundamental differences significantly contribute to the disparities between these two attention paradigms, which is demonstrated by our substantial empirical validation in the paper. In addition, more experiment results indicate that linear attention, as long as endowed with these two properties, can outperform Softmax attention across various tasks while maintaining lower computation complexity. Code is available at https://github.com/LeapLabTHU/InLine.",
          "authors": [
            "Dongchen Han",
            "Yifan Pu",
            "Zhuofan Xia",
            "Yizeng Han",
            "Xuran Pan",
            "Xiu Li",
            "Jiwen Lu",
            "Shiji Song",
            "Gao Huang"
          ],
          "decision": null,
          "pdate": 1727287702885,
          "cdate": 1715173511687
        }
      ]
    }
  },
  {
    "paper_id": 22253,
    "title": "ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors",
    "abstract": "3D point clouds directly collected from objects through sensors are often incomplete due to self-occlusion. Conventional methods for completing these partial point clouds rely on manually organized training sets and are usually limited to object categories seen during training.\nIn this work, we propose a test-time framework for completing partial point clouds across unseen categories without any requirement for training.\nLeveraging point rendering via Gaussian Splatting, we develop techniques of Partial Gaussian Initialization, Zero-shot Fractal Completion, and Point Cloud Extraction that utilize priors from pre-trained 2D diffusion models to infer missing regions and extract uniform completed point clouds.\nExperimental results on both synthetic and real-world scanned point clouds demonstrate that our approach outperforms existing methods in completing a variety of objects. Our project page is athttps://tianxinhuang.github.io/projects/ComPC/.",
    "introduction": "3D point clouds have always been an important perceptual approach for the physical world, finding extensive use in various applications such as SLAM(Cadena et al.,2016)or 3D detection(Geiger et al.,2013; Reddy et al.,2018). However, point clouds are often captured from specific camera viewpoints(Yuan et al.,2018; Kasten et al.,2024)in real applications, which may lead to the incompleteness of collected points due to the self-occlusion.\nRobust completion for partial point clouds can greatly reduce the cost for data collection, and are useful for subsequent 3D perception.\n\n\nAs illustrated in Fig.1-(a), most existing completion methods(Yuan et al.,2018; Zhao et al.,2021; Zhou et al.,2022; Yu et al.,2023)adopt well-designed deep neural networks to directly generate complete point clouds from partial ones. These methods are usually trained on specific point cloud datasets(Yuan et al.,2018; Yu et al.,2023)and demonstrate outstanding performances on their respective test sets.\nHowever, they face challenges in handling data that differs from what they were trained on, such as unseen object categories or real-world scans. This limitation significantly hinders the practical deployment of these point cloud completion methods.\n\n\nLeveraging the impressive capabilities of 2D diffusion models(Rombach et al.,2022; Saharia et al.,2022; Ho et al.,2020), SDS-complete(Kasten et al.,2024)firstly propose a test-time point cloud completion methods utilizing text-to-3D generative models(Poole et al.,2022; Wang et al.,2023).\nAs shown in Fig.1-(b), this method optimizes aNeuralsurface(Yariv et al.,2021)guided by Score Distillation Sampling (SDS)(Poole et al.,2022)of the text-conditioned Stable Diffusion(Rombach et al.,2022).\nThe Neural surface, modeled as a Signed Distance Field (SDF) followingVolSDFYariv et al. (2021), incorporates the geometric details from the partial points by setting their SDF values to zero. The completed points are then generated from the optimized surface for assessment.\nBy tapping into the extensive 2D knowledge provided by diffusion models, SDS-complete(Kasten et al.,2024)manages to achieve significantly robust point cloud completion without any training on specific training sets.\nHowever, a notable limitation of the method proposed by SDS-complete(Kasten et al.,2024)is its dependency on manually created text prompts for each point cloud to guide the completion. This requirement can encounter a challenge in real-world applications, where providing detailed and accurate text descriptions for incomplete point clouds is not always feasible.\n\n\nIn view of the above-mentioned issues, we propose a novel test-time point cloud completion framework that eliminates the need for any extra manually provided information such as text descriptions.As discussed in PCN(Yuan et al.,2018)and SDS-complete(Kasten et al.,2024), existing completion methods concentrate mainly on point clouds incomplete due to self-occlusion, which means that these point clouds often appear nearly complete from at least one viewpoint.\nInspired by the amodal perception(Lehar,1999; Breckon & Fisher,2005), we aim to complete a point cloud by utilizing the observation from a reference viewpoint that provides the most complete view of the point cloud.\n\n\nAs illustrated in Fig.1-(c), we estimate such a viewpoint and acquire a reference image of the partial point cloud.Inspired by the capability of novel view synthetic diffusion model, e.g., Zero 1-to-3(Liu et al.,2023), we propose to use the reference image as a condition for guidance from the diffusion model to infer the missing regions.\nUtilizing 3D Gaussian Splatting (GS)(Kerbl et al.,2023), which can render 2D images from discrete 3D Gaussians initialized from point clouds, we can effectively render the reference image. This approach also allows us to incorporate 2D diffusion priors into the process of modifying 3D geometry. Consequently, we can complete the missing regions by optimizing the 3D Gaussians with guidance from the 2D diffusion model.Moreover, we propose Preservation Constraint to maintain the geometric integrity of partial point clouds. The completed point clouds would be finally acquired from the 3D Gaussian centers.\n\n\nOurmain contributionscan be summarized as below:•We propose the Partial Gaussian Initialization to generate a reference image for partial points, which is observed from an estimated reference viewpoint;•Based on the reference image, we develop the Zero-shot Fractal Completion to complete the missing regions by introducing 2D diffusion priors;•We propose Point Cloud Extraction to extract uniform point clouds from 3D Gaussians;•Through comprehensive evaluation across various data, we demonstrate that our approach surpasses conventional completion methods in handling both synthetic and real-world scanned point clouds.\n\n\nFigure 1:Different point cloud completion methods. (a) Existing network-based completion methods; (b) Test-time SDS-complete(Kasten et al.,2024)with text prompts to guide Neural surface for completion; (c) Our method based on 3D Gaussian Splatting (GS) guided by the diffusion model from Zero 1-to-3(Liu et al.,2023)conditioned on the reference image rendered from partial points.\n\n",
    "related_work": "2.13D Generation via 2D priorsSince the notable success of 2D diffusion models in text-to-image generation(Rombach et al.,2022; Saharia et al.,2022; Ho et al.,2020), text-to-3D and image-to-3D generation have attracted the attention of an increasing number of researchers.\nTo achieve robust and generalizable 3D generation, researchers propose to lift 2D priors for 3D generation(Poole et al.,2022; Wang et al.,2023; Mohammad Khalid et al.,2022; Michel et al.,2022).\nThese works usually optimize specific 3D representations by guidance from 2D diffusion models under different viewpoints, where the guidance is calculated with Score Distillation Sampling (SDS)(Poole et al.,2022)through rendered images.Score Distillation Sampling (SDS) guides a target model (e.g., NeRF) by using gradients from a pre-trained diffusion model. This aligns the target model’s output with the diffusion model’s learned distribution, enabling high-quality generation in specialized domains.Zero 1-to-3(Liu et al.,2023)achieve remarkable 3D generation quality by using SDS guidance from their pre-trained novel view synthesis diffusion model explicitly conditioned on the reference image and camera transformation.Conditioned on a single image, Zero 1-to-3 predicts an image consistent with plausible 3D shapes for any given camera pose.However, its reliance on NeRF representation leads to prolonged optimization times.3D Gaussian Splatting (GS)(Kerbl et al.,2023)is an efficient 3D representation that encodes both geometrical and appearance information using a set of 3D Gaussians. Each Gaussian is defined by attributes such as 3D coordinates, scaling, opacity, rotation, and spherical harmonics parameters. By optimizing these attributes, information from 2D images can be incorporated into the Gaussians, enabling efficient novel-view rendering.\nDreamgaussian(Tang et al.,2023)offers a solution by optimizing 3D Gaussians through SDS from Zero 1-to-3, achieving a balance between high-quality outputs and acceptable optimization durations.Motivated by Dreamgaussian, we recognize the potential of GS to refine 3D coordinates of Gaussian centers using guidance from 2D diffusion models.\nThis insight presents an opportunity to apply 2D diffusion priors to tasks related to 3D point clouds, such as point cloud completion.\n\n\n2.2Point Cloud CompletionPoint cloud completion aims to recover completed point clouds from partial input point clouds.\nEver since PCN(Yuan et al.,2018)firstly applied deep neural networks to predict complete point clouds from partial inputs, numerous advancements(Zhang et al.,2020; Xie et al.,2020; Huang et al.,2020; Yu et al.,2021; Wang et al.,2020; Xiang et al.,2022; Wen et al.,2021)have been made to enhance the accuracy of point cloud completion by altering network architectures. For example, GRNet(Xie et al.,2020)converts point clouds into grid formats and employs 3D CNNs for predicting the completed structures, while PFNet(Huang et al.,2020)adopts a fractal approach to better preserve existing shape details.The Fractal approach focuses on predicting only the missing regions of point clouds, preserving existing details by retaining the shapes from the partial input.RFNet(Huang et al.,2021)utilizes a differentiable layer to merge existing geometrical details from partial point clouds into completed results.More recent approaches(Wang et al.,2024; Zhu et al.,2023; Li et al.,2023; Yu et al.,2021; Xiang et al.,2022; Zhou et al.,2022; Yu et al.,2023; Yan et al.,2022)integrate carefully-designed transformers to improve completion accuracy by considering broader geometric relationships.DiffCompleteChu et al. (2023)is a diffusion-based model for 3D shape completion, leveraging probabilistic modeling to predict missing parts of 3D shapes while preserving structural coherence and diversity.However, the effectiveness of these point cloud completion methods diminishes when applied to data that differ from their training sets, such as point clouds from unseen categories or other datasets.\nSDS-complete(Kasten et al.,2024)proposed a test-time completion framework that employsVolSDF(Yariv et al.,2021)for rendering, drawing on priors from pre-trained text-to-image 2D diffusion models(Rombach et al.,2022). This approach maintains the original shapes by constraining the Signed Distance Field (SDF) values of the partial inputs.\nYet, this strategy’s reliance on text-to-image diffusion models for guidance necessitates well-defined text prompts for each partial point cloud, which may not be practical in real-world applications.\nMoreover, the optimization of SDS-Complete is quite time-consuming, which may take more than 1000 minutes for one point cloud.In this study, we propose to leverage 3D Gaussian Splatting (GS)(Kerbl et al.,2023)to bridge point clouds with priors from 2D diffusion models.\nBy generating a reference image of the partial point cloud to serve as a condition for guidance from Zero 1-to-3(Liu et al.,2023), our method can extract uniform and completed point clouds from the 3D Gaussian centers. Since our method exclusively utilizes information gathered from the incomplete point cloud for completion, it eliminates the need for any additional manually specified prompts for each point cloud.\nDue to the efficient rendering from 3D GS, and stronger priors from Zero 1-to-3, our method can achieve much higher optimization efficiency than SDS-Complete(Kasten et al.,2024).\n\n",
    "origin_citations": [
      "High-resolution image synthesis with latent diffusion models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Denoising diffusion probabilistic models.",
      "Dreamfusion: Text-to-3d using 2d diffusion.",
      "Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation.",
      "Clip-mesh: Generating textured meshes from text using pretrained image-text models.",
      "Text2mesh: Text-driven neural stylization for meshes.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "Dreamgaussian: Generative gaussian splatting for efficient 3d content creation.",
      "Pcn: Point completion network.",
      "Detail preserved point cloud completion via separated feature aggregation.",
      "Grnet: Gridding residual network for dense point cloud completion.",
      "Pf-net: Point fractal network for 3d point cloud completion.",
      "Pointr: Diverse point cloud completion with geometry-aware transformers.",
      "Cascaded refinement network for point cloud completion.",
      "Snowflake point deconvolution for point cloud completion and generation with skip-transformer.",
      "Pmp-net: Point cloud completion by learning multi-step point moving paths.",
      "Rfnet: Recurrent forward network for dense point cloud completion.",
      "Pointattn: You only need attention for point cloud completion.",
      "Svdformer: Complementing point cloud via self-view augmentation and self-structure dual-generator.",
      "Proxyformer: Proxy alignment assisted point cloud completion with missing part sensitive transformer.",
      "Seedformer: Patch seeds based point cloud completion with upsample transformer.",
      "Adapointr: Diverse point cloud completion with adaptive geometry-aware transformers.",
      "Shapeformer: Transformer-based shape completion via sparse representation.",
      "Diffcomplete: Diffusion-based generative 3d shape completion.",
      "Point cloud completion with pretrained text-to-image diffusion models.",
      "Volume rendering of neural implicit surfaces."
    ],
    "citations": [
      2240,
      5926,
      4827,
      176110,
      223722,
      500898,
      152144,
      495077,
      180969,
      477674,
      5609,
      201880,
      5610,
      5611,
      454163,
      5607,
      174879,
      110149,
      503077,
      470708,
      909345,
      421160,
      302498,
      56637,
      47482,
      50099,
      393442
    ],
    "extract_rate": 0.9642857142857143,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 85,
      "Final_Score": 84.9,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 90\nSynthesis_Analysis: 85\nFinal_Score: 85"
    },
    "overall_score": 85.54,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "SoUwcVplq4",
        "forum": "SoUwcVplq4",
        "number": 2610,
        "title": "ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors",
        "normalized_title": "compc completing a 3d point cloud with 2d diffusion priors",
        "abstract": "3D point clouds directly collected from objects through sensors are often incomplete due to self-occlusion. Conventional methods for completing these partial point clouds rely on manually organized training sets and are usually limited to object categories seen during training. In this work, we propose a test-time framework for completing partial point clouds across unseen categories without any requirement for training. Leveraging point rendering via Gaussian Splatting, we develop techniques of Partial Gaussian Initialization, Zero-shot Fractal Completion, and Point Cloud Extraction that utilize priors from pre-trained 2D diffusion models to infer missing regions and extract uniform completed point clouds. Experimental results on both synthetic and real-world scanned point clouds demonstrate that our approach outperforms existing methods in completing a variety of objects. Our project page is at \\url{https://tianxinhuang.github.io/projects/ComPC/}.",
        "authors": [
          "Tianxin Huang",
          "Zhiwen Yan",
          "Yuyang Zhao",
          "Gim Hee Lee"
        ],
        "decision": null,
        "pdate": 1737562284872,
        "cdate": 1727011640635
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "SoUwcVplq4",
          "forum": "SoUwcVplq4",
          "number": 2610,
          "title": "ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors",
          "normalized_title": "compc completing a 3d point cloud with 2d diffusion priors",
          "abstract": "3D point clouds directly collected from objects through sensors are often incomplete due to self-occlusion. Conventional methods for completing these partial point clouds rely on manually organized training sets and are usually limited to object categories seen during training. In this work, we propose a test-time framework for completing partial point clouds across unseen categories without any requirement for training. Leveraging point rendering via Gaussian Splatting, we develop techniques of Partial Gaussian Initialization, Zero-shot Fractal Completion, and Point Cloud Extraction that utilize priors from pre-trained 2D diffusion models to infer missing regions and extract uniform completed point clouds. Experimental results on both synthetic and real-world scanned point clouds demonstrate that our approach outperforms existing methods in completing a variety of objects. Our project page is at \\url{https://tianxinhuang.github.io/projects/ComPC/}.",
          "authors": [
            "Tianxin Huang",
            "Zhiwen Yan",
            "Yuyang Zhao",
            "Gim Hee Lee"
          ],
          "decision": null,
          "pdate": 1737562284872,
          "cdate": 1727011640635
        }
      ]
    }
  },
  {
    "paper_id": 27637,
    "title": "BERTs are Generative In-Context Learners",
    "abstract": "While in-context learning is commonly associated with causal language models, such as GPT, we demonstrate that this capability also ‘emerges’ in masked language models. Through anembarrassingly simpleinference technique, we enable an existing masked model, DeBERTa, to perform generative tasks without additional training or architectural changes. Our evaluation reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. These complementary strengths suggest that the field’s focus on causal models for in-context learning may be limiting – both architectures can develop these capabilities, but with distinct advantages; pointing toward promising hybrid approaches that combine the strengths of both objectives.",
    "introduction": "Masked language models used to dominate the field of natural language processing due to their adaptability across diverse tasks and their superior performance compared to causal language models(Radford et al.,2018;Devlin et al.,2019). Between 2018 and 2020, the field witnessed a surge in the development of these models(Devlin et al.,2019;Liu et al.,2019;Lan et al.,2020, inter alia). However, the field dramatically shifted with GPT-3 and its introduction ofin-context learning– the ability to infer and perform tasks from prompts and examples without any finetuning(Brown et al.,2020). This capability eliminated the need for task-specific training data and deep-learning expertise, making such models far more practical for real-world applications. This perceived advantage led many researchers and practitioners to abandon masked language models in favor of GPT-style architectures.\n\n\nFigure 1:The average 1-shot performance across four groups of NLP tasksWe compare the scaling abilities of DeBERTa (four sizes in red) with GPT-3 (eight sizes in blue). Even though these models rely on different training objectives, they scale in a similar log-linear manner overall. Yet, on a task-by-task basis, the pretraining methods lead to substantial differences between them.\n\n\nPrevious studies of ‘emergent’ in-context learning abilities have focused almost exclusively on causal language models, creating a widespread assumption that this capability is unique to them(Saunshi et al.,2021;Olsson et al.,2022;Wei et al.,2022;Wang et al.,2023, inter alia). In this paper, we challenge this assumption by demonstrating that in-context learning can emerge in masked language models as well. In-context learning is a more general phenomenon and should not be studied with a singular pretraining objective in mind. Moreover, the assumed inability of masked language models to perform (generative) in-context learning has rendered them outdated – as explicitly noted byTay et al.(2023):“BERT-style models are very restricted in their generative capabilities. Because of the cumbersomeness of task-specific classification heads, we strongly do not recommend using this class of autoencoding models moving forward and consider them somewhat deprecated.”\n\n\nIn this paper, we challenge these prevailing assumptions about masked language models (MLMs). We present empirical evidence showing that DeBERTa, an MLM released just one month after GPT-3, is equally adept at in-context learning. Our findings suggest that the capacity for in-context learning is not tied to the training objective, but can be achieved across different types of language models. To our surprise, we found that DeBERTa does not simply mimic the performance of GPT-3 – the two model behave very differently – DeBERTa is clearly much better on tasks such as language understanding, and, on the other hand, much worse on tasks such as closed-book question answering. This suggests that masked and causal language modeling are two complementary training objectives and that there is a great potential for a training method that combines the strengths of both objectives. Finally,scaling(performance improvement with increased size of pretrained language models) is a crucial feature of modern language models; we demonstrate that MLMs doscaleon in-context learning (Figure1).\n\n\nWe introduce a simple inference technique that transforms an MLM into a generative model without any further training. Using publicly available DeBERTa checkpoints, we show that the MLM training objective not only provides a versatile way of encoding text, but is also competitive in text generation and text completion ranking. This claim is tested by following the same evaluation suite as GPT-3, speculating on an ‘alternative reality’ in which a masked language model is the first model reported to achieve the so-called ‘emergent’ in-context learning abilities. While other masked language models could potentially demonstrate similar capabilities, we deliberately target DeBERTa because of its large size and its length-generalization abilities. Ultimately, our goal is to demonstrate that MLMscanperform in-context learning and that theycanbe surprisingly good at doing so.\n\n\nOutlineFirst,Section2(Method) describes the inference methods used to evaluate the in-context learning abilities of an off-the-shelf masked language model. ThenSection3(DeBERTa) describes the details of the particular model used in this study.Section4(Evaluation) details the evaluation setup and compares DeBERTa with GPT-3. Finally,Section5(Related work) talks about other relevant work within this topic, and the paper concludes withSection6(Conclusion).Figure 2:Illustration of the proposed methods for using a masked language model for text generation and text rankingWe show how to adapt a masked language model for in-context-learning tasks through simple input reformatting, requiring no additional training.Left: Text generation is achieved by 1) appending[MASK]tokens to the input prompt, 2) predicting the next token for the first mask, and 3) iteratively appending new masks and predicting tokens.Right: A similar approach is used to retrieve a pseudo-log-likelihood score of a text sequence that can be used to rank multiple sequences by their individual likelihoods. Both methods maintain the model’s original architecture while enabling new capabilities through careful input formatting.\n\n",
    "related_work": "Few-shot finetuning with masked language modelsWhile our work demonstrates the emergence of in-context learning in masked language models, prior research has explored different approaches to few-shot learning with these architectures. The dominant paradigm has been few-shotfinetuning, where the model’s weights are updated using a small number of examples. Studies bySchick and Schütze(2021),Gao et al.(2021), andXia et al.(2022)showed promising results with this approach. However, these methods require additional training steps with a complicated training objective, making them more complex to implement compared to the simple prompting-based in-context learning demonstrated in our work. Despite the generally lower performance of in-context learning compared to few-shot finetuning(Liu et al.,2022), its simplicity and immediacy have made it the preferred choice in many practical applications.\n\n\nOther large masked language modelsOur choice of DeBERTa for this study was motivated by its unique combination of size and capability to handle extended context lengths. While larger masked language models exist, such as Megatron BERT with 3.9 billion parameters(unfortunately not publicly available;Shoeybi et al.,2019)and XLM-RoBERTa with 10.7 billion parameters(Goyal et al.,2021), they have limitations that make them less suitable for studying in-context learning. Megatron BERT lacks mechanisms for length generalization, which is crucial for processing long prompts with multiple examples, while XLM-RoBERTa’s multilingual nature and restricted sequence length of 512 tokens would confound our analysis. DeBERTa’s architecture, particularly its relative positional embeddings, makes it an ideal candidate for exploring how masked language models scale with in-context learning.\n\n\nHybrid masked-causal modelsOur empirical findings, particularly the complementary strengths of masked and causal models demonstrated inSection4, suggest significant potential in combining these approaches. Several architectures have already explored this direction, even if inadvertently: T5(Raffel et al.,2020), BART(Lewis et al.,2020)and GLM(Du et al.,2022)introduced autoregressive fill-in-the-blank objectives; CM3 developed a causal-mask approach(Aghajanyan et al.,2022); and PrefixLM implemented a partially bidirectional causal model(Dong et al.,2019;Raffel et al.,2020). These efforts align with our observations about the distinct advantages of masked and causal objectives. The recent work byDing et al.(2024)provides theoretical support for this direction, demonstrating that prefix language models, which combine aspects of both architectures, are particularly well-suited for in-context learning.\n\n",
    "origin_citations": [
      "It’s not just size that matters: Small language models are also few-shot learners.",
      "Making pre-trained language models better few-shot learners.",
      "Prompting ELECTRA: Few-shot learning with discriminative pre-trained models.",
      "Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning.",
      "Megatron-LM: Training multi-billion parameter language models using model parallelism.",
      "Larger-scale transformers for multilingual masked language modeling.",
      "Exploring the limits of transfer learning with a unified text-to-text transformer.",
      "BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.",
      "GLM: General language model pretraining with autoregressive blank infilling.",
      "CM3: A causal masked multimodal model of the internet.",
      "Unified language model pre-training for natural language understanding and generation.",
      "CausalLM is not optimal for in-context learning."
    ],
    "citations": [
      8745,
      8744,
      145345,
      377481,
      1469,
      251552,
      9,
      69558,
      2212,
      57784,
      50,
      179637
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 83,
      "Final_Score": 85.8,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 83\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 82,
      "Final_Score": 85.3,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 88\nSynthesis_Analysis: 82\nFinal_Score: 85"
    },
    "overall_score": 85.5,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "BCA9NMZkLS",
        "forum": "BCA9NMZkLS",
        "number": 20723,
        "title": "BERTs are Generative In-Context Learners",
        "normalized_title": "berts are generative in context learners",
        "abstract": "While in-context learning is commonly associated with causal language models, such as GPT, we demonstrate that this capability also 'emerges' in masked language models. Through an embarrassingly simple inference technique, we enable an existing masked model, DeBERTa, to perform generative tasks without additional training or architectural changes. Our evaluation reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. These complementary strengths suggest that the field's focus on causal models for in-context learning may be limiting – both architectures can develop these capabilities, but with distinct advantages; pointing toward promising hybrid approaches that combine the strengths of both objectives.",
        "authors": [
          "David Samuel"
        ],
        "decision": null,
        "pdate": 1727288240234,
        "cdate": 1715798029955
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "BCA9NMZkLS",
          "forum": "BCA9NMZkLS",
          "number": 20723,
          "title": "BERTs are Generative In-Context Learners",
          "normalized_title": "berts are generative in context learners",
          "abstract": "While in-context learning is commonly associated with causal language models, such as GPT, we demonstrate that this capability also 'emerges' in masked language models. Through an embarrassingly simple inference technique, we enable an existing masked model, DeBERTa, to perform generative tasks without additional training or architectural changes. Our evaluation reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. These complementary strengths suggest that the field's focus on causal models for in-context learning may be limiting – both architectures can develop these capabilities, but with distinct advantages; pointing toward promising hybrid approaches that combine the strengths of both objectives.",
          "authors": [
            "David Samuel"
          ],
          "decision": null,
          "pdate": 1727288240234,
          "cdate": 1715798029955
        }
      ]
    }
  },
  {
    "paper_id": 34264,
    "title": "FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention",
    "abstract": "Video diffusion models have made substantial progress in various video generation applications. However, training models for long video generation tasks require significant computational and data resources, posing a challenge to developing long video diffusion models.\nThis paper investigates a straightforward and training-free approach to extend an existing short video diffusion model (e.g., pre-trained on 16-frame videos) for consistent long video generation (e.g., 128 frames).\nOur preliminary observation has found that directly applying the short video diffusion model to generate long videos can lead to severe video quality degradation. Further investigation reveals that this degradation is primarily due to the distortion of high-frequency components in long videos, characterized by a decrease in spatial high-frequency components and an increase in temporal high-frequency components. Motivated by this, we propose a novel solution named FreeLong to balance the frequency distribution of long video features during the denoising process.\nFreeLong blends the low-frequency components of global video features, which encapsulate the entire video sequence, with the high-frequency components of local video features that focus on shorter subsequences of frames. This approach maintains global consistency while incorporating diverse and high-quality spatiotemporal details from local videos, enhancing both the consistency and fidelity of long video generation.\nWe evaluated FreeLong on multiple base video diffusion models and observed significant improvements. Additionally, our method supports coherent multi-prompt generation, ensuring both visual coherence and seamless transitions between scenes. Project page:https://yulu.net.cn/freelong",
    "introduction": "Figure 1:Results of Short and Long Videos.The first row of each case shows 16-frame videos generated using short video diffusion models (LaVie[1]and VideoCrafter2[2]). Directly extending these models to longer videos, like those with 128 frames, preserves temporal consistency but lacks fine spatial-temporal details. In contrast, our proposed FreeLong adapts short video diffusion models to create consistent long videos with high fidelity.\n\n\nVideo diffusion models[1,2,3,4,5,6,7]trained on vast video-text datasets[8,9]have demonstrated impressive capabilities in generating high-quality videos. Inspired by Sora[10], multiple studies[11,12,13]have concentrated on training these models to create longer videos using extensive, long video-text datasets[14,15,16,17,18]. However, these methods demand significant computational resources and data annotations.\n\n\nA more practical approach involves adapting pre-trained short video models to generate consistent longer video sequences without retraining. Recent research[19,20]has explored sliding window temporal attention to ensure smooth transitions between video clips in the generation of long videos. Nonetheless, these techniques often struggle to maintain global temporal consistency across extended sequences and require multiple passes of temporal attention.\n\n\nIn this study, we propose a simple, training-free method to adapt existing short video diffusion models (e.g., pretrained on 16 frames) for generating consistent long videos (e.g., 128 frames). Initially, we examine the direct application of short video diffusion models for long video generation. As depicted in Figure1, straightforwardly using a 16-frame video diffusion model to produce 128-frame sequences yields globally consistent yet low-quality results.\n\n\nTo delve further into these issues, we conducted a frequency analysis of the generated long videos. As shown in Figure2(a), the low-frequency components remain stable as the video length increases, while the high-frequency components exhibit a noticeable decline, leading to a drop in video quality. The findings indicate that although the overall temporal structure is preserved, fine-grained details suffer notably in longer sequences. Specifically, there is a decrease in high-frequency spatial components (Figure2(b)) and an increase in high-frequency temporal components (Figure2(c)). This high-frequency distortion poses a challenge in maintaining high fidelity over extended sequences. As illustrated in the middle row of each case in Figure1, intricate textures like forest paths or sunrises become blurred and less defined, while temporal flickering and sudden changes disrupt the video’s narrative flow.\n\n\nTo tackle these challenges, we introduce FreeLong, a novel framework that employs SpectralBlend Temporal Attention (SpectralBlend-TA) to balance the frequency distribution of long video features in the denoising process. SpectralBlend-TA integrates global and local features via two parallel streams, enhancing the fidelity and consistency of long video generation. The global stream deals with the entire video sequence, capturing extensive dependencies and themes for narrative continuity. Meanwhile, the local stream focuses on shorter frame subsequences to retain fine details and smooth transitions, preserving high-frequency spatial and temporal information. SpectralBlend-TA combines global and local video features in the frequency domain, improving both consistency and fidelity by blending low-frequency global components with high-frequency local components. Our method is entirely training-free and allows for the easy integration of FreeLong into existing video diffusion models by adjusting the original temporal attention of video diffusion models. Comparative experiments demonstrate significant improvements in temporal consistency and video fidelity when applying our method to generate long video sequences.\n\n\nOur contributions can be summarized as follows:1)We conduct a frequency analysis on the direct application of short video models for longer video generation and identify high-frequency distortions in the longer videos.2)We devise a SpectralBlend Temporal Attention mechanism to merge the consistent low-frequency components of global videos with the high-fidelity high-frequency components of local videos.3)Our training-free approach, FreeLong, outperforms existing state-of-the-art models in both temporal consistency and video fidelity.\n\n",
    "related_work": "Text-to-Video Diffusion Models:Text-to-video (T2V) generation has progressed significantly from early variational autoencoders[21,22]and GANs[23]to advanced diffusion-based techniques[3,4,24,25,26], marking a major leap in synthesis methods. Modern video diffusion models build on pre-trained image-to-text diffusion models[27,28,29], incorporating temporal transformers in the diffusion UNet to capture temporal relationships. These models achieve impressive video generation results through post-training on video-text data[14,9,8,1], enhancing coherence and fidelity. However, due to computational constraints and limited dataset availability, current video diffusion models are typically trained on fixed-length short videos (e.g., 16 frames), limiting their ability to produce longer videos.\nIn this paper, we propose extending these short video diffusion models to generate long and consistent videos without requiring any additional training videos.\n\n\nLong-video Generation:Generating long videos is challenging due to temporal complexity, resource constraints, and the need for content consistency. Recent advancements focus on improving temporal coherence and visual quality using GAN-based[30,31]and diffusion-based techniques[32,33,34,35,36]. For instance, Nuwa-XL[36]employs a parallel diffusion process, while StreamingT2V[11]uses an autoregressive approach with a short-long memory block to improve the consistency of long video sequences. Despite their effectiveness, these methods require substantial computational resources and large-scale datasets.\nRecent research has explored training-free adaptations using short video diffusion models for long video generation. Gen-L-Video[37]extends videos by merging overlapping sub-segments with a sliding-window method during denoising. FreeNoise[19]employs sliding-window temporal attention and a noise initialization strategy to maintain temporal consistency. However, these approaches focus on smooth transitions between video clips and fail to capture global consistency across long video sequences. This paper proposes FreeLong, a novel approach that blends global and local video features during the denoising process to enhance both global temporal consistency and visual quality in long video generation.\n\n",
    "origin_citations": [
      "Video generation from text.",
      "Storygan: A sequential conditional GAN for story visualization.",
      "Generative adversarial networks.",
      "Cogvideo: Large-scale pretraining for text-to-video generation via transformers.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Magicvideo: Efficient video generation with latent diffusion models.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Freeinit: Bridging initialization gap in video diffusion models.",
      "High-resolution image synthesis with latent diffusion models.",
      "Gligen: Open-set grounded text-to-image generation.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Panda-70m: Captioning 70m videos with multiple cross-modality teachers.",
      "Frozen in time: A joint video and image encoder for end-to-end retrieval.",
      "Internvid: A large-scale video-text dataset for multimodal understanding and generation.",
      "Lavie: High-quality video generation with cascaded latent diffusion models.",
      "Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2.",
      "Generating long videos of dynamic scenes.",
      "Flexible diffusion modeling of long videos.",
      "Mcvd-masked conditional video diffusion for prediction, generation, and interpolation.",
      "Consistent view synthesis with pose-guided diffusion models.",
      "Denoising diffusion probabilistic models.",
      "Nuwa-xl: Diffusion over diffusion for extremely long video generation.",
      "Streamingt2v: Consistent, dynamic, and extendable long video generation from text.",
      "Gen-l-video: Multi-text to long video generation via temporal co-denoising.",
      "Freenoise: Tuning-free longer video diffusion via noise rescheduling."
    ],
    "citations": [
      505804,
      359655,
      7217,
      145593,
      185878,
      244720,
      176197,
      271105,
      2240,
      214623,
      5926,
      1024928,
      30138,
      184971,
      478403,
      535582,
      5905,
      147199,
      148022,
      492166,
      4827,
      494613,
      425504,
      256996,
      313502
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 83,
      "Final_Score": 85.8,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 83\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 85,
      "Final_Score": 85.3,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 85\nFinal_Score: 85"
    },
    "overall_score": 85.5,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "X9Fga52OOv",
        "forum": "X9Fga52OOv",
        "number": 347,
        "title": "FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention",
        "normalized_title": "freelong training free long video generation with spectralblend temporal attention",
        "abstract": "Video diffusion models have made substantial progress in various video generation applications. However, training models for long video generation tasks require significant computational and data resources, posing a challenge to developing long video diffusion models.\nThis paper investigates a straightforward and training-free approach to extend an existing short video diffusion model (e.g. pre-trained on 16-frame videos) for consistent long video generation (e.g. 128 frames).  Our preliminary observation has found that directly applying the short video diffusion model to generate long videos can lead to severe video quality degradation. Further investigation reveals that this degradation is primarily due to the distortion of high-frequency components in long videos, characterized by a decrease in spatial high-frequency components and an increase in temporal high-frequency components. Motivated by this, we propose a novel solution named FreeLong to balance the frequency distribution of long video features during the denoising process. FreeLong blends the low-frequency components of global video features, which encapsulate the entire video sequence, with the high-frequency components of local video features that focus on shorter subsequences of frames. This approach maintains global consistency while incorporating diverse and high-quality spatiotemporal details from local videos, enhancing both the consistency and fidelity of long video generation. We evaluated FreeLong on multiple base video diffusion models and observed significant improvements. Additionally, our method supports coherent multi-prompt generation, ensuring both visual coherence and seamless transitions between scenes. Our project page is at: https://yulu.net.cn/freelong.",
        "authors": [
          "Yu Lu",
          "Yuanzhi Liang",
          "Linchao Zhu",
          "Yi Yang"
        ],
        "decision": null,
        "pdate": 1727287635541,
        "cdate": 1713872204379
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "X9Fga52OOv",
          "forum": "X9Fga52OOv",
          "number": 347,
          "title": "FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention",
          "normalized_title": "freelong training free long video generation with spectralblend temporal attention",
          "abstract": "Video diffusion models have made substantial progress in various video generation applications. However, training models for long video generation tasks require significant computational and data resources, posing a challenge to developing long video diffusion models.\nThis paper investigates a straightforward and training-free approach to extend an existing short video diffusion model (e.g. pre-trained on 16-frame videos) for consistent long video generation (e.g. 128 frames).  Our preliminary observation has found that directly applying the short video diffusion model to generate long videos can lead to severe video quality degradation. Further investigation reveals that this degradation is primarily due to the distortion of high-frequency components in long videos, characterized by a decrease in spatial high-frequency components and an increase in temporal high-frequency components. Motivated by this, we propose a novel solution named FreeLong to balance the frequency distribution of long video features during the denoising process. FreeLong blends the low-frequency components of global video features, which encapsulate the entire video sequence, with the high-frequency components of local video features that focus on shorter subsequences of frames. This approach maintains global consistency while incorporating diverse and high-quality spatiotemporal details from local videos, enhancing both the consistency and fidelity of long video generation. We evaluated FreeLong on multiple base video diffusion models and observed significant improvements. Additionally, our method supports coherent multi-prompt generation, ensuring both visual coherence and seamless transitions between scenes. Our project page is at: https://yulu.net.cn/freelong.",
          "authors": [
            "Yu Lu",
            "Yuanzhi Liang",
            "Linchao Zhu",
            "Yi Yang"
          ],
          "decision": null,
          "pdate": 1727287635541,
          "cdate": 1713872204379
        }
      ]
    }
  },
  {
    "paper_id": 52784,
    "title": "Object-Centric Pretraining via Target Encoder Bootstrapping",
    "abstract": "Object-centric representation learning has recently been successfully applied to real-world datasets. This success can be attributed to pretrained non-object-centric foundation models, whose features serve as reconstruction targets for slot attention. However, targets must remain frozen throughout the training, which sets an upper bound on the performance object-centric models can attain. Attempts to update the target encoder by bootstrapping result in large performance drops, which can be attributed to its lack of object-centric inductive biases, causing the object-centric model’s encoder to drift away from representations useful as reconstruction targets.\nTo address these limitations, we proposeObject-CEntric Pretraining by Target EncoderBOotstrapping, a self-distillation setup for training object-centric models from scratch, on real-world data, for the first time ever. In OCEBO, the target encoder is updated as an exponential moving average of the object-centric model, thus explicitly being enriched with object-centric inductive biases introduced by slot attention while removing the upper bound on performance present in other models. We mitigate the slot collapse caused by random initialization of the target encoder by introducing a novel cross-view patch filtering approach that limits the supervision to sufficiently informative patches. When pretrained on 241k images from COCO, OCEBO achieves unsupervised object discovery performance comparable to that of object-centric models with frozen non-object-centric target encoders pretrained on hundreds of millions of images. The code and pretrained models are publicly available athttps://github.com/djukicn/ocebo.",
    "introduction": "In recent years, large-scale foundation models have become ubiquitous across many domains of deep learning. This is mainly thanks to the development of self-supervised learning (SSL) techniques that enable pretraining on massive amounts of unlabeled data. In computer vision, several families of SSL techniques have emerged, focusing on (i) imposing cross-view invariance at the global(Grill et al.,2020; Caron et al.,2021), patch(Zhou et al.,2022; Stegmüller et al.,2023)or cluster(Wen et al.,2022; Lebailly et al.,2024)level, or (ii) reconstruction in the pixel(He et al.,2021)or latent space(Zhou et al.,2022; Assran et al.,2023). On the other hand, research in cognitive psychology(Benjamin Peters,2021)suggests that human visual perception reasons about visual inputs by decomposing scenes into sets of objects with corresponding representations. In the past years, efforts have been made to mimic this behavior(Locatello et al.,2020; Seitzer et al.,2022; Kakogeorgiou et al.,2024; Didolkar et al.,2024)and move past global and patch representations towards object-centric representations. Object-centric models are not only more biologically plausible(Benjamin Peters,2021), but can provably achieve compositional generalization(Wiedemer et al.,2024)and have already proven useful on several downstream tasks, such as robotic control(Zadaianchuk et al.,2021; Haramati et al.,2024), compositional generation(Singh et al.,2022; Wu et al.,2023; Jiang et al.,2023; Sajjadi et al.,2022), and visual question answering(Xu et al.,2024). While incredibly promising, object-centric models have not yet been successfully pretrained on a large-scale real-world dataset.\n\n\nA slot attention-based object-centric model consists of an image encoder, slot attention encoder and slot decoder. Initially, these were trained end-to-end with a reconstruction objective in the pixel-space(Locatello et al.,2020), resulting in good scene decomposition on synthetic datasets but slot collapse111Slot collapse refers to a scenario where slots learn to attach to positionally coherent blocks rather than actual objects.on real-world data.Seitzer et al. (2022)introduce a reconstruction objective in the latent space of a pretrained encoder (e.g., DINO(Caron et al.,2021), DINOv2(Oquab et al.,2023), etc.), showing that the presence of informative reconstruction targets is crucial for avoiding slot collapse. As this encoder provides reconstruction targets, we refer to it as atarget encoder. Moreover,Seitzer et al. (2022)initialize the encoder of their object-centric model with the same pretrained encoder and keep it frozen alongside the target encoder, effectively training only the slot attention encoder and decoder.\n\n\nWhile the framework with the frozen target encoder achieves unprecedented performance on real-world datasets(Seitzer et al.,2022), it has a significant limitation. Namely, the target encoder cannot be updated, meaning that there exists an upper bound on what the object-centric model can learn from it. In fact, despite the target encoder being pretrained on millions or even hundreds of millions of images(Caron et al.,2021; Oquab et al.,2023), the object-centric model can efficiently consume all its knowledge and reach the upper limit. This is demonstrated in an experiment byDidolkar et al. (2024), where the performance plateaus with∼similar-to\\sim∼16k images from MS COCO(Lin et al.,2015)and does not further increase, even with an order of magnitude more data. This clashes with the scalability trends observed in other (self-supervised) representation learning families. A question arises —how to update the target encoder, thus removing the performance upper limit and enabling large-scale pretraining of object-centric models?\n\n\nKakogeorgiou et al. (2024)show for the first time that unfreezing a few final layers of the object-centric model’s encoder is possible and further improves the performance.Didolkar et al. (2024)further extend this finding by fine-tuning the full encoder. A straightforward approach would be updating the target encoder as an exponential moving average (EMA) of the object-centric model’s encoder. However, this has been shown to result in huge performance drops(Didolkar et al.,2024). We hypothesize that the reason for this is thelack of object-centric inductive biasesin the target encoder. Most often trained for cross-view consistency, SSL models organize their representations in terms of semantics, rather than object instances. For example, semantically similar parts of different objects will be close in the representation space and farther from less similar parts of the same object. When the object-centric model tries to reconstruct such features, its inductive bias focuses on assigning these semantically similar regions to different objects (i.e., instances), thus updating its encoder to focus slightly less on semantics and more on positions in the image. EMA update of the target encoder then results in informative knowledge slowly leaking at the cost of more positional information, which ultimately degrades the performance.\n\n\nIn this work, we argue that a way to overcome the current limitations of object-centric learning and unlock its full potential is by large-scale pretraining from scratch. We pose object-centric learning as a self-distillation bootstrapping problem(Caron et al.,2021), in which the object-centric model is distilled from the target encoder that is in turn updated as an EMA of the object-centric model’s encoder. We propose theobject-centric self-distillation loss, a patch-level loss that acts as a reconstruction objective. We apply it in a cross-view consistency fashion to enrich slot attention with augmentation invariance. To prevent slot collapse stemming from the lack of informative reconstruction targets due to random initialization of the target encoder, we propose across-view patch filteringprocedure that uses cross-view correspondences to determine which patches to reconstruct at a particular training stage. The resulting model, named OCEBO allows pretraining object-centric models on real-world datasets for the first time. The EMA updates not only remove the upper bound from object-centric models but explicitly introduce object-centric inductive biases into the target encoder. By training from scratch, we allow the model to learn informative instance-based features rather than trying to reorganize semantic features from a pretrained non-object-centric encoder. When pretrained on∼similar-to\\sim∼118k or∼similar-to\\sim∼241k images from MS COCO(Lin et al.,2015), not only does OCEBO avoid slot collapse, but also achieves performance comparable to that of object-centric models with a DINOv2(Oquab et al.,2023)target encoder pretrained on 142M images, proving the object-centric inductive biases highly beneficial for the target encoder. Moreover, OCEBO demonstrates scalability well beyond a few thousand training images, unlike other object-centric models. Altogether, we believe this work paves a way towardslarge-scale pretraining of object-centric foundation models.\n\n",
    "related_work": "Self-distillation.Cross-view consistency has emerged as a predominant self-supervised learning (SSL) task in computer vision. In this framework, two views of an input image, obtained through data augmentations, are used to enforce cross-view representation consistency while avoiding representation collapse. While some works avoid collapse via the use of negative samples(Chen et al.,2020; He et al.,2020), self-distillation approaches rely on positive samples only, in combination with mechanisms such as branch asymmetry(Grill et al.,2020), stop-gradients(Chen & He,2021; Caron et al.,2021), momentum encoders(Grill et al.,2020; Caron et al.,2021), etc. Moreover, self-distillation methods can enforce consistency at different levels of granularity. In global or image-level self-supervision(Grill et al.,2020; Caron et al.,2021), consistency is enforced between global representations that capture the whole input image. In patch-level or local self-supervision(Zhou et al.,2022; Stegmüller et al.,2023), consistency is enforced between pairs of corresponding patches in both views. Finally, consistency has been enforced between semantically coherent groups of patches(Wen et al.,2022; Lebailly et al.,2024), i.e., clusters (also referred to as cluster-level or object-level self-supervision).\n\n\nObject-centric representation learning.Image-, patch- and cluster-level self-distillation methods have achieved impressive performance across computer vision. In parallel, attempts at object-centric representation learning have been made. An object-centric model decomposes a scene into a set of objects with corresponding representations, which is in line with theories about underlying mechanisms of human visual perception(Benjamin Peters,2021). Unlike object-level self-distillation methods that rely on soft clustering(Wen et al.,2022; Lebailly et al.,2024), which can often group semantically similar parts of different instances, object-centric models explicitly use inductive biases that enforce instance-level scene decomposition. Slot attention(Locatello et al.,2020)has emerged as a particularly successful inductive bias. It introduces an attention mechanism that promotes competition among slots for different parts of input. In the original work(Locatello et al.,2020), an autoencoding model consisting of an image encoder, slot encoder and slot decoder trained end-to-end has been proposed. It achieved impressive performance on well-structured synthetic datasets but resulted in slot collapse on real-world datasets.Seitzer et al. (2022)replaced the pixel space reconstruction objective with reconstruction of features produced by a frozen pretrained non-object-centric SSL model, enabling the training of object-centric models on real-world data. Subsequent works focused on different decoding strategies(Kakogeorgiou et al.,2024)or fine-tuning the encoder while keeping the target encoder frozen(Didolkar et al.,2024). Diffusion-based objectivesWu et al. (2023); Jiang et al. (2023)have been proposed as an alternative to reconstruction-based objective. However, these still rely on pretrained SSL models. To the best of our knowledge, direct pretraining of object-centric models on real-world datasets has not been achieved yet.\n\n\nNote that SlotCon(Wen et al.,2022)also introduces the concept of slots. However, these serve as learnable queries whose assignment to patch representations of both views is used as a supervision signal. In this context, a larger number of slots corresponding to the number of object types in the training dataset is initialized, such that each slot learns to represent a particular object type. In scenes with multiple instances of a same object type, SlotCon would not produce one representation per instance, but rather a single slot attached to patches of all instances. This is fundamentally different from OCEBO and other object-centric works, where instance-level decomposition is achieved.\n\n",
    "origin_citations": [
      "A simple framework for contrastive learning of visual representations.",
      "Momentum contrast for unsupervised visual representation learning, 2020.",
      "Bootstrap your own latent: A new approach to self-supervised learning.",
      "Exploring simple siamese representation learning, 2021.",
      "Emerging properties in self-supervised vision transformers.",
      "ibot: Image bert pre-training with online tokenizer.",
      "Croc: Cross-view online clustering for dense visual representation learning.",
      "Self-supervised visual representation learning with semantic grouping.",
      "CrIBo: Self-supervised learning via cross-image object-level bootstrapping.",
      "Capturing the objects of vision with neural networks.",
      "Object-centric learning with slot attention.",
      "Bridging the gap to real-world object-centric learning.",
      "Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers.",
      "Zero-shot object-centric representation learning, 2024.",
      "Slotdiffusion: Object-centric generative modeling with diffusion models.",
      "Object-centric slot diffusion."
    ],
    "citations": [
      7000,
      122,
      2517,
      2518,
      2514,
      1017696,
      494222,
      145315,
      320682,
      171699,
      167570,
      176167,
      274563,
      903884,
      261149,
      495332
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 75,
      "Final_Score": 84.80000000000001,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 75\nFinal_Score: 83"
    },
    "overall_score": 85.48,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "7d2JwGbxhA",
        "forum": "7d2JwGbxhA",
        "number": 12641,
        "title": "Object-Centric Pretraining via Target Encoder Bootstrapping",
        "normalized_title": "object centric pretraining via target encoder bootstrapping",
        "abstract": "Object-centric representation learning has recently been successfully applied to real-world datasets. This success can be attributed to pretrained non-object-centric foundation models, whose features serve as reconstruction targets for slot attention. However, targets must remain frozen throughout the training, which sets an upper bound on the performance object-centric models can attain. Attempts to update the target encoder by bootstrapping result in large performance drops, which can be attributed to its lack of object-centric inductive biases, causing the object-centric model’s encoder to drift away from representations useful as reconstruction targets. To address these limitations, we propose **O**bject-**CE**ntric Pretraining by Target Encoder **BO**otstrapping, a self-distillation setup for training object-centric models from scratch, on real-world data, for the first time ever. In OCEBO, the target encoder is updated as an exponential moving average of the object-centric model, thus explicitly being enriched with object-centric inductive biases introduced by slot attention while removing the upper bound on performance present in other models. We mitigate the slot collapse caused by random initialization of the target encoder by introducing a novel cross-view patch filtering approach that limits the supervision to sufficiently informative patches. When pretrained on 241k images from COCO, OCEBO achieves unsupervised object discovery performance comparable to that of object-centric models with frozen non-object-centric target encoders pretrained on hundreds of millions of images. The code and pretrained models are publicly available at https://github.com/djukicn/ocebo.",
        "authors": [
          "Nikola Đukić",
          "Tim Lebailly",
          "Tinne Tuytelaars"
        ],
        "decision": null,
        "pdate": 1737562916432,
        "cdate": 1727482792187
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "7d2JwGbxhA",
          "forum": "7d2JwGbxhA",
          "number": 12641,
          "title": "Object-Centric Pretraining via Target Encoder Bootstrapping",
          "normalized_title": "object centric pretraining via target encoder bootstrapping",
          "abstract": "Object-centric representation learning has recently been successfully applied to real-world datasets. This success can be attributed to pretrained non-object-centric foundation models, whose features serve as reconstruction targets for slot attention. However, targets must remain frozen throughout the training, which sets an upper bound on the performance object-centric models can attain. Attempts to update the target encoder by bootstrapping result in large performance drops, which can be attributed to its lack of object-centric inductive biases, causing the object-centric model’s encoder to drift away from representations useful as reconstruction targets. To address these limitations, we propose **O**bject-**CE**ntric Pretraining by Target Encoder **BO**otstrapping, a self-distillation setup for training object-centric models from scratch, on real-world data, for the first time ever. In OCEBO, the target encoder is updated as an exponential moving average of the object-centric model, thus explicitly being enriched with object-centric inductive biases introduced by slot attention while removing the upper bound on performance present in other models. We mitigate the slot collapse caused by random initialization of the target encoder by introducing a novel cross-view patch filtering approach that limits the supervision to sufficiently informative patches. When pretrained on 241k images from COCO, OCEBO achieves unsupervised object discovery performance comparable to that of object-centric models with frozen non-object-centric target encoders pretrained on hundreds of millions of images. The code and pretrained models are publicly available at https://github.com/djukicn/ocebo.",
          "authors": [
            "Nikola Đukić",
            "Tim Lebailly",
            "Tinne Tuytelaars"
          ],
          "decision": null,
          "pdate": 1737562916432,
          "cdate": 1727482792187
        }
      ]
    }
  },
  {
    "paper_id": 49912,
    "title": "UDora: A Unified Red Teaming Framework against LLM Agents by Dynamically Hijacking Their Own Reasoning",
    "abstract": "Large Language Model (LLM) agents equipped with external tools have become increasingly powerful for handling complex tasks such as web shopping, automated email replies, and financial trading. However, these advancements also amplify the risks of adversarial attacks, particularly when LLM agents can access sensitive external functionalities. Moreover, because LLM agents engage in extensive reasoning or planning before executing final actions, manipulating them into performing targeted malicious actions or invoking specific tools remains a significant challenge. Consequently, directly embedding adversarial strings in malicious instructions or injecting malicious prompts into tool interactions has become less effective against modern LLM agents. In this work, we presentUDora, a unified red teaming framework designed for LLM Agents that dynamically leverages the agent’s own reasoning processes to compel it toward malicious behavior. Specifically,UDorafirst samples the model’s reasoning for the given task, then automatically identifies multiple optimal positions within these reasoning traces to insert targeted perturbations. Subsequently, it uses the modified reasoning as the objective to optimize the adversarial strings. By iteratively applying this process, the LLM agent will then be induced to undertake designated malicious actions or to invoke specific malicious tools. Our approach demonstrates superior effectiveness compared to existing methods across three LLM agent datasets.",
    "introduction": "Figure 1:A red-teaming example involves a real AI email agent based on Microsoft’s AutoGen, designed for automated email replies. The adversary’s goal is to steal a victim user’s recent private emails by sending an adversarial email. The target tool here for UDora to trigger is GMAIL_FETCH_EMAILS.\n\n\nFigure 2:OurUDoraframework for attacking LLM agents. We explore two red-teaming scenarios:malicious environments, where the adversarial string is inserted into the observation, andmalicious instructions, where the string is directly inserted into the adversary’s instruction. The optimization process involves:Step 1, gathering the initial response from the LLM agent;Step 2, identifying the optimal position in the response for noise insertion (e.g., target item name or function name);Step 3, optimizing the string to maximize the likelihood of the noise within the modified response. These steps are repeated until the adversarial string successfully misleads the agent into performing the target malicious action within its reasoning style.\n\n\nLarge Language Models (LLMs)(Brown et al.,2020; Achiam et al.,2023; Touvron et al.,2023; Jiang et al.,2023)have demonstrated remarkable capabilities by training on massive, web-crawled datasets that inevitably include some malicious or harmful content. To mitigate the risk of generating unsafe outputs, these models are typically further refined through alignment steps before deployment, with the goal of preventing them from complying with malicious requests. Nevertheless, adversaries can still continually seek to bypass these alignment mechanisms—commonly referred to as “jailbreak.” One notable example is the Greedy Coordinate Gradient (GCG) attack(Zou et al.,2023), which appends an adversarial suffix to a malicious prompt (e.g., “how to make a bomb?”), optimizing for an affirmative response (e.g., “Sure, here is …”). Although such attacks can occasionally succeed in eliciting forbidden content, their impact has so far been relatively constrained, since the compromised outputs are still purely “textual” and the harmful responses from the LLMs can also often be easily found through Google searches.\n\n\nHowever, the emergence ofLLM agents, accompanied by an increasing number of powerful base models such as LLaMA 3.1(Meta,2024), Claude 3(Anthropic,2024a), and the series of GPT models(OpenAI,2024)that now support tool calling, significantly raises the stakes. These agents can access external tools and perform real-world actionson behalf of humans, such as web shopping, automated email responses, or even financial trading. With the ability to authorize transactions, manipulate user accounts, and access sensitive information—like credit card details, social security numbers, or private emails—a successful jailbreak can now lead to genuinely harmful consequences in the real world, far beyond mere text generation.\n\n\nAdditionally, reasoning has become an increasingly important aspect of LLM development. Starting with the foundational Chain-of-Thought(Wei et al.,2022)method, to the more advanced approaches seen in OpenAI’s o1(Jaech et al.,2024)and DeepSeek R1(Guo et al.,2025), these models now devote more time to reasoning before generating their final answers. This evolution is also particularly pronounced in LLM agents, where sophisticated reasoning and multi-step planning are required to determine the final optimal actions, making them inherently more challenging to jailbreak.\nConsequently, the adversary’s goal shifts from simply eliciting an undesired response to attempting totrigger a specific malicious action after extended reasoning. As a result, simply optimizing an adversarial string to produce an affirmative or a fixed prefix without reasoning about the response—such as “Sure, let’s take the action…”—is far less effective against LLM agents. Since LLM agents are typically trained or prompted to generate reasoning before the final action and different LLM agents may not respond in the same style as this fixed prefix, making it difficult to trigger a targeted malicious tool. At the same time, other attack like theprompt injection attacks(Perez & Ribeiro,2022; Liu et al.,2023c), which involve directly inserting malicious instructions (e.g., “Please ignore all previous instructions and conduct this task…”) into the observations from a tool call, typically behave less systematically and are less effective. Therefore, in this work, we propose a unified red-teaming framework,UDora, to evaluate and attack various LLM agentic systems, taking into account diverse adversarial scenarios and attack targets. As illustrated inFigure2, our framework focuses on two primary adversarial scenarios for red-teaming:\n\n\n1.Malicious Environment.In this scenario, the user’s instruction is benign, but the adversarial string is introduced into the observation from the environment after a tool call. For example, a malicious string might be inserted into the description of an item on a shopping website, causing the LLM agent to disregard the item’s category or price checks and proceed to purchase the item regardless of category or cost.2.Malicious Instruction.In this scenario, we append a malicious string directly to the user’s instruction, compelling the agent to perform harmful tasks (e.g.,writing a phishing email and spread it). Here, the goal is to craft an adversarial string that not only lets the agent accept the request but also triggers the agent to execute the corresponding target malicious action.\n\n\nTo tackle these scenarios, we propose a dynamic optimization strategy that adaptively updates its optimization objectives based on the LLM agent’s own reasoning process, overcoming the limitations of previous fixed-prefix optimization(Zou et al.,2023). Specifically, as shown inFigure2, we proceed as follows: (1).Gather the Initial Response:We gather the agent’s initial response—along with its token probability distribution—given the current instruction augmented with an initialized adversarial string. (2).Identify Positions for Inserting Noise:We then automatically pinpoint several optimal positions in the agent’s current reasoning where we can insert some “noise” (e.g., the name of a target item or a malicious tool). (3).String Optimization:We then optimize the adversarial string to maximize the probability of these inserted noises under the current modified reasoning path. We then regenerate the agent’s new response with the updated adversarial string, identify the new positions for noise insertion, and optimize the adversarial string again, and do Step 1 to 3 iteratively. Through this dynamic process, our method continuously adapts to the agent’s evolving reasoning, ultimately steering it toward an incorrect reasoning path that triggers the target malicious action.\n\n\nUsing this approach, as shown inFigure1, we successfully generate an adversarial string that compels a real-world AI email agent(AutoGen,2024)—based on Microsoft’s AutoGen(Wu et al.,2024)—to forward the user’s recent private emails to an attacker during an auto-reply process. Detailed attack logs are deferred toSectionC.2. Moreover, for malicious-observation benchmarks such as InjecAgent(Zhan et al.,2024)and WebShop(Yao et al.,2022), we attain an Attack Success Rate (ASR) of99%and61.75%with LLaMA 3.1(Meta,2024), and97%and91.50%with Ministral(MistralAI,2024), respectively. Additionally, on the malicious-instruction benchmark AgentHarm, we achieve an ASR of97.7%on LLaMA 3.1, and100.0%on Ministral.\n\n\nOverall, our contributions can be summarized as follows:•We presentUDora, a unified red-teaming framework that spans two different adversarial scenarios on LLM agentic systems.•We introduce a novel optimization strategy for adversarial strings that leverages the agent’s own reasoning process, strategically inserting “noise” and iteratively optimizing the adversarial string to adaptively attack LLM agents, regardless of the reasoning style of the underlying base LLMs.•We also achieve the highest ASR across three diverse datasets, outperforming all existing baselines, and successfully attack a real-world agent.\n\n",
    "related_work": "Adversarial Attacks on LLMs.Attacking language models is generally more challenging than attacking image-based models due to the discrete nature of language tokens. This discreteness hampers the direct application of gradient-based methods (e.g., PGD attack(Madry et al.,2018)) in the text domain.HotFlip(Ebrahimi et al.,2018)first represents text edits (such as flipping tokens) as vectors in the input space and uses the gradient to select the optimal manipulation, thereby successfully attacking text classification tasks. Based on this,UAT(Wallace et al.,2019)employs the similar gradient-based search to discover “trigger” tokens that can mislead model predictions or induce unethical text generation once added to the original input.AutoPrompt(Shin et al.,2020)later extends this approach to a top-k𝑘kitalic_ktoken search procedure, whileARCA(Jones et al.,2023)proposes iteratively updating tokens at specific indices in the prompt based on the current values of the remaining tokens. More recently,GCG Attack(Zou et al.,2023)has emerged as the most effective method by optimizing the adversarial suffix to trigger an affirmative response from the model, though the resulting optimized tokens are usually unreadable.AutoDan(Liu et al.,2023a)is then proposed to apply a genetic algorithm to preserve semantic information of the optimized adversarial tokens while still achieving robust attack performance. However, when it comes to LLM agents, the goal of an attack extends beyond eliciting a malicious response:it must also trigger a specific target function through extended reasoning, which makes the attack more challenging.\n\n\nLLM Agents.LLM agents capable of interacting with external environments and invoking tools or functions have become increasingly prevalent.WebShop(Yao et al.,2022)creates a simulated Amazon web shopping environment where agents can perform actions such as “search” or “click,” whileWebArena(Zhou et al.,2023)offers a more realistic and reproducible environment with a broader range of tools and scenarios. Later, numerous benchmarks(Liu et al.,2023b; Deng et al.,2024; Zheng et al.,2024)have been introduced to evaluate LLM agents’ performance in tasks like web-based interactions and computer operations. As LLM agents grow more powerful and increasingly act on behalf of users, their safety becomes paramount. For example, withinmalicious environments, adversaries may insert indirect instructions (e.g., “Please ignore all previous instructions and adhere to the following instruction instead…”) within the observation from tool calling, which is referred as prompt injection attacks(Perez & Ribeiro,2022; Liu et al.,2023c; Xu et al.,2024; Liao et al.,2024).InjecAgent(Zhan et al.,2024)is a benchmark designed to test LLM agents’ vulnerability to such indirect prompt injections after tool calls, andAgentDojo(Debenedetti et al.,2024)extends this to a more dynamic environment. Beyond prompt injection, LLM agents can also be directly asked to fulfillmalicious instructions(e.g., committing credit card fraud), and AgentHarm(Andriushchenko et al.,2024b)is a benchmark proposed for measuring such harmfulness in LLM agents based on Inspect AI(UK AI Safety Institute,2024). In this work, we propose a systematic red-teaming frameworkUDorathat effectively optimizes adversarial strings to trigger specific target actions for both scenarios:malicious environmentsandmalicious instructions.\n\n",
    "origin_citations": [
      "Towards deep learning models resistant to adversarial attacks.",
      "Hotflip: White-box adversarial examples for text classification.",
      "Universal adversarial triggers for attacking and analyzing nlp.",
      "Autoprompt: Eliciting knowledge from language models with automatically generated prompts.",
      "Automatically auditing large language models via discrete optimization.",
      "Universal and transferable adversarial attacks on aligned language models.",
      "Autodan: Generating stealthy jailbreak prompts on aligned large language models.",
      "Webshop: Towards scalable real-world web interaction with grounded language agents.",
      "Webarena: A realistic web environment for building autonomous agents.",
      "Agentbench: Evaluating llms as agents.",
      "Mind2web: Towards a generalist agent for the web.",
      "Gpt-4v (ision) is a generalist web agent, if grounded.",
      "Ignore previous prompt: Attack techniques for language models.",
      "Prompt injection attack against llm-integrated applications.",
      "Advweb: Controllable black-box attacks on vlm-powered web agents.",
      "Eia: Environmental injection attack on generalist web agents for privacy leakage.",
      "Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents.",
      "Agentdojo: A dynamic environment to evaluate attacks and defenses for llm agents.",
      "Agentharm: A benchmark for measuring harmfulness of llm agents.",
      "Inspect ai: Framework for large language model evaluations, May 2024."
    ],
    "citations": [
      917,
      889,
      7586,
      7787,
      77350,
      3027,
      323411,
      3046,
      7674,
      3095,
      3049,
      485717,
      3099,
      3106,
      741327,
      699991,
      291191,
      1062919,
      1067495
    ],
    "extract_rate": 0.95,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 87,
      "Final_Score": 84.19999999999999,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 87\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 80,
      "Final_Score": 86.30000000000001,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 80\nFinal_Score: 86"
    },
    "overall_score": 85.46000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2025/Conference",
        "venue": "ICML",
        "year": "2025",
        "openreview_id": "pRmxQHgjb1",
        "forum": "pRmxQHgjb1",
        "number": 15515,
        "title": "UDora: A Unified Red Teaming Framework against LLM Agents by Dynamically Hijacking Their Own Reasoning",
        "normalized_title": "udora a unified red teaming framework against llm agents by dynamically hijacking their own reasoning",
        "abstract": "Large Language Model (LLM) agents equipped with external tools have become increasingly powerful for complex tasks such as web shopping, automated email replies, and financial trading. However, these advancements amplify the risks of adversarial attacks, especially when agents can access sensitive external functionalities. Nevertheless, manipulating LLM agents into performing targeted malicious actions or invoking specific tools remains challenging, as these agents extensively reason or plan before executing final actions. In this work, we present UDora, a unified red teaming framework designed for LLM agents that dynamically hijacks the agent's reasoning processes to compel malicious behavior. Specifically, UDora first generates the model’s reasoning trace for the given task, then automatically identifies optimal points within this trace to insert targeted perturbations. The resulting perturbed reasoning is then used as a surrogate response for optimization. By iteratively applying this process, the LLM agent will then be induced to undertake designated malicious actions or to invoke specific malicious tools. Our approach demonstrates superior effectiveness compared to existing methods across three LLM agent datasets. The code is available at https://github.com/AI-secure/UDora.",
        "authors": [
          "Jiawei Zhang",
          "Shuang Yang",
          "Bo Li"
        ],
        "decision": null,
        "pdate": 1746105708315,
        "cdate": 1737705093360
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2025/Conference",
          "venue": "ICML",
          "year": "2025",
          "openreview_id": "pRmxQHgjb1",
          "forum": "pRmxQHgjb1",
          "number": 15515,
          "title": "UDora: A Unified Red Teaming Framework against LLM Agents by Dynamically Hijacking Their Own Reasoning",
          "normalized_title": "udora a unified red teaming framework against llm agents by dynamically hijacking their own reasoning",
          "abstract": "Large Language Model (LLM) agents equipped with external tools have become increasingly powerful for complex tasks such as web shopping, automated email replies, and financial trading. However, these advancements amplify the risks of adversarial attacks, especially when agents can access sensitive external functionalities. Nevertheless, manipulating LLM agents into performing targeted malicious actions or invoking specific tools remains challenging, as these agents extensively reason or plan before executing final actions. In this work, we present UDora, a unified red teaming framework designed for LLM agents that dynamically hijacks the agent's reasoning processes to compel malicious behavior. Specifically, UDora first generates the model’s reasoning trace for the given task, then automatically identifies optimal points within this trace to insert targeted perturbations. The resulting perturbed reasoning is then used as a surrogate response for optimization. By iteratively applying this process, the LLM agent will then be induced to undertake designated malicious actions or to invoke specific malicious tools. Our approach demonstrates superior effectiveness compared to existing methods across three LLM agent datasets. The code is available at https://github.com/AI-secure/UDora.",
          "authors": [
            "Jiawei Zhang",
            "Shuang Yang",
            "Bo Li"
          ],
          "decision": null,
          "pdate": 1746105708315,
          "cdate": 1737705093360
        }
      ]
    }
  },
  {
    "paper_id": 28033,
    "title": "Language Guided Skill Discovery",
    "abstract": "Skill discovery methods enable agents to learn diverse emergent behaviors without explicit rewards. To make learned skills useful for downstream tasks, obtaining a semantically diverse repertoire of skills is crucial.\nWhile some approaches use discriminators to acquire distinguishable skills and others focus on increasing state coverage, the direct pursuit of ‘semantic diversity’ in skills remains underexplored. We hypothesize that leveraging the semantic knowledge of large language models (LLM) can lead us to improve semantic diversity of resulting behaviors.\nIn this sense, we introduce Language Guided Skill Discovery (LGSD), a skill discovery framework that aims to directly maximize the semantic diversity between skills. LGSD takes user prompts as input and outputs a set of semantically distinctive skills. The prompts serve as a means to constrain the search space into a semantically desired subspace, and the generated LLM outputs guide the agent to visit semantically diverse states within the subspace.\nWe demonstrate that LGSD enables legged robots to visit different user-intended areas on a plane by simply changing the prompt. Furthermore, we show that language guidance aids in discovering more diverse skills compared to five existing skill discovery methods in robot-arm manipulation environments. Lastly, LGSD provides a simple way of utilizing learned skills via natural language.",
    "introduction": "One of the key capabilities of intelligent agents is to autonomously learn useful skills applicable to downstream tasks without task-specific objectives.\nConsider how human infants acquire manipulation skills through simple play with toys, which can later be applied to a broader set of tasks, such as using a spoon or holding a bottle. Our research aims to emulate this process of skill learning. Prior works in unsupervised skill discovery suggest that maximizing diversity of behaviors can be a way to develop such skills(Gregor et al.,2016; Eysenbach et al.,2018; Kwon,2020). These approaches operate under the assumption that acquiring a wide range of diverse behaviors may lead to the development of useful skills for downstream tasks. Building on this premise, studies likeSharma et al. (2019); Hansen et al. (2019); Liu & Abbeel (2021a); Laskin et al. (2022)have associated behaviors with random vectors by maximizing Mutual Information between them to acquire diverse skills.\nAdditionally, there are other approaches aimed at increasing coverage of the state space(Liu & Abbeel,2021b; Yarats et al.,2021; Park et al.,2021),\nor exploration-based strategies that leverage errors in predictive models to enhance learning(Burda et al.,2018; Pathak et al.,2019; Park et al.,2023a).\n\n\nHowever, we contend that these measures are proxies for what we term “semantic diversity” and they do not necessarily reflect the semantic diversity of a repertoire of skills. For instance, in high degrees of freedom (DOF) systems, such as robot-arm manipulation tasks, unstructured movements of each DOF might exhibit high diversity in terms of state coverage or distinctiveness from a neural discriminator, but nonetheless lack meaningful semantic diversity. To more accurately measure semantic diversity, we propose the use of Large Language Models (LLM) as they are well-suited for understanding and assessing the semantic significance of different behaviors in a way that better aligns with human judgment.\n\n\nFigure 1:We proposed LGSD  which can discover a semantically distinctive set of skills. We showcase four sample skills acquired from a single training run. Our approach successfully learned skills that manipulate only ‘edible’ objects (banana and meat_can) from a total of four objects.\n\n\nIn this work, we presentLanguage Guided Skill Discovery (LGSD), a skill discovery algorithm that utilizes the guidance of LLMs to learn semantically diverse skills. LGSD aims to fulfill three desiderata for skill discovery. (i) Firstly, we want to discover skills that have diverse semantic meanings. We use LLMs to generate descriptions for each of agent states. Based on these descriptions, we measure the semantic difference between states usinglanguage-distance, and train skills to maximize this distance. Because LLMs are equipped with a semantic understanding of the world, this provides a way to maximizing semantic differences between skills. Fig.1illustrates skills learned from LGSD, showcasing the semantic diversity of resultant behaviors. (ii) Secondly, we seek to restrict the search space for skills into a desired semantic subspace.\nIntuitively, restricting the skill space is akin to enforcing an infant to grab toys only with their hands, and not with their mouth or feet. We implement this desideratum by utilizing language prompts. Humans provide prompts to LLMs, and the generated descriptions are then constrained by these prompts. This will allow our method to ignore differences outside the intended semantic space.\n(iii) Lastly, we aim for the learned skills to be easy to use. When the skill space is continuous, merely selecting the best skills for target tasks becomes non-trivial. LGSD supports a zero-shot language instruction-following capability by training separate networkψ𝜓\\psiitalic_ψfor skill inference. Humans can provide a natural language description of the desired state, and then LGSD can infer which skill should be used to reach that goal state.\n\n\nThe primary contributions of this work are as follows. 1) We propose a skill discovery framework that enables the discovery of semantically diverse skills by utilizing language guidance from LLMs. To the best of our knowledge, this is the first work to incorporate semantic diversity into skill discovery. 2) We present a method of constraining the skill search space to a user-defined semantic space using language prompts. This allows users to control the resulting skills by simply specifying different prompts. We emphasize that, prior to training, existing methods only offer limited control over the skills to be learned.\n3) We provide a theoretical proof of howlanguage-distance, a proxy for semantic distance, can be employed as a validpseudometric. 4) We propose a method that enables an agent to reach a goal state specified by natural language, facilitating the convenient use of learned skills. 5) We demonstrate that our method can train semantically diverse skills. Through extensive experiments, we show that LGSD outperforms five different skill discovery baseline methods on both locomotion and manipulation tasks, in terms of both diversity and sample efficiency.\n\n",
    "related_work": "2.1Unsupervised skill discoveryMutual information based approachCommon approaches to associate skills with corresponding behaviors are based on maximizing the mutual information (MI) between statesS𝑆Sitalic_Sand skillsZ𝑍Zitalic_Z, orℐ⁢(S;Z)ℐ𝑆𝑍\\mathcal{I}(S;Z)caligraphic_I ( italic_S ; italic_Z )(Gregor et al.,2016; Eysenbach et al.,2018; Sharma et al.,2019; Hansen et al.,2019; Liu & Abbeel,2021a; Laskin et al.,2022). MI can be viewed as a dependency measure that quantifies the association between two random variables. To maximize this quantity, a popular early formulation utilizes the decompositionℐ⁢(S;Z)=ℋ⁢(Z)−ℋ⁢(Z|S)ℐ𝑆𝑍ℋ𝑍ℋconditional𝑍𝑆\\mathcal{I}(S;Z)=\\mathcal{H}(Z)-\\mathcal{H}(Z|S)caligraphic_I ( italic_S ; italic_Z ) = caligraphic_H ( italic_Z ) - caligraphic_H ( italic_Z | italic_S ), whereℋ⁢[⋅]ℋdelimited-[]⋅\\mathcal{H}[\\cdot]caligraphic_H [ ⋅ ]refers to Shannon entropy. Since directly computingℋ⁢(Z|S)ℋconditional𝑍𝑆\\mathcal{H}(Z|S)caligraphic_H ( italic_Z | italic_S )is intractable,Eysenbach et al. (2018)introduce a variational posteriorqϕ⁢(z|s)subscript𝑞italic-ϕconditional𝑧𝑠q_{\\phi}(z|s)italic_q start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPT ( italic_z | italic_s )to optimize the lower bound. Similarly, CIC(Laskin et al.,2022)maximizesℐ⁢((s,s′);Z)ℐ𝑠superscript𝑠′𝑍\\mathcal{I}((s,s^{\\prime});Z)caligraphic_I ( ( italic_s , italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT ) ; italic_Z ), promoting distinct state transitions across different skills. However, a significant limitation of using MI as the objective arises from the use of Kullback–Leibler (KL) divergence to define MI:ℐ(S;Z)=defDK⁢L(P(S,Z)||P(S)P(Z)).\\mathcal{I}(S;Z)\\stackrel{{\\scriptstyle\\text{def}}}{{=}}D_{KL}(P(S,Z)||P(S)P(Z%\n)).caligraphic_I ( italic_S ; italic_Z ) start_RELOP SUPERSCRIPTOP start_ARG = end_ARG start_ARG def end_ARG end_RELOP italic_D start_POSTSUBSCRIPT italic_K italic_L end_POSTSUBSCRIPT ( italic_P ( italic_S , italic_Z ) | | italic_P ( italic_S ) italic_P ( italic_Z ) ) .Maximization of MI using KL divergence often results in skills with less distinctive behaviors. This is because KL divergence is fully maximized when two densities have no overlap, posing a problem as there is no further incentive to distinguish the densities beyond this point. This is problematic because having no overlap between two densities does not necessarily mean two skills are noticeably distinctive. In practice, this becomes a severe shortcoming of DIAYN(Eysenbach et al.,2018), where its objective can be fully optimized as long as the discriminatorqϕsubscript𝑞italic-ϕq_{\\phi}italic_q start_POSTSUBSCRIPT italic_ϕ end_POSTSUBSCRIPTcan distinguish them perfectly. Neural networks can easily distinguish minor numerical differences, so the skills learned with this objective often have only these minute differences.Distance maximization approachTo overcome these shortcomings, recent skill discovery works(He et al.,2022; Park et al.,2021;2023a;2023b)propose using the Wasserstein Dependency Measure (WDM,Ozair et al. (2019)):ℐ𝒲(S;Z)=def𝒲(P(S,Z)||P(S)P(Z))\\mathcal{I_{W}}(S;Z)\\stackrel{{\\scriptstyle\\text{def}}}{{=}}\\mathcal{W}(P(S,Z)%\n||P(S)P(Z))caligraphic_I start_POSTSUBSCRIPT caligraphic_W end_POSTSUBSCRIPT ( italic_S ; italic_Z ) start_RELOP SUPERSCRIPTOP start_ARG = end_ARG start_ARG def end_ARG end_RELOP caligraphic_W ( italic_P ( italic_S , italic_Z ) | | italic_P ( italic_S ) italic_P ( italic_Z ) )WDM is defined using the Earth-Mover (EM) distance and retains the advantage of always providing an incentive to maximize the distance between distributions, even when they are already distinguishable.An essential characteristic ofℐ𝒲subscriptℐ𝒲\\mathcal{I_{W}}caligraphic_I start_POSTSUBSCRIPT caligraphic_W end_POSTSUBSCRIPTis that it must be defined within ametricspace.\nTherefore, the choice ofmetricto measure the difference between skills comes to the gist of the algorithm since it ultimately governs the resulting behavior of each skills. LSD(Park et al.,2021)employs the Euclidean distance in state space as its metric. Maximizingℐ𝒲subscriptℐ𝒲\\mathcal{I_{W}}caligraphic_I start_POSTSUBSCRIPT caligraphic_W end_POSTSUBSCRIPTunder this metric encourages agents to visit states that are are far apart in terms of Euclidean distances.\nConversely, CSD(Park et al.,2023a)points out the problem of using Euclidean distance as the metric. They argue that maximizing Euclidean distance tends to focus on “easy” changes in the state space, such as moving the robot arm itself rather than the target object.\nThus, CSD proposed to maximize “controllability-aware distance” instead. This metric measures the negative log likelihood of a transition dynamics−log⁡p⁢(s′|s)𝑝conditionalsuperscript𝑠′𝑠-\\log{p(s^{\\prime}|s)}- roman_log italic_p ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT | italic_s ). Using this metric means the algorithm favors visiting rare transitions. LGSD adopts a similar approach but differentiates itself by employing a language-based distance metric within the WDM framework, to maximize the semantic difference.\n\n\n2.2Using language as guidance for behavior learningRecent work has studied using semantic knowledge in large pre-trained models for control tasks in several ways. A popular approach is to use the common-sense reasoning capabilities of LLMs to produce high-level plans over known, existing skills through the interface of natural language(Sharma et al.,2021; Ahn et al.,2022; Singh et al.,2023; Huang et al.,2022; Ha et al.,2023)or generated programs(Liang et al.,2022; Zeng et al.,2022; Vemprala et al.,2024). In addition, LLMs have also been shown to be effective in directly producing low-level actions that can be executed by an agent(Wang et al.,2023; Kwon et al.,2023; Driess et al.,2023; Brohan et al.,2023). Our work does not use LLMs to generateactionsat any level of abstraction but to provide a meaningful distance measure to guide learning towards discovering semantically meaningful skills.Several works have used LLMs to guide agents to learn semantically meaningful behavior by using them to design reward functions for either a given desired task(Yu et al.,2023; Ma et al.,2023). This showcases LLMs’ capacity to understand specific physical embodiments assorting it with semantic knowledge to reason about given tasks. We note that these works aim to solve a single specific task, whereas our method aims to learn a repertoire of diverse behaviors without explicit supervision.The work most related to ours isDu et al. (2023), which employs an LLM to guide exploration. Their approach uses the LLM to suggest plausible set of goals for the agent, maximizing the cosine similarity between the language embeddings of the goals and the current state. While they focus on using an LLM to reason with a low-dimensional, discrete set of high-level actions in complex environments like Crafter and Housekeep, our work focus on discovering meaningful behaviors in complex, continuous control spaces where actions (e.g. joint targets) are not as readily interpretable. Furthermore, their reward function, based on language similarity, lacks theoretical support as it is not a validmetric. In contrast, our method effectively handles thelanguage-distancemeasure and provides a theoretical framework for its use. We elaborate on this in Section4.3.\n\n",
    "origin_citations": [
      "Variational intrinsic control.",
      "Diversity is all you need: Learning skills without a reward function.",
      "Dynamics-aware unsupervised discovery of skills.",
      "Fast task inference with variational intrinsic successor features.",
      "Aps: Active pretraining with successor features.",
      "Unsupervised reinforcement learning with contrastive intrinsic control.",
      "Wasserstein unsupervised reinforcement learning.",
      "Lipschitz-constrained unsupervised skill discovery.",
      "Controllability-aware unsupervised skill discovery.",
      "Metra: Scalable unsupervised rl with metric-aware abstraction.",
      "Wasserstein dependency measure for representation learning.",
      "Skill induction and planning with latent language.",
      "Do as i can, not as i say: Grounding language in robotic affordances.",
      "Progprompt: Generating situated robot task plans using large language models.",
      "Inner monologue: Embodied reasoning through planning with language models.",
      "Scaling up and distilling down: Language-guided robot skill acquisition.",
      "Code as policies: Language model programs for embodied control.",
      "Socratic models: Composing zero-shot multimodal reasoning with language.",
      "Chatgpt for robotics: Design principles and model abilities.",
      "Prompt a robot to walk with large language models.",
      "Language models as zero-shot trajectory generators.",
      "Palm-e: An embodied multimodal language model.",
      "Rt-2: Vision-language-action models transfer web knowledge to robotic control.",
      "Language to rewards for robotic skill synthesis.",
      "Eureka: Human-level reward design via coding large language models.",
      "Guiding pretraining in reinforcement learning with large language models."
    ],
    "citations": [
      26739,
      389262,
      159461,
      268342,
      1004429,
      705199,
      90918,
      489789,
      320160,
      400667,
      343381,
      7658,
      177711,
      2997,
      468075,
      3031,
      5575,
      227363,
      480946,
      315335,
      2963,
      7672,
      304212,
      314683,
      489039
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 78,
      "Final_Score": 80.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 78\nFinal_Score: 81"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 90,
      "Final_Score": 88.5,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 90\nFinal_Score: 89"
    },
    "overall_score": 85.42,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "i3e92uSZCp",
        "forum": "i3e92uSZCp",
        "number": 7461,
        "title": "Language Guided Skill Discovery",
        "normalized_title": "language guided skill discovery",
        "abstract": "Skill discovery methods enable agents to learn diverse emergent behaviors without explicit rewards. To make learned skills useful for downstream tasks, obtaining a semantically diverse repertoire of skills is crucial. While some approaches use discriminators to acquire distinguishable skills and others focus on increasing state coverage, the direct pursuit of ‘semantic diversity’ in skills remains underexplored. We hypothesize that leveraging the semantic knowledge of large language models (LLM) can lead us to improve semantic diversity of resulting behaviors. In this sense, we introduce Language Guided Skill Discovery (LGSD), a skill discovery framework that aims to directly maximize the semantic diversity between skills. LGSD takes user prompts as input and outputs a set of semantically distinctive skills. The prompts serve as a means to constrain the search space into a semantically desired subspace, and the generated LLM outputs guide the agent to visit semantically diverse states within the subspace. We demonstrate that LGSD enables legged robots to visit different user-intended areas on a plane by simply changing the prompt. Furthermore, we show that language guidance aids in discovering more diverse skills compared to five existing skill discovery methods in robot-arm manipulation environments. Lastly, LGSD provides a simple way of utilizing learned skills via natural language.",
        "authors": [
          "Seungeun Rho",
          "Laura Smith",
          "Tianyu Li",
          "Sergey Levine",
          "Xue Bin Peng",
          "Sehoon Ha"
        ],
        "decision": null,
        "pdate": 1737562530709,
        "cdate": 1727367042735
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "i3e92uSZCp",
          "forum": "i3e92uSZCp",
          "number": 7461,
          "title": "Language Guided Skill Discovery",
          "normalized_title": "language guided skill discovery",
          "abstract": "Skill discovery methods enable agents to learn diverse emergent behaviors without explicit rewards. To make learned skills useful for downstream tasks, obtaining a semantically diverse repertoire of skills is crucial. While some approaches use discriminators to acquire distinguishable skills and others focus on increasing state coverage, the direct pursuit of ‘semantic diversity’ in skills remains underexplored. We hypothesize that leveraging the semantic knowledge of large language models (LLM) can lead us to improve semantic diversity of resulting behaviors. In this sense, we introduce Language Guided Skill Discovery (LGSD), a skill discovery framework that aims to directly maximize the semantic diversity between skills. LGSD takes user prompts as input and outputs a set of semantically distinctive skills. The prompts serve as a means to constrain the search space into a semantically desired subspace, and the generated LLM outputs guide the agent to visit semantically diverse states within the subspace. We demonstrate that LGSD enables legged robots to visit different user-intended areas on a plane by simply changing the prompt. Furthermore, we show that language guidance aids in discovering more diverse skills compared to five existing skill discovery methods in robot-arm manipulation environments. Lastly, LGSD provides a simple way of utilizing learned skills via natural language.",
          "authors": [
            "Seungeun Rho",
            "Laura Smith",
            "Tianyu Li",
            "Sergey Levine",
            "Xue Bin Peng",
            "Sehoon Ha"
          ],
          "decision": null,
          "pdate": 1737562530709,
          "cdate": 1727367042735
        }
      ]
    }
  },
  {
    "paper_id": 29188,
    "title": "BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models",
    "abstract": "Large Language Models (LLMs) often suffer from overconfidence during inference, particularly when adapted to downstream domain-specific tasks with limited data.\nPrevious work addresses this issue by employing approximate Bayesian estimationafterthe LLMs are trained, enabling them to quantify uncertainty.\nHowever, such post-training approaches’ performance is severely limited by the parameters learnedduringtraining.\nIn this paper, we go beyond post-training Bayesianization and proposeBayesianLow-Rank Adaptation byBackpropagation (BLoB), an algorithm that continuously and jointly adjusts both the mean and covariance of LLM parameters throughout the whole fine-tuning process.\nOur empirical results verify the effectiveness of BLoB in terms of generalization and uncertainty estimation, when evaluated on both in-distribution and out-of-distribution data.\nCode is available athttps://github.com/Wang-ML-Lab/bayesian-peft.",
    "introduction": "Despite the recent advancements in Large Language Models (LLMs)[9,106,105,65,15,4,91,92,77,12,1,71], addressing the challenges of reliability and responsibility remains imperative[41,100,99]. LLMs often produce overconfident responses detached from factual grounding, posing potential harm to users[3,107,43,42,87,49,7,118,111,120,32,68,115,44]. Therefore, accurately estimating response confidence (or uncertainty) is crucial to preemptively intervene before harm occurs. Current research predominantly focuses on eliciting the internal capability of uncertainty estimation of LLMs. For example, studies suggest that verbalized uncertainty yields better-calibrated results compared to conditional probability[87,44].\n\n\nWhile effective, the aforementioned methods do not offer a universal solution for expressing LLM uncertainty across all scenarios, especially when adapted[113]to domain-specific corpora, human preferences, or downstream tasks[43]. Even a well-calibrated LLM may struggle to estimate uncertainty during fine-tuning due to catastrophic forgetting of general knowledge[84]. Moreover, when applied to limited-scale downstream tasks, excessively over-parameterized LLMs can rapidly overfit, leading to overconfidence. Thus, enabling accurate uncertainty estimation of LLMs is vital for their reliable and responsible deployment.\n\n\nBayesian methods emerge as a natural solution for learning uncertainty estimation abilities among their counterparts[88,11,101,98,29,45,50,60,97,57,102,20,110]. These methods model predictive uncertaintyP⁢(𝒚|𝒙,𝒟)𝑃conditional𝒚𝒙𝒟P({\\bm{y}}|{\\bm{x}},{\\mathcal{D}})italic_P ( bold_italic_y | bold_italic_x , caligraphic_D )by marginalizing the posterior parameter distributionP⁢(𝜽|𝒟)𝑃conditional𝜽𝒟P({\\bm{\\theta}}|{\\mathcal{D}})italic_P ( bold_italic_θ | caligraphic_D )after observing the dataset𝒟𝒟{\\mathcal{D}}caligraphic_D:P⁢(𝒚|𝒙,𝒟)𝑃conditional𝒚𝒙𝒟\\displaystyle P({\\bm{y}}|{\\bm{x}},{\\mathcal{D}})italic_P ( bold_italic_y | bold_italic_x , caligraphic_D )=∫P⁢(𝒚|𝒙,𝜽)⁢P⁢(𝜽|𝒟)⁢𝑑𝜽.absent𝑃conditional𝒚𝒙𝜽𝑃conditional𝜽𝒟differential-d𝜽\\displaystyle=\\int P({\\bm{y}}|{\\bm{x}},{\\bm{\\theta}})P({\\bm{\\theta}}|{\\mathcal%\n{D}})d{\\bm{\\theta}}.= ∫ italic_P ( bold_italic_y | bold_italic_x , bold_italic_θ ) italic_P ( bold_italic_θ | caligraphic_D ) italic_d bold_italic_θ .(1)However, adapting the Bayesian framework to LLMs poses significant challenges. LLM architectures typically incorporate complex components, including non-linear activation functions, rendering exact Bayesian inference of parameter posteriors intractable, i.e., unable to compute the integral precisely. Consequently, finding an accurate approximation algorithm for the true posterior distribution becomes a primary challenge. Additionally, modeling parameter posterior distributions demands extra memory space, imposing a prohibitive burden on systems due to the massive scale of LLMs.\n\n\nContemporary methods leverage Parameter-Efficient Fine-Tuning (PEFT) to reduce the number of tunable parameters, thus alleviating computational and storage resource burdens[23,40,26,121,55,52]. Built on this, recent research explores Bayesianizing only the PEFT module during fine-tuning to calibrate LLMs[8,103,116,69], somewhat relieving the burden of introducing more parameters for posterior approximation. However, initial investigations suggest that straightforward combinations of PEFT and basic Bayesian techniques like Monte-Carlo Dropout (MCD,[29]) or Deep Ensemble (ENS,[50,8,103]) yield only marginal improvements in generalization and uncertainty estimation. The most promising results to date involve Kronecker factorized Laplace approximation, applied after maximum a posteriori (MAP) estimation provided by any optimization algorithm[116]. Nevertheless, we argue that such post-training procedures bifurcate posterior approximation into two stages, inevitably leading to suboptimal estimation.\n\n\nTo address this challenge, we proposeBayesianLow-Rank Adaptation byBackpropagation (BLoB), a Bayesian Deep Learning framework for fine-tuning LLMs with LoRA.\nBLoB jointly estimates the low-rank variational distributions’ mean and covariance throughout the entire fine-tuning stage via backpropagation. Unlike methods relying on post-training approximation, BLoB enables simultaneous estimation of both the parameter mode (i.e., the mean if one assumes Gaussian distributions) and the parameter variance. Random sampling of model parameters based on variance estimation can enhance mode estimation. It thereby improves model performance in terms of accuracy and uncertainty estimation on both in-distribution and out-of-distribution datasets, as verified by our extensive experiments across multiple datasets.\nIn summary, our contributions are:•We propose a principled Bayesianization framework for Low-Rank Adaptation (LoRA) in Large Language Models (LLMs) by assuming that full weights’ approximate posterior distribution has a low-rank structure containing a linear combination of independent Gaussian distributions.•We show that, under mild conditions, optimization of the full-weight variational distribution can be done efficiently in the low-rank space of the weight update matrices.•We introduce BLoB, a variational Bayesian low-rank adaptation framework for LLMs that jointly learns the mean and covariance of the variational distribution during fine-tuning.•Extensive evaluations demonstrate the superiority of BLoB in terms of generalization and uncertainty estimation across different scenarios.\n\n",
    "related_work": "Parameter-Efficient Fine-Tuning (PEFT) for LLMs.Due to the prohibitively large size of LLMs, parameter-efficient fine-tuning has become a trending topic. Computational paradigms in this area include adapter-based fine-tuning[39,35,80,75,61], prompt-based fine-tuning[33,53,58,56,94,6], and partial fine-tuning[119,124,5,112,31]. Among these, LoRA[40]has gained significant attention due to its simplicity and effectiveness. Building on LoRA, numerous studies have aimed to further optimize parameter efficiency when fine-tuning large models[26,34,22,21]. For instance, KronA models weight updates as the Kronecker product of two smaller matrices without decreasing the update rank[26], and SVDiff performs Singular Value Decomposition (SVD) on the original weight matrices, fine-tuning only the singular values[34].\nHowever, in this paper, we focus solely on Bayesianizing LoRA due to its widespread application in existing works. We also note that BLoB can be naturally adapted to handle different LoRA variants.\n\n\nUncertainty Estimation in Large Language Models.Large-scale pre-trained models are well-calibrated during pre-training[43], but fail to accurately express predictive uncertainty during inference[3,107,43,42,87], especially after fine-tuning[8,103,116,69]. This indicates that measures effective during pre-training[93,114,16,122,14]may lose their power of uncertainty estimation after fine-tuning for domain-specific knowledge.\nTo address this issue,[27,123]define priors and approximate posteriors on the full attention weights during fine-tuning, achieving better uncertainty estimation but at a significant cost in time and space.\nConsequently, recent work integrates Bayesian methods and PEFT for efficient uncertainty estimation. For instance,[8,103]train and store multiple copies of different LoRAs, ensembling their outputs during inference to achieve somewhat better results.[116]applies Kronecker factorized Laplace approximation on fine-tuned LoRA. However, such post-training procedures bifurcate posterior approximation into two stages, leading to suboptimal estimation.\nIn contrast, our BLoB enables simultaneous estimation of both the mean and covariance of LLM parameters in a single fine-tuning stage, substantially improving performance.\n\n",
    "origin_citations": [
      "Parameter-efficient transfer learning for nlp.",
      "Towards a unified view of parameter-efficient transfer learning.",
      "Adapterdrop: On the efficiency of adapters in transformers.",
      "Adapterfusion: Non-destructive task composition for transfer learning.",
      "Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks.",
      "Warp: Word-level adversarial reprogramming.",
      "The power of scale for parameter-efficient prompt tuning.",
      "Gpt understands, too.",
      "Prefix-tuning: Optimizing continuous prompts for generation.",
      "Spot: Better frozen model adaptation through soft prompt transfer.",
      "Attempt: Parameter-efficient multi-task tuning via attentional mixtures of soft prompts.",
      "Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models.",
      "Masking as an efficient alternative to finetuning for pretrained language models.",
      "Composable sparse fine-tuning for cross-lingual transfer.",
      "Raise a child in large language model: Towards effective and generalizable fine-tuning.",
      "Parameter-efficient transfer learning with diff pruning.",
      "LoRA: Low-rank adaptation of large language models.",
      "Krona: Parameter efficient tuning with kronecker adapter.",
      "Svdiff: Compact parameter space for diffusion fine-tuning.",
      "Sparse low-rank adaptation of pre-trained language models.",
      "Qlora: Efficient finetuning of quantized llms.",
      "Language models (mostly) know what they know, 2022.",
      "Concrete problems in ai safety, 2016.",
      "Ethical and social risks of harm from language models, 2021.",
      "A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023.",
      "Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback.",
      "Uncertainty quantification in fine-tuned llms using lora ensembles.",
      "Lora ensembles for large language model fine-tuning, 2023.",
      "Bayesian low-rank adaptation for large language models.",
      "Gaussian stochastic weight averaging for bayesian low-rank adaptation of large language models, 2024.",
      "Bayesian layers: A module for neural network uncertainty.",
      "Bayesian transformer language models for speech recognition.",
      "Pathologies in priors and inference for bayesian transformers.",
      "Cyclical stochastic gradient mcmc for bayesian deep learning.",
      "Calibrating transformers via sparse gaussian processes.",
      "Bayesian attention modules.",
      "Bayesian attention belief networks."
    ],
    "citations": [
      7567,
      342105,
      436360,
      419165,
      194758,
      4201,
      8536,
      7789,
      1591,
      146972,
      394266,
      191234,
      340587,
      170626,
      108490,
      1589,
      139643,
      495091,
      25001,
      2253,
      2391,
      1356,
      2494,
      294654,
      258979,
      466191,
      477320,
      404799,
      704633,
      359279,
      118993,
      342253,
      515180,
      78245,
      436968,
      194401
    ],
    "extract_rate": 0.972972972972973,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 88,
      "Final_Score": 84.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 82\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.42,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "MaDykgj4Ru",
        "forum": "MaDykgj4Ru",
        "number": 8582,
        "title": "BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models",
        "normalized_title": "blob bayesian low rank adaptation by backpropagation for large language models",
        "abstract": "Large Language Models (LLMs) often suffer from overconfidence during inference, particularly when adapted to downstream domain-specific tasks with limited data. Previous work addresses this issue by employing approximate Bayesian estimation after the LLMs are trained, enabling them to quantify uncertainty. However, such post-training approaches' performance is severely limited by the parameters learned during training. In this paper, we go beyond post-training Bayesianization and propose Bayesian Low-Rank Adaptation by Backpropagation (BLoB), an algorithm that continuously and jointly adjusts both the mean and covariance of LLM parameters throughout the whole fine-tuning process. Our empirical results verify the effectiveness of BLoB in terms of generalization and uncertainty estimation, when evaluated on both in-distribution and out-of-distribution data.",
        "authors": [
          "Yibin Wang",
          "Haizhou Shi",
          "Ligong Han",
          "Dimitris N. Metaxas",
          "Hao Wang"
        ],
        "decision": null,
        "pdate": 1727287884371,
        "cdate": 1715664858063
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "MaDykgj4Ru",
          "forum": "MaDykgj4Ru",
          "number": 8582,
          "title": "BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models",
          "normalized_title": "blob bayesian low rank adaptation by backpropagation for large language models",
          "abstract": "Large Language Models (LLMs) often suffer from overconfidence during inference, particularly when adapted to downstream domain-specific tasks with limited data. Previous work addresses this issue by employing approximate Bayesian estimation after the LLMs are trained, enabling them to quantify uncertainty. However, such post-training approaches' performance is severely limited by the parameters learned during training. In this paper, we go beyond post-training Bayesianization and propose Bayesian Low-Rank Adaptation by Backpropagation (BLoB), an algorithm that continuously and jointly adjusts both the mean and covariance of LLM parameters throughout the whole fine-tuning process. Our empirical results verify the effectiveness of BLoB in terms of generalization and uncertainty estimation, when evaluated on both in-distribution and out-of-distribution data.",
          "authors": [
            "Yibin Wang",
            "Haizhou Shi",
            "Ligong Han",
            "Dimitris N. Metaxas",
            "Hao Wang"
          ],
          "decision": null,
          "pdate": 1727287884371,
          "cdate": 1715664858063
        }
      ]
    }
  },
  {
    "paper_id": 34212,
    "title": "SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain",
    "abstract": "In this paper, we introduceSaulLM-54BandSaulLM-141B, two large language models (LLMs) tailored for the legal sector. These models, which feature architectures of 54 billion and 141 billion parameters, respectively, are based on the Mixtral architecture. The development ofSaulLM-54BandSaulLM-141Bis guided by large-scale domain adaptation, divided into three strategies: (1) the exploitation of continued pretraining involving a base corpus that includes over540540540540billion of legal tokens, (2) the implementation of a specialized legal instruction-following protocol, and (3) the alignment of model outputs with human preferences in legal interpretations. The integration of synthetically generated data in the second and third steps enhances the models’ capabilities in interpreting and processing legal texts, effectively reaching state-of-the-art performance and outperforming previous open-source models onLegalBench-Instruct. This work explores the trade-offs involved in domain-specific adaptation at this scale, offering insights that may inform future studies on domain adaptation using strong decoder models. Building uponSaulLM-7B, this study refines the approach to produce an LLM better equipped for legal tasks. We are releasing base, instruct, and aligned versions on top ofSaulLM-54BandSaulLM-141Bunder the MIT License to facilitate reuse and collaborative research.",
    "introduction": "LLMs have demonstrated exceptional capabilities across various domains[1,65,58,73,37,38,74,7,26], excelling in tasks such as language translation[6], medical diagnostics[16,11,12], and automated code generation[4,40,30], among others. These achievements highlight the potential for human-like communication through large language models (LLMs). Despite the significant potential benefits, the adaptation of most recent LLMs for legal tasks has not been extensively examined, with only two recent studies cited from[18,53,83], and its impact on society could be substantial. Indeed, at a time when legal systems in many countries are overburdened[69], the development of robust and high-performing legal LLMs could provide critical support to lawyers and judicial systems[63,10]. However, adapting LLMs to the legal domain presents unique challenges, particularly because of the vast scale involved, with hundreds of billions of existing legal data available.\n\n\nPrevious efforts to tailor LLMs to the legal sector have encountered significant challenges[18,53,83]: first, a limited model scale, capped at7/127127/127 / 12B parameters, which is considerably smaller than the largest open-source models[7,38]; second, training datasets restricted to no more than30303030billion tokens, significantly fewer than potentially available tokens[28,52]. Given the importance of scale and breadth in effectively adapting LLMs to new domains, this paper aims to answer the following research question:How much can we improve the specialization of generic LLMs for legal tasks by scaling up both model and corpus size?\n\n\nIn this paper, we present an empirical study on the scalability and domain adaptation of LLMs in the legal sector. Relying on a corpus exceeding500500500500B tokens and models up to141141141141B parameters, our research seeks to address the gaps in the examination of legal applications. A novel aspect of our approach is the adaptation of large-scale Mixture of Experts (MoE) models with54545454B and141141141141B parameters, which have gained significant traction in recent months[90,22,43,87,84,59]. Formally, this study makes two principal contributions:1. Comprehensive Analysis of Domain Adaptation Strategies for Legal LLMsDomain adaptation for legal LLMs remains a challenging and somewhat underexplored area. This work advances the field by specializing each step in the process of developing modern LLMs, from continued pretraining to instruction fine-tuning and alignment, relying on both synthetic and real data. This paper offers a fresh perspective on the efficacy of each step and its value for adapting to the legal domain, potentially guiding further research in the legal domain as well as in other expert domains.2.SaulLM-54B&SaulLM-141B: Joining SaulLM-7b to form a Family of Legal LLMs under Permissive License111Model will be made available athttps://huggingface.co/.We specialize general-purpose, large-scale LLMs for the law. This work represents an ambitious advancement in terms of scale and leveraging the increasingly popular MoE architecture. While this architecture is widely used, its specific applications within focused domains, particularly the legal sector, are still largely unexplored. By releasing these models, we aim to foster further research in legal NLP and contribute to unlocking the full potential of LLMs.\n\n",
    "related_work": "2.1Domain Specialization For Large Language ModelsThe process of domain specialization for LLMs has demonstrated promising results in areas such as medicine[16], science[71], translation[6,5], or code[64,40,4]. Models like SciBERT[9], PubMedBERT[72], Galactica[71]and Meditron[16]have been specifically trained on domain-related corpora to enhance their performance. Studies have identified that both the scale of the model and the size of the in-domain data are crucial for achieving strong domain adaptation[16,64].In the legal domain, earlier models such as LegalBERT[15], InCaseLawBERT[57], and SaulLM-7B[18], among others, while pioneering, have been constrained by their relatively small scale and the specificity of their training data, which covers a limited number of documents and jurisdictions. Our work aims to build on these efforts by deploying LLMs at an unprecedented scale, utilizing models of up to 141B parameters and a base corpus exceeding 500 billion tokens to significantly enhance the depth and breadth of legal language comprehension and generation.\n\n\n2.2Legal Domain Adaptation for Modern LLMThe field of legal domain adaptation has traditionally concentrated on refining models through pretraining on specialized corpora[15,18,21]. Yet, in the current paradigm, pretraining represents just one aspect of the solution, as LLMs often utilize techniques like instruction fine-tuning and alignment, employing algorithms such as DPO[61], PPO[66]or RLHF[55,42].Recent domain-adapted models, such as SaulLM or Legal-FLAN-T5 (a closed model), have tried to improve alignment with legal instructions. However, SaulLM is a smaller model, and Legal-FLAN-T5, is based on an outdated architecture and does not leverage the extreme scale pretraining that contemporary models do. Moreover, it not being publicly available stymies progress vital for advancing research and applications in the legal sector.We believe this work pioneers a holistic approach to domain adaptation by training modern LLMs specifically for the legal domain, from pretraining to instruction fine-tuning and legal preference alignment. We demonstrate that synthetic data can be effectively utilized for alignment, advancing beyond SaulLM-7B’s use solely of instruction fine-tuning. The resulting models,SaulLM-54BandSaulLM-141B, lay the groundwork for further research and development, and expand access to high-performance legal LLMs.\n\n",
    "origin_citations": [
      "Meditron-70b: Scaling medical pretraining for large language models.",
      "Galactica: A large language model for science.",
      "Tower: An open multilingual large language model for translation-related tasks.",
      "Steering large language models for machine translation with finetuning and in-context learning.",
      "Code llama: Open foundation models for code.",
      "Starcoder: may the source be with you!",
      "Santacoder: don’t reach for the stars!",
      "Scibert: A pretrained language model for scientific text.",
      "Fine-tuning large neural language models for biomedical natural language processing.",
      "Legal-bert: The muppets straight out of law school.",
      "Pre-trained language models for the legal domain: A case study on indian law.",
      "Saullm-7b: A pioneering large language model for law.",
      "Chatlaw: Open-source legal large language model with integrated external knowledge bases.",
      "Direct preference optimization: Your language model is secretly a reward model.",
      "Proximal policy optimization algorithms.",
      "Training language models to follow instructions with human feedback.",
      "Reinforcement learning with human feedback: Learning dynamic choices via pessimism."
    ],
    "citations": [
      22950,
      5560,
      463291,
      314434,
      404859,
      7561,
      302951,
      2483,
      704245,
      356572,
      228604,
      290566,
      7229,
      5930,
      2219,
      364,
      257255
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 88,
      "Final_Score": 84.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 82\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.42,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "NLUYZ4ZqNq",
        "forum": "NLUYZ4ZqNq",
        "number": 14773,
        "title": "SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain",
        "normalized_title": "saullm 54b saullm 141b scaling up domain adaptation for the legal domain",
        "abstract": "In this paper, we introduce SaulLM-medium and SaulLM-large, two large language models (LLMs) families tailored for the legal sector. These models, which feature architectures of 54 billion and 140 billion parameters, respectively, are based on the Mixtral architecture. The development of SaulLM-54B and SaulLM-140B is guided by large-scale domain adaptation,  divided into strategies: (1) the exploitation of continued pretaining involving a legal corpus that includes over $400$ billion tokens, (2) the implementation of a specialized legal instruction-following protocol, and (3) the alignment of model outputs with human preferences in legal interpretations. The integration of synthetically generated data in the second and third steps enhances the models' capabilities in interpreting and processing legal texts, effectively reaching state-of-the-art performance and outperforming all previous open-source models on LegalBench Instruct. This research thoroughly explores the trade-offs involved in domain-specific adaptation at this scale, offering insights that may inform future studies on domain adaptation using strong decoder models. Building upon SaulLM-7B, this study refines the approach to produce an LLM better equipped for legal tasks and domains. Additionally, we release base, instruct and aligned versions on top of SaulLM-medium and SaulLM-large under the MIT License to facilitate reuse and collaborative research.",
        "authors": [
          "Pierre Colombo",
          "Telmo Pires",
          "Malik Boudiaf",
          "Rui Melo",
          "Gabriel Hautreux",
          "Etienne Malaboeuf",
          "Johanne Charpentier",
          "Dominic Culver",
          "Michael Desa"
        ],
        "decision": null,
        "pdate": 1727288078871,
        "cdate": 1715753562292
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "NLUYZ4ZqNq",
          "forum": "NLUYZ4ZqNq",
          "number": 14773,
          "title": "SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain",
          "normalized_title": "saullm 54b saullm 141b scaling up domain adaptation for the legal domain",
          "abstract": "In this paper, we introduce SaulLM-medium and SaulLM-large, two large language models (LLMs) families tailored for the legal sector. These models, which feature architectures of 54 billion and 140 billion parameters, respectively, are based on the Mixtral architecture. The development of SaulLM-54B and SaulLM-140B is guided by large-scale domain adaptation,  divided into strategies: (1) the exploitation of continued pretaining involving a legal corpus that includes over $400$ billion tokens, (2) the implementation of a specialized legal instruction-following protocol, and (3) the alignment of model outputs with human preferences in legal interpretations. The integration of synthetically generated data in the second and third steps enhances the models' capabilities in interpreting and processing legal texts, effectively reaching state-of-the-art performance and outperforming all previous open-source models on LegalBench Instruct. This research thoroughly explores the trade-offs involved in domain-specific adaptation at this scale, offering insights that may inform future studies on domain adaptation using strong decoder models. Building upon SaulLM-7B, this study refines the approach to produce an LLM better equipped for legal tasks and domains. Additionally, we release base, instruct and aligned versions on top of SaulLM-medium and SaulLM-large under the MIT License to facilitate reuse and collaborative research.",
          "authors": [
            "Pierre Colombo",
            "Telmo Pires",
            "Malik Boudiaf",
            "Rui Melo",
            "Gabriel Hautreux",
            "Etienne Malaboeuf",
            "Johanne Charpentier",
            "Dominic Culver",
            "Michael Desa"
          ],
          "decision": null,
          "pdate": 1727288078871,
          "cdate": 1715753562292
        }
      ]
    }
  },
  {
    "paper_id": 49215,
    "title": "Distilling Reinforcement Learning Algorithms for In-Context Model-Based Planning",
    "abstract": "Recent studies have shown that Transformers can perform in-context reinforcement learning (RL) by imitating existing RL algorithms, enabling sample-efficient adaptation to unseen tasks without parameter updates.\nHowever, these models also inherit the suboptimal behaviors of the RL algorithms they imitate.\nThis issue primarily arises due to the gradual update rule employed by those algorithms.\nModel-based planning offers a promising solution to this limitation by allowing the models to simulate potential outcomes before taking action, providing an additional mechanism to deviate from the suboptimal behavior.\nRather than learning a separate dynamics model, we propose Distillation for In-Context Planning (DICP), an in-context model-based RL framework where Transformers simultaneously learn environment dynamics and improve policy in-context.\nWe evaluate DICP across a range of discrete and continuous environments, including Darkroom variants and Meta-World.\nOur results show that DICP achieves state-of-the-art performance while requiring significantly fewer environment interactions than baselines, which include both model-free counterparts and existing meta-RL methods.\nThe code is available athttps://github.com/jaehyeon-son/dicp.",
    "introduction": "Since the introduction of Transformers(Vaswani et al.,2017), their versatility in handling diverse tasks has been widely recognized across various domains(Brown et al.,2020; Dosovitskiy et al.,2021; Bubeck et al.,2023).\nA key aspect of their success is in-context learning(Brown et al.,2020), which enables models to acquire knowledge rapidly without explicit parameter updates through gradient descent.\nRecently, this capability has been explored in reinforcement learning (RL)(Chen et al.,2021; Schulman et al.,2017; Lee et al.,2022; Reed et al.,2022), where acquiring skills in a sample-efficient manner is crucial.\nThis line of research naturally extends to meta-RL, which focuses on leveraging prior knowledge to quickly adapt to novel tasks.\n\n\nIn this context,Laskin et al. (2023)introduce Algorithm Distillation (AD), an in-context RL approach where Transformers sequentially model the entire learning histories of a specific RL algorithm across various tasks.\nThe goal is for the models to replicate the exploration-exploitation behaviors of the source RL algorithm, enabling them to tackle novel tasks purely in-context.\nBeyond replication,Laskin et al. (2023)show that this approach can enhance sample efficiency by bypassing intermediate learning steps or distilling histories from multiple actors.\nThis combination of an off-the-shelf RL algorithm with the in-context learning capability of Transformers has demonstrated strong potential for improving the adaptability of meta-RL(Lee et al.,2023a; Liu & Abbeel,2023; Huang et al.,2024; Zisman et al.,2024; Sinii et al.,2024).\n\n\n(a) Previous approaches(b) Distillation for In-Context Planning (Ours)Figure 1:Comparison between previous approaches(Laskin et al.,2023; Lee et al.,2023a; Huang et al.,2024)and our proposed Distillation for In-Context Planning (DICP).\nUnlike the previous approaches that directly predict actions without modeling dynamics, our approach leverages the in-context learned dynamics model for planning.\n\n\nHowever, prior in-context RL approaches have a notable limitation: they tend to replicate the suboptimal behaviors of the source algorithm.\nThe source RL algorithm updates its policy gradually through gradient descent, sometimes deliberately preventing abrupt changes(Schulman et al.,2015;2017).\nAs a result, it may take multiple iterations to fully integrate newly discovered information, leading to repeated suboptimal actions during this process.\nWithout a mechanism to deviate from the source algorithm’s behavior, existing in-context RL methods(Laskin et al.,2023; Huang et al.,2024; Sinii et al.,2024)inherit these inefficiencies.\n\n\nTo address this limitation, we introduceDistillation for In-Context Planning (DICP), an in-context model-based RL framework where the dynamics model is learned in-context alongside policy improvement.\nUnlike in-context policy improvement, in-context learning of the dynamics model does not inherit the suboptimal behaviors of the source algorithm, as modeling the environment’s dynamics is independent of the inefficiency.\nBy planning with this in-context learned dynamics model, our framework provides the agent with an additional mechanism to overcome the suboptimal behaviors of the source algorithm.\nFurthermore, simulating potential outcomes before executing actions allows for more deliberate decision-making based on predicted future returns.\nTo the best of our knowledge, ours is the first model-based approach on in-context RL using Transformers to imitate the source algorithm.\n\n\nTo demonstrate the effectiveness of our framework, we conduct experiments in both discrete and continuous environments, including Darkroom variants(Laskin et al.,2023)and Meta-World benchmark suite(Yu et al.,2019).\nThe results indicate that our approach significantly surpasses both in-context model-free counterparts (e.g., AD(Laskin et al.,2023), DPT(Lee et al.,2023a), and IDT(Huang et al.,2024)) and existing meta-RL methods (e.g., RL2(Duan et al.,2016), MAML(Finn et al.,2017), PEARL(Rakelly et al.,2019), MACAW(Mitchell et al.,2021), FOCAL(Li et al.,2021), BOReL(Dorfman et al.,2021), MoSS(Wang et al.,2023b), and IDAQ(Wang et al.,2023a)).\nNotably, our method achieves state-of-the-art results on the Meta-World ML1 and ML10 benchmarks, while requiring significantly fewer environmental interactions compared to the baselines.\n\n",
    "related_work": "RL as sequence modeling.With the advent of Transformers, which can learn from much larger datasets than what an agent can typically collect online, their application to offline RL(Levine et al.,2020)has gained prominence(Chen et al.,2021; Janner et al.,2021; Lee et al.,2022; Reed et al.,2022).\nDespite these advances,Laskin et al. (2023)point out a key limitation: these approaches struggle to improve their policy in-context through trial and error.\nThe primary reason is that they are designed to imitate the dataset policy, which makes them unsuitable for performing in-context RL on novel tasks.\nTo address this,Laskin et al. (2023)propose Algorithm Distillation (AD), an in-context RL approach where a Transformer is trained to distill learning histories from a source RL algorithm across diverse tasks.\nNotably, AD becomes effective when the context length of Transformers exceeds the episode horizon.\nInstead of imitating the source algorithm’s actions, Decision-Pretrained Transformer (DPT;Lee et al. (2023a)) is designed to predictoptimalactions.\nIn-context Decision Transformer (IDT;Huang et al. (2024)) implements a hierarchical approach to in-context RL, where the Transformer predicts high-level decisions that guide a sequence of actions, rather than predicting each action individually.\nIn addition,Sinii et al. (2024)propose an in-context RL approach for variable action spaces, whileZisman et al. (2024)synthesize learning histories by gradually denoising policies rather than directly executing a source algorithm.\n\n\nMeta-RL.Deep meta-RL began with online approaches, where RNNs are employed to learn RL algorithms that generalize across environments(Duan et al.,2016; Wang et al.,2016).\nIn parallel, gradient-based approaches(Finn et al.,2017; Nichol et al.,2018)aim to discover parameter initializations that rapidly adapt to new tasks.\nRecently,offlinemeta-RL has gained attention, leveraging pre-collected meta-training datasets to address the meta-RL problem(Mitchell et al.,2021; Dorfman et al.,2021; Yuan & Lu,2022; Wang et al.,2023a;b).\nVarious methods have been explored for this problem, including gradient-based(Mitchell et al.,2021), Bayesian(Dorfman et al.,2021), and contrastive learning approaches(Yuan & Lu,2022).\nFurthermore, recent in-context RL approaches(Laskin et al.,2023; Lee et al.,2023a; Huang et al.,2024; Sinii et al.,2024), including our own, fall within the category of offline meta-RL.\n\n\nModel-based meta-RL.In the prior research on model-based meta-RL,Zintgraf et al. (2020;2021); Dorfman et al. (2021)focus on learning to infer belief states about environment dynamics, whileWang et al. (2023b)aim to infer task representations.Wang et al. (2023a)quantify uncertainty using an ensemble of meta-training dynamics when facing novel meta-test tasks.Nagabandi et al. (2019); Pinon et al. (2022); Rimon et al. (2024)learn to construct dynamics models for planning.\nOur approach aligns with the latter category but stands out by integrating both the policy and the dynamics model within the same sequence model.\nThis contrasts with prior work that either relies on separate modules for modeling dynamics(Nagabandi et al.,2019; Rimon et al.,2024), or omits a policy network while depending solely on the dynamics model(Pinon et al.,2022).\nWe provide a more comprehensive overview of related works in App.C.\n\n",
    "origin_citations": [
      "Offline reinforcement learning: Tutorial, review, and perspectives on open problems.",
      "Decision transformer: Reinforcement learning via sequence modeling.",
      "Offline reinforcement learning as one big sequence modeling problem.",
      "Multi-game decision transformers.",
      "A generalist agent.",
      "In-context reinforcement learning with algorithm distillation.",
      "Supervised pretraining can learn in-context reinforcement learning.",
      "In-context decision transformer: Reinforcement learning via hierarchical chain-of-thought.",
      "In-context reinforcement learning for variable action spaces.",
      "Emergence of in-context reinforcement learning from noise distillation.",
      "RL2: Fast Reinforcement Learning via Slow Reinforcement Learning.",
      "Learning to reinforcement learn.",
      "Model-agnostic meta-learning for fast adaptation of deep networks.",
      "On first-order meta-learning algorithms.",
      "Offline meta-reinforcement learning with advantage weighting.",
      "Offline meta reinforcement learning - identifiability challenges and effective data collection strategies.",
      "Robust task representations for offline meta-reinforcement learning via contrastive learning.",
      "Offline meta reinforcement learning with in-distribution online adaptation.",
      "Meta-reinforcement learning based on self-supervised task representation learning.",
      "Varibad: A very good method for bayes-adaptive deep RL via meta-learning.",
      "Exploration in approximate hyper-state space for meta reinforcement learning.",
      "Learning to adapt in dynamic, real-world environments through meta-reinforcement learning.",
      "A model-based approach to meta-reinforcement learning: Transformers and tree search.",
      "Mamba: an effective world model approach for meta-reinforcement learning."
    ],
    "citations": [
      6170,
      196331,
      196027,
      145335,
      2564,
      334411,
      48090,
      1009123,
      348123,
      348546,
      8050,
      7753,
      1695,
      7739,
      160637,
      81207,
      638808,
      51455,
      71491,
      357290,
      3850,
      67515,
      287536
    ],
    "extract_rate": 0.9583333333333334,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 91,
      "Synthesis_Analysis": 78,
      "Final_Score": 84.69999999999999,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 91\nSynthesis_Analysis: 78\nFinal_Score: 84"
    },
    "overall_score": 85.41999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "BfUugGfBE5",
        "forum": "BfUugGfBE5",
        "number": 2042,
        "title": "Distilling Reinforcement Learning Algorithms for In-Context Model-Based Planning",
        "normalized_title": "distilling reinforcement learning algorithms for in context model based planning",
        "abstract": "Recent studies have shown that Transformers can perform in-context reinforcement learning (RL) by imitating existing RL algorithms, enabling sample-efficient adaptation to unseen tasks without parameter updates. However, these models also inherit the suboptimal behaviors of the RL algorithms they imitate. This issue primarily arises due to the gradual update rule employed by those algorithms. Model-based planning offers a promising solution to this limitation by allowing the models to simulate potential outcomes before taking action, providing an additional mechanism to deviate from the suboptimal behavior. Rather than learning a separate dynamics model, we propose Distillation for In-Context Planning (DICP), an in-context model-based RL framework where Transformers simultaneously learn environment dynamics and improve policy in-context. We evaluate DICP across a range of discrete and continuous environments, including Darkroom variants and Meta-World. Our results show that DICP achieves state-of-the-art performance while requiring significantly fewer environment interactions than baselines, which include both model-free counterparts and existing meta-RL methods.",
        "authors": [
          "Jaehyeon Son",
          "Soochan Lee",
          "Gunhee Kim"
        ],
        "decision": null,
        "pdate": 1737562253947,
        "cdate": 1726807905186
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "BfUugGfBE5",
          "forum": "BfUugGfBE5",
          "number": 2042,
          "title": "Distilling Reinforcement Learning Algorithms for In-Context Model-Based Planning",
          "normalized_title": "distilling reinforcement learning algorithms for in context model based planning",
          "abstract": "Recent studies have shown that Transformers can perform in-context reinforcement learning (RL) by imitating existing RL algorithms, enabling sample-efficient adaptation to unseen tasks without parameter updates. However, these models also inherit the suboptimal behaviors of the RL algorithms they imitate. This issue primarily arises due to the gradual update rule employed by those algorithms. Model-based planning offers a promising solution to this limitation by allowing the models to simulate potential outcomes before taking action, providing an additional mechanism to deviate from the suboptimal behavior. Rather than learning a separate dynamics model, we propose Distillation for In-Context Planning (DICP), an in-context model-based RL framework where Transformers simultaneously learn environment dynamics and improve policy in-context. We evaluate DICP across a range of discrete and continuous environments, including Darkroom variants and Meta-World. Our results show that DICP achieves state-of-the-art performance while requiring significantly fewer environment interactions than baselines, which include both model-free counterparts and existing meta-RL methods.",
          "authors": [
            "Jaehyeon Son",
            "Soochan Lee",
            "Gunhee Kim"
          ],
          "decision": null,
          "pdate": 1737562253947,
          "cdate": 1726807905186
        }
      ]
    }
  },
  {
    "paper_id": 47381,
    "title": "Communicating Activations Between Language Model Agents",
    "abstract": "Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate viaactivations; concretely, we pause an LMB𝐵Bitalic_B’s computation at an intermediate layer, combine its current activation with another LMA𝐴Aitalic_A’s intermediate activation via some functionf𝑓fitalic_f, then passf𝑓fitalic_f’s output into the next layer ofB𝐵Bitalic_Band continue the forward pass till decoding is complete. This approach scales up LMs on new tasks withzeroadditional parameters and data, and saves asubstantial amount of computeover natural language communication. We test our method with various functional formsf𝑓fitalic_fon two experimental setups—multi-player coordination games and reasoning benchmarks—and find that it achieves up to27.0%percent27.027.0\\%27.0 %improvement over natural language communication across datasets with<<<1/4141/41 / 4the compute, illustrating the superiority and robustness of activations as an alternative “language” for communication between LMs.",
    "introduction": "Language is for the purpose of communication. As large language models (LLMs) have been increasingly used to power autonomous, goal-driven agents capable of reasoning, tool usage, and adaptive decision-making(Yao et al.,2023; Xi et al.,2023; Wang et al.,2024; Ahn et al.,2022; Schick et al.,2023; Shen et al.,2023; Park et al.,2023; Nakano et al.,2022), communication between multiple cooperating agents has emerged as an intuitive approach to amplify the reasoning capabilities of LLMs(Wu et al.,2023). Explicit communication in natural language between multiple LLMs has been shown to encourage divergent thinking(Liang et al.,2023), improve factuality and reasoning(Du et al.,2023), enable integration of cross-domain knowledge(Sukhbaatar et al.,2024), and allow for modular composition of abilities in a complementary manner(Wu et al.,2023; Prasad et al.,2023).\n\n\nA critical problem with natural language communication, however, is that it incurs extremely high inference costs that scale quickly with the number of agents as well as length and number of messages(Du et al.,2023; Yang et al.,2023; Wu et al.,2023). Restricting LLM communication to natural language also raises the question: as LLMs are increasingly capable of handling larger, more complex tasks (sometimes with “super-human” ability)(Wei et al.,2022; Burns et al.,2023), might they communicate more effectively in representations of higher dimension than natural language? While using natural language as a communicative medium is appealing due to its interpretability, we claim that it may not be optimal for inter-LLM communication. Natural language generation uses only one token to represent the model’s belief over the entire vocabulary, which risks losing information embedded within the model output logits(Pham et al.,2024); furthermore, a model’s belief over the entire vocabulary is itself not always better (for communicative purposes) than the model’s (often richer) representation of the input in earlier layers. Indeed,Hernandez et al. (2024)find that by around the halfway point of an LM’s computation, it has developed “enriched entity representations” of the input, where entities in the prompt are populated with additional facts about that entity encoded in the model’s weights; but by the later layers these embeddings are transformed into a representation of the next word which leverages only parts of the previous, richer representations, when that full embedding would be quite useful for communication.\n\n\nMotivated by these concerns, this work outlines a simple technique whereby LLM agents communicate viaactivations, thus enabling more efficient (i.e., higher-entropy) communication at a fraction of the number of forward passes required at inference time. Concretely, we (1) pause a Transformer LMB𝐵Bitalic_B’s computation at intermediate layerj𝑗jitalic_jin the residual stream; (2) combine its post-layerj𝑗jitalic_jactivation with another LMA𝐴Aitalic_A’s post-layerk𝑘kitalic_kactivation via some functionf𝑓fitalic_f; and then (3) passf𝑓fitalic_f’s output into the next layerj+1𝑗1j+1italic_j + 1ofB𝐵Bitalic_Band continue its forward pass till decoding is complete. This approach scales up LLMs on new tasks by leveraging existing, frozen LLMs along withzerotask-specific parameters and data, applying to diverse domains and settings. Furthermore, in requiring only a partial forward pass throughA𝐴Aitalic_Aand one forward pass throughB𝐵Bitalic_B, this method saves asubstantial amount of computeover traditional natural language communication, which we quantify inSection 3.2.\n\n\nWe validate our method by testing this approach with various functional formsf𝑓fitalic_fon two experimental setups: two multi-player coordination games, whereB𝐵Bitalic_Bis asked to complete a task requiring information provided in a prompt toA𝐴Aitalic_A; and seven reasoning benchmarks spanning multiple domains: Biographies(Du et al.,2023), GSM8k(Cobbe et al.,2021), MMLU High School Psychology, MMLU Formal Logic, MMLU College Biology, MMLU Professional Law, and MMLU Public Relations(Hendrycks et al.,2021). Our activation communication protocol exhibits up to27.0%percent27.027.0\\%27.0 %improvement over natural language communication across these datasets, using<<<1/4141/41 / 4the compute. Critically, unlike prior work which test inter-LLM communication only on large-scale (>>>70707070B) models(Du et al.,2023; Liang et al.,2023), we find that our approach generalizes across a wide array of LLM suites and sizes, enabling even smaller LLMs to unlock the benefits of communication.\n\n\nIn summary, our contributions are two-fold:•We propose a novel inter-model communication protocol for LLM agents that is purely activation-based.•We perform comprehensive experiments to validate the improved performance of activation communication over traditional natural language communication. We also formally quantify our approach’s compute savings over natural language communication, illustrating the superiority and robustness of activations as an alternative “language” for communication between LMs.\n\n",
    "related_work": "Multi-agent communicationThe field of multi-agent communication has a long-standing history. Notably, prior works on emergent communication have showed that agents can autonomously\nevolve communication protocols when deployed in multi-agent environments that enable cooperative and competitive game-play(Sukhbaatar et al.,2016; Foerster et al.,2016; Lazaridou et al.,2017). However, recent experiments have demonstrated that learning meaningful languages from scratch, even with centralized training, remains difficult(Lowe et al.,2020; Chaabouni et al.,2019; Jaques et al.,2019).With the emergence of large pre-trained language models, allowing communication between LLMs in natural language has hence become a promising approach to enable coordination among multiple LLM agents(Li et al.,2023). Recent works have demonstrated that such conversations enable integration of cross-domain knowledge(Sukhbaatar et al.,2024), modular composition of abilities in a complementary manner(Wu et al.,2023), and improved task performance via splitting into subtasks(Prasad et al.,2023). Most notable is multiagent debate introduced byDu et al. (2023), where LLMs provide\ninitial responses and then make refinements by iteratively considering inputs from peers. While such methods have been shown to improve performance on various tasks over vanilla and majority-vote(Wang et al.,2023)style prompting, these experiments have only focused on large models (GPT-3.5/4,LLaMA2-70Band up), leaving the efficacy of debate on smaller, open-source models underexplored; our study addresses this gap by reimplementingDu et al. (2023)in experiments with smaller-scale (1−701701-701 - 70B) models. More crucially, debate and similar natural language communication methods areextremely computationally expensive, which this work addresses(Yang et al.,2023; Wu et al.,2023).Notably,Pham et al. (2024)propose CIPHER, which usesinput (tokenizer) embeddings(as opposed to activations) to enable multi-agent communication; specifically, CIPHER passes the average tokenizer embedding (weighted by the LLM’s next-token probabilities) between models. While(Pham et al.,2024)show this approach outperforms natural language debate, it (i) still faces substantial information loss relative to the modelactivationsand (ii) does not save compute, as the number of these “average embeddings” passed between models is the same as the number of tokens passed between models in natural language communication.\n\n\nActivation engineeringActivation engineering involves editing an LLM’s intermediate layer representations during a forward pass to create desired changes to output text(Li et al.,2024; Turner et al.,2023). Past work has explored extracting latent steering vectors from a frozen LLM to control quality and content of completions(Subramani et al.,2022), as well as using “direction” vectors (computed as the difference in activations between two prompts) that enable inference-time control over high-level properties of generations(Li et al.,2024; Turner et al.,2023). This work involves activation editing that is similar to such prior works at a high level, though for the purpose of communication between LLM agents.\n\n\nModel composition and graftingComposing expert models has been a recurring strategy to improve large models, with different methods imposing different restrictions on the types of base LLMs that can be combined. Mixture of Experts(Shazeer et al.,2017)requires that all experts are\ntrained simultaneously using the same data; Branch-Train-Mix(Sukhbaatar et al.,2024)trains a single base LM multiple times on different datasets, then learns a router on outputs. Crucially, these methods do not work when neither model can do the task at hand well (i.e., they solve the problem of choosing which of several outputs is best, not that of generating a high-quality output by recombining the disparate abilities of the various base LMs).Model grafting, in contrast, seeks to merge different models immediately prior to or at inference-time. Past works have explored this at the parameter level (e.g., task vector averaging as inIlharco et al. (2023), which requires that the base models be well aligned), probability distribution / token level as inShen et al. (2024)(which imposes few restrictions on the relationship between the base models, but by virtue of being token-based can result in cascading errors during decoding), and activation level (e.g., CALM(Bansal et al.,2024)which learns an attention layer on top of two models’ intermediate layer activations and thus enables broader integration of model abilities than token-level methods, but requires re-tuning of the attention mechanism for every model pair). In this work, we seek to unify CALM and other activation-level grafting techniques under a single framework, parameterized by the functionf𝑓fitalic_fused to combine activations; crucially, we explore simple forms off𝑓fitalic_f(e.g., sum, mean) that—unlikeBansal et al. (2024)—requirezero additional task-specific parameters and data, and are far more compute-efficient.\n\n",
    "origin_citations": [
      "Learning multiagent communication with backpropagation, 2016.",
      "Learning to communicate with deep multi-agent reinforcement learning, 2016.",
      "Multi-agent cooperation and the emergence of (natural) language, 2017.",
      "Multi-agent actor-critic for mixed cooperative-competitive environments, 2020.",
      "Word-order biases in deep-agent emergent communication.",
      "Social influence as intrinsic motivation for multi-agent deep reinforcement learning, 2019.",
      "Camel: Communicative agents for ”mind” exploration of large language model society, 2023.",
      "Branch-train-mix: Mixing expert llms into a mixture-of-experts llm, 2024.",
      "Autogen: Enabling next-gen llm applications via multi-agent conversation, 2023.",
      "Adapt: As-needed decomposition and planning with language models, 2023.",
      "Improving factuality and reasoning in language models through multiagent debate, 2023.",
      "Self-consistency improves chain of thought reasoning in language models, 2023.",
      "Auto-gpt for online decision making: Benchmarks and additional opinions, 2023.",
      "Let models speak ciphers: Multiagent debate through embeddings, 2024.",
      "Inference-time intervention: Eliciting truthful answers from a language model.",
      "Activation addition: Steering language models without optimization, 2023.",
      "Extracting latent steering vectors from pretrained language models, 2022.",
      "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017.",
      "Editing models with task arithmetic, 2023.",
      "Learning to decode collaboratively with multiple language models, 2024.",
      "Llm augmented llms: Expanding capabilities through composition, 2024."
    ],
    "citations": [
      279273,
      3450,
      7083,
      3442,
      374618,
      1349,
      288513,
      407255,
      294808,
      2454,
      2191,
      433,
      321431,
      2448,
      406160,
      377689,
      680,
      222300,
      290575,
      485320
    ],
    "extract_rate": 0.9523809523809523,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.36,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2025/Conference",
        "venue": "ICML",
        "year": "2025",
        "openreview_id": "W6RPXUUFic",
        "forum": "W6RPXUUFic",
        "number": 12901,
        "title": "Communicating Activations Between Language Model Agents",
        "normalized_title": "communicating activations between language model agents",
        "abstract": "Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate via *activations*; concretely, we pause an LM $B$'s computation at an intermediate layer, combine its current activation with another LM $A$'s intermediate activation via some function $f$, then pass $f$'s output into the next layer of $B$ and continue the forward pass till decoding is complete. This approach scales up LMs on new tasks with *zero* additional parameters and data, and saves a *substantial amount of compute* over natural language communication. We test our method with various functional forms $f$ on two experimental setups—multi-player coordination games and reasoning benchmarks—and find that it achieves up to $27$% improvement over natural language communication across datasets with $<$$1/4$ the compute, illustrating the superiority and robustness of activations as an alternative \"language\" for communication between LMs.",
        "authors": [
          "Vignav Ramesh",
          "Kenneth Li"
        ],
        "decision": null,
        "pdate": 1746105621267,
        "cdate": 1737666193657
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2025/Conference",
          "venue": "ICML",
          "year": "2025",
          "openreview_id": "W6RPXUUFic",
          "forum": "W6RPXUUFic",
          "number": 12901,
          "title": "Communicating Activations Between Language Model Agents",
          "normalized_title": "communicating activations between language model agents",
          "abstract": "Communication between multiple language model (LM) agents has been shown to scale up the reasoning ability of LMs. While natural language has been the dominant medium for inter-LM communication, it is not obvious this should be the standard: not only does natural language communication incur high inference costs that scale quickly with the number of both agents and messages, but also the decoding process abstracts away too much rich information that could be otherwise accessed from the internal activations. In this work, we propose a simple technique whereby LMs communicate via *activations*; concretely, we pause an LM $B$'s computation at an intermediate layer, combine its current activation with another LM $A$'s intermediate activation via some function $f$, then pass $f$'s output into the next layer of $B$ and continue the forward pass till decoding is complete. This approach scales up LMs on new tasks with *zero* additional parameters and data, and saves a *substantial amount of compute* over natural language communication. We test our method with various functional forms $f$ on two experimental setups—multi-player coordination games and reasoning benchmarks—and find that it achieves up to $27$% improvement over natural language communication across datasets with $<$$1/4$ the compute, illustrating the superiority and robustness of activations as an alternative \"language\" for communication between LMs.",
          "authors": [
            "Vignav Ramesh",
            "Kenneth Li"
          ],
          "decision": null,
          "pdate": 1746105621267,
          "cdate": 1737666193657
        }
      ]
    }
  },
  {
    "paper_id": 51924,
    "title": "Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection",
    "abstract": "In pursuit of detecting unstinted objects that extend beyond predefined categories, prior arts of open-vocabulary object detection (OVD) typically resort to pretrained vision-language models (VLMs) for base-to-novel category generalization. However, to mitigate the misalignment between upstream image-text pretraining and downstream region-level perception, additional supervisions are indispensable,e.g., image-text pairs or pseudo annotations generated via self-training strategies. In this work, we propose CCKT-Det trainedwithoutany extra supervision. The proposed framework constructs a cyclic and dynamic knowledge transfer from language queries and visual region features extracted from VLMs, which forces the detector to closely align with the visual-semantic space of VLMs. Specifically, 1) we prefilter and inject semantic priors to guide the learning of queries, and 2) introduce a regional contrastive loss to improve the awareness of queries on novel objects. CCKT-Det can consistently improve performance as the scale of VLMs increases, all while requiring the detector at a moderate level of computation overhead. Comprehensive experimental results demonstrate that our method achieves performance gain of+2.9%percent2.9+2.9\\%+ 2.9 %and+10.2%percent10.2+10.2\\%+ 10.2 %AP50subscriptAP50\\text{AP}_{50}AP start_POSTSUBSCRIPT 50 end_POSTSUBSCRIPTover previous state-of-the-arts on the challenging COCO benchmark, bothwithoutandwitha stronger teacher model. The code is provided athttps://github.com/ZCHUHan/CCKT-Det.",
    "introduction": "Object detection, a fundamental perception task in computer vision, entails both localization and classification of objects on given images. The last decade has witnessed great success in object detection including advanced architectures(Ren,2015; Carion et al.,2020), tailored loss objectives(Lin et al.,2017; Rezatofighi et al.,2019), and large-scale datasets(Gupta et al.,2019; Shao et al.,2019). However, canonical close-set detectors are constrained by small-scale and predefined categories, rendering them less capable of detecting new concepts(Bansal et al.,2018; Gu et al.,2021). To deploy detectors in real-world scenarios with countless concepts, many of which have no annotations in the training set, open-vocabulary object detection is formulated to facilitate the recognition of novel objects(Zareian et al.,2021).\n\n\nDrawing on the impressive image-level zero-shot learning capabilities of vision-language models like CLIP(Radford et al.,2021), recent initiatives replace the learnable classifier weights in traditional detectors with frozen text embeddings generated from VLMs text encoder that takes as input template prompts filled with category names. To address the challenges posed by the misalignment between image-text pre-training and region-level perception, as shown in Figure1, existing methods heavily rely on additional data such as: 1) caption datasets,e.g., CC3M(Sharma et al.,2018)and COCO Caption(Chen et al.,2015), or 2) meticulously designed self-training strategies to generate pseudo annotations. These extra supervisions are typically coarse and noisy, leading to inaccurate region-word alignment. This raises the question of whether a moderate level of weak supervision is sufficient while effectively maximizing the knowledge transfer from VLMs and multi-modal large language models (MLLMs)(Wang et al.,2023).\n\n\nFigure 1:The evolution of extracting novel concepts in OVD models. Compared to existing methods by using extra captions in (a) or pseudo annotations & self-training strategies in (b), we propose leveraging semantic priors to reveal novel concepts and employing contrastive knowledge distillation paradigm in (c) to align the enriched teacher space with the region-aware student space.\n\n\nIn this paper, as illustrated in Figure2, we introduce a novel framework termed as CCKT-Detwithout any additional supervisionto optimize the pipeline for efficient knowledge distillation in OVD. We argue that a cyclic and dynamic knowledge transfer from VLMs can efficiently mimick their well-aligned visual-semantic space and close the gap with methods that leverage extra data. Concretely, we first construct language queries by integrating semantic priors into the input of the detector,i.e., object queries, which makes the model novel-concepts aware and accelerates model convergence (refSec.3.2). To leverage the enriched semantics, we further propose a regional contrastive knowledge distillation loss that aligns the region-level visual features of queries from the student model with that from the teacher model (refSec.3.3). The semantic priors and regional contrastive knowledge distillation loss together forma cross-modal cycle with detector in between that can dynamically integrate semantic and visual knowledge from VLMs. The cycle forces detector to tightly align with VLMs as breaking either visual or semantic alignment with VLMs would fail this cyclic knowledge transfer as evidenced by our ablation study. Unlike existing methods that enhance performance by utilizing a more complex student backbone, our method exhibits consistent performance improvements as the strength and reliability of the teacher model increase without introducing any additional computational overhead during inference. We validate the effectiveness of the proposed CCKT-Det across the COCO(Lin et al.,2014), LVIS(Gupta et al.,2019), and Objects365(Shao et al.,2019)transfer benchmarks. On the OV-COCO(Zareian et al.,2021)benchmark, CCKT-Det enhances the AP50for novel categories by 2.9% compared to previous methods. This enhancement consistently scales with a stronger teacher model, culminating in a new state-of-the-art accuracy of 46.0% AP50on novel classes, all without the need for additional training data. On OV-LVIS(Gupta et al.,2019), CCKT-Det attains a competitive average precision of 18.2% average AP on rare without the incorporation of additional training data. Furthermore, when the model trained on the LVIS dataset is directly evaluated on the COCO and Objects365 datasets, CCKT-Det++ yields a performance of 38.5% AP and 15.2% AP, narrowing the performance difference compared to supervised models on COCO (-17%) and Objects365 (-40%), respectively.\n\n\nThe main contributions are summarized as the following:\n\n\n•We propose semantic priors as guidance and regional contrastive knowledge distillation loss to effectively align with the visual-semantic space of VLMs.•With the proposed two components, CCKT-Det constructs a cross-modal cycle that the detector in between can dynamically choose to transfer knowledge from.•The performance steadily improves as the teacher VLMs and MLLMs scale, and the proposed method achieves state-of-the-art in OVD without relying on any additional training data.\n\n",
    "related_work": "Contrastive Representation Learning.Recent developments in self-supervised pretraining, particularly through contrastive learning, have highlighted its ability to reduce reliance on labeled datasets. These methods primarily focus on the extraction of robust visual representations on image-level tasks(He et al.,2020; Chen et al.,2020). While effective at capturing general features, such global representations face significant challenges when applied to dense perception tasks. To mitigate, region-level contrastive learning has been introduced with different variants, such as sliding windows(Xiao et al.,2021), object proposals(Wei et al.,2021; Hénaff et al.,2021), and random query patches(Dai et al.,2021)to capture local visual clues. These strategies transform and adapt representations from image-level to pixel- or region-level, enhancing spatial reasoning capability in dense perception tasks. At the core of these methods is the definition of regional-level pre-text tasks. In our work, we leverage contrastive learning to distill rich semantic knowledge from a teacher model into the regional representations of a student model, tightly integrated with Hungarian matching to establish regional-level pairs. The proposed method captures higher-order correlations and dependencies in representation space, offering a more diverse set of negative and positive samples for distillation, thereby alleviating the constraints of learning exclusively from positive pairs of base categories.\n\n\nVision-Language Models (VLMs) and Multimodal Large Language Models (MLLMs).VLMs and MLLMs have profoundly transformed the integration of linguistic information in visual recognition tasks. As a prominent example of VLMs, CLIP(Radford et al.,2021), successfully aligns which image goes with which text on billion-scale image-text pairs, demonstrating strong generalization on novel categories. Concurrently, MLLMs,e.g., BLIP-2(Li et al.,2023b), MiniGPT-4(Zhu et al.,2023), and Vicuna(Zheng et al.,2023)connect image encoders with large language models to interpret visual clues and respond to human instructions effectively. Pretraining on web-scale data has endowed VLMs and MLLMs with a remarkable capability across various downstream tasks and benchmarks(Li et al.,2023a; Liu et al.,2024; Zeng et al.,2024). In this study, we adapt VLMs and MLLMs as teachers to convey novel semantic knowledge to our proposed detector.\n\n\nOpen Vocabulary Object Detection (OVD).To transfer the rich image-level semantic knowledge in VLMs to region-level, many works(Zhu & Chen,2024)seek to utilize additional supervisions,e.g., region-word alignment on image-text pairs in a weakly supervised manner which is noisy, or pseudo annotations generated via a self-training strategy which needs to know novel categories during training. Another methodology distills from teacher VLMs to student detectors for semantic knowledge transfer. The pioneering work ViLD(Gu et al.,2021)distills region embeddings from teacher VLMs in an element-wise manner, rendering the cross-modal similarity comparison toward the well-aligned visual-semantic space of VLMs. Building upon ViLD, BARON(Wu et al.,2023a)distills from a single region to a bag-of-regions level, effectively mining the co-occurrence and compositional structure of visual concepts inherently captured by VLMs. In addition, the pretrained image encoder of VLMs can be directly leveraged as a detector backbone while adding detection heads on top of it. The backbone is either frozen(Kuo et al.,2022; Zhong et al.,2022), or fine-tuned(Kim et al.,2023b)on detection datasets. To address the distribution gap between image-level pretraining and region-level perception, miscellaneous techniques including prompt tuning(Du et al.,2022; Feng et al.,2022)have been proposed. In particular, CORA(Wu et al.,2023c)utilizes region prompting to refine CLIP image encoder for improved localization. DITO(Kim et al.,2024)introduces a shifted-window learning technique to the region-centric image-text pretraining.\n\n",
    "origin_citations": [
      "Momentum contrast for unsupervised visual representation learning.",
      "Improved baselines with momentum contrastive learning.",
      "Region similarity representation learning.",
      "Aligning pretraining for detection via object-level contrastive learning.",
      "Efficient visual pretraining with contrastive detection.",
      "Up-detr: Unsupervised pre-training for object detection with transformers.",
      "Learning transferable visual models from natural language supervision.",
      "Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.",
      "Minigpt-4: Enhancing vision-language understanding with advanced large language models.",
      "Judging llm-as-a-judge with mt-bench and chatbot arena, 2023.",
      "Seed-bench: Benchmarking multimodal llms with generative comprehension.",
      "Visual instruction tuning.",
      "What matters in training a gpt4-style language model with multimodal inputs?",
      "A survey on open-vocabulary detection and segmentation: Past, present, and future.",
      "Open-vocabulary object detection via vision and language knowledge distillation.",
      "Aligning bag of regions for open-vocabulary object detection.",
      "F-vlm: Open-vocabulary object detection upon frozen vision and language models.",
      "Regionclip: Region-based language-image pretraining.",
      "Detection-oriented image-text pretraining for open-vocabulary detection.",
      "Learning to prompt for open-vocabulary object detection with vision-language model.",
      "Promptdet: Towards open-vocabulary detection using uncurated images.",
      "Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching.",
      "Region-centric image-language pretraining for open-vocabulary detection, 2024."
    ],
    "citations": [
      122,
      133,
      507524,
      195711,
      508417,
      432980,
      1639,
      212324,
      2243,
      2205,
      3522,
      2238,
      186958,
      470365,
      5325,
      225488,
      1088728,
      537669,
      477224,
      500137,
      499400,
      494302,
      906367
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 87,
      "Final_Score": 84.19999999999999,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 87\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 75,
      "Final_Score": 86.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 95\nSynthesis_Analysis: 75\nFinal_Score: 85"
    },
    "overall_score": 85.28,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "JU9oHs7ivN",
        "forum": "JU9oHs7ivN",
        "number": 4226,
        "title": "Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection",
        "normalized_title": "cyclic contrastive knowledge transfer for open vocabulary object detection",
        "abstract": "In pursuit of detecting unstinted objects that extend beyond predefined categories, prior arts of open-vocabulary object detection (OVD) typically resort to pretrained vision-language models (VLMs) for base-to-novel category generalization. However, to mitigate the misalignment between upstream image-text pretraining and downstream region-level perception, additional supervisions are indispensable, e.g., image-text pairs or pseudo annotations generated via self-training strategies. In this work, we propose CCKT-Det trained without any extra supervision. The proposed framework constructs a cyclic and dynamic knowledge transfer from language queries and visual region features extracted from VLMs, which forces the detector to closely align with the visual-semantic space of VLMs. Specifically, 1) we prefilter and inject semantic priors to guide the learning of queries, and 2) introduce a regional contrastive loss to improve the awareness of queries on novel objects. CCKT-Det can consistently improve performance as the scale of VLMs increases, all while requiring the detector at a moderate level of computation overhead. Comprehensive experimental results demonstrate that our method achieves performance gain of +2.9% and +10.2% AP_{50} over previous state-of-the-arts on the challenging COCO benchmark, both without and with a stronger teacher model.",
        "authors": [
          "Chuhan ZHANG",
          "Chaoyang Zhu",
          "Pingcheng Dong",
          "Long Chen",
          "Dong Zhang"
        ],
        "decision": null,
        "pdate": 1737562352698,
        "cdate": 1727246284466
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "JU9oHs7ivN",
          "forum": "JU9oHs7ivN",
          "number": 4226,
          "title": "Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection",
          "normalized_title": "cyclic contrastive knowledge transfer for open vocabulary object detection",
          "abstract": "In pursuit of detecting unstinted objects that extend beyond predefined categories, prior arts of open-vocabulary object detection (OVD) typically resort to pretrained vision-language models (VLMs) for base-to-novel category generalization. However, to mitigate the misalignment between upstream image-text pretraining and downstream region-level perception, additional supervisions are indispensable, e.g., image-text pairs or pseudo annotations generated via self-training strategies. In this work, we propose CCKT-Det trained without any extra supervision. The proposed framework constructs a cyclic and dynamic knowledge transfer from language queries and visual region features extracted from VLMs, which forces the detector to closely align with the visual-semantic space of VLMs. Specifically, 1) we prefilter and inject semantic priors to guide the learning of queries, and 2) introduce a regional contrastive loss to improve the awareness of queries on novel objects. CCKT-Det can consistently improve performance as the scale of VLMs increases, all while requiring the detector at a moderate level of computation overhead. Comprehensive experimental results demonstrate that our method achieves performance gain of +2.9% and +10.2% AP_{50} over previous state-of-the-arts on the challenging COCO benchmark, both without and with a stronger teacher model.",
          "authors": [
            "Chuhan ZHANG",
            "Chaoyang Zhu",
            "Pingcheng Dong",
            "Long Chen",
            "Dong Zhang"
          ],
          "decision": null,
          "pdate": 1737562352698,
          "cdate": 1727246284466
        }
      ]
    }
  },
  {
    "paper_id": 16251,
    "title": "Neural Language of Thought Models",
    "abstract": "The Language of Thought Hypothesis suggests that human cognition operates on a structured, language-like system of mental representations. While neural language models can naturally benefit from the compositional structure inherently and explicitly expressed in language data, learning such representations from non-linguistic general observations, like images, remains a challenge. In this work, we introduce the Neural Language of Thought Model (NLoTM), a novel approach for unsupervised learning of LoTH-inspired representation and generation. NLoTM comprises two key components: (1) the Semantic Vector-Quantized Variational Autoencoder, which learns hierarchical, composable discrete representations aligned with objects and their properties, and (2) the Autoregressive LoT Prior, an autoregressive transformer that learns to generate semantic concept tokens compositionally, capturing the underlying data distribution. We evaluate NLoTM on several 2D and 3D image datasets, demonstrating superior performance in downstream tasks, out-of-distribution generalization, and image generation quality compared to patch-based VQ-VAE and continuous object-centric representations. Our work presents a significant step towards creating neural networks exhibiting more human-like understanding by developing LoT-like representations and offers insights into the intersection of cognitive science and machine learning.",
    "introduction": "The Language of Thought Hypothesis (LoTH)(Fodor et al.,1975)suggests that human cognition is based on a structured, language-like system of mental representations, often referred to as “Mentalese”. Mentalese comprises word-like units that form sentence-like structures, which convey meaning. The meaning of these mental “sentences” is systematically determined by the meanings of their constituent “words” and their specific arrangement. From a computational viewpoint, while neural language models(Bengio et al.,2000; Brown et al.,2020; Bommasani et al.,2021)can benefit from the compositional and symbolic structure inherently expressed in the language data they are trained on, it remains unclear how we can learn such LoT-like structure from non-linguistic general observations, such as images, videos, and audio signals. The significance of this ability is further highlighted by the fact that infants learn these structures from observing objects and events before they acquire language skills(Spelke,2022).\n\n\nHow can we create neural networks that learn to develop such language of thought representations in an unsupervised way?To address this, we outline the following three properties as the desired characteristics of a neural language of thought model.\n\n\nFirst, when perceiving a visual scene, humans do not simply represent it as a monolithic vector of features. Instead, we view the scene structurally and semantically, recognizing it as acomposition of meaningful componentssuch as objects and their attributes, including shape, color, and position(Palmer,1977; Singer,2007; Spelke & Kinzler,2007). Our observation here is that in line with the LoTH, these visual attributes can be likened to words, objects to sentences, and the scene to a paragraph. Recent works, particularly those focused on object-centric representations(Greff et al.,2020), have demonstrated that this structural decomposition facilitates the benefits associated with the LoTH such as relational reasoning(Wu et al.,2021; Yoon et al.,2023; Webb et al.,2023a;b)and out-of-distribution generalization(Dittadi et al.,2022; Yoon et al.,2023)due to increased compositional generalization.\n\n\nMoreover, these structured and semantic representations can be categorized and conceptualized, resulting insymbol-like discrete concept abstraction. Such an ability is critical for organizing and comprehending the complexity of the environment, e.g., via language, as well as for implementing modularity(Andreas et al.,2016)or symbolic reasoning(Lake et al.,2016). Such discrete representations are also useful to leverage powerful generative models like autoregressive transformers(Vaswani et al.,2017). One of the most popular models for discrete representation learning is VQ-VAE(van den Oord et al.,2017). It has been shown to be beneficial for image generation(Razavi et al.,2019; Esser et al.,2021)and probability density modeling(Van den Oord et al.,2016). However, VQ-VAE and its variants, such as dVAE(Ramesh et al.,2021; Singh et al.,2022a)and VQ-GAN(Esser et al.,2021),\nrepresent a scene as a grid of small patches, lacking the capability to capture the scene’s holistic structure and semantics.\n\n\nBesides, the ability tocompositionally and probabilistically generate samplesthat adhere to the distribution of prior beliefs, constructed from observation data, is crucial for endowing AI with the capabilities to imagine and simulate. These abilities are essential for tasks such as planning(Mattar & Lengyel,2022; Hafner et al.,2019)and reasoning. However, this ability, related to probabilistic Language of Thought (PLoT)(Goodman et al.,2015)and productivity, is supported by only a certain class of representation learning models. While models like Slot Attention(Locatello et al.,2020)and SysBinder(Singh et al.,2023)offer structured, object-centric compositional representations, in their original form it is unclear how they support density-based sampling. In contrast, VAE-based models support this ability to sample from a prior distribution. However, they either do not provide object-centric structures or are limited to patch-based discrete abstractions (VQ-VAE).\n\n\nTable 1:Desiderata for Neural Language of Thought Models and Related ModelsVAEVQ-VAESlot AttentionSysBinderNLoTM (Ours)Compositionality(Semantic Scene Decomposition)Factor✘ObjectObject & FactorObject & FactorSymbolic(Discrete Concept Abstraction)✘✓(Patch Concept)✘✘✓(Semantic Concept)Productivity(Probabilistic Compositional Generation)✓✓(Patch Stitching)✘✘✓(Semantic Composition)\n\n\nIn this work, we present the Neural Language of Thought Model (NLoTM), the first model that satisfies the aforementioned criteria summarized in Table1.\nNLoTM comprises two novel components: the Semantic Vector-Quantized (SVQ) Variational Autoencoder and the Autoregressive LoT Prior (ALP). SVQ achieves discrete semantic decomposition of a scene by learning hierarchical, composable factors that closely align with the objects and the properties of objects in visual scenes, similar to the role of words and sentences in LoTH. ALP, analogous to language models trained on text tokens, is an autoregressive transformer trained to learn a probabilistic generation of semantic concept tokens. However, unlike VQ-VAE which stitches a grid of patches, ALP composes semantic concepts such as objects and their attributes to represent a scene.\n\n\nIn our experiments, we demonstrate the following practical benefits of our method.\nFirst, we find that for multi-object scenes, NLoTM is able to model the prior distribution better than patch-based methods, as measured by the quality of the samples generated.\nSecond, we find that NLoTM representations outperform patch-based VQ representations on downstream tasks that require knowledge of the different properties of the objects in the scene.\nWe also find evidence that NLoTM representations can generalize better to out-of-distribution tasks compared to patch-based VQ representations and SysBinder continuous representations.\nLastly, we show that despite introducing a discrete bottleneck, NLoTM can work on the challenging CLEVRTex(Karazija et al.,2021)dataset, one of the most complex datasets used in recent unsupervised object-centric representation learning models.\n\n\nOur contributions are as follows: First, we introduce NLoTM, the first neural network model implementing the LoTH for unstructured observations. Second, we propose the SVQ model to obtain object-centric semantic neural discrete representations. Third, we propose the ALP to model the probabilistic prior capable of capturing the underlying data distribution and compositional generation of new samples. Lastly, we evaluate our model on several 2D and 3D datasets including the challenging CLEVRTex dataset, showing superior downstream task performance and image generation quality.\n\n",
    "related_work": "2.1Vector-Quantized Variational Autoencoder (VQ-VAE)The VQ-VAE(van den Oord et al.,2017)is a model that learns to compress high-dimensional data into a discretized latent space.\nThe latent space is maintained by a codebook of prototype vectorse∈ℝK×desuperscriptℝ𝐾𝑑\\textbf{e}\\in\\mathbb{R}^{K\\times d}e ∈ blackboard_R start_POSTSUPERSCRIPT italic_K × italic_d end_POSTSUPERSCRIPTwhereK𝐾Kitalic_Kis the size of the codebook andd𝑑ditalic_dis the dimensionality of each prototype vector.\nAn inputxis first passed through encoderE⁢(x)𝐸xE(\\textbf{x})italic_E ( x )to obtain latentsze∈ℝdsubscriptz𝑒superscriptℝ𝑑\\textbf{z}_{e}\\in\\mathbb{R}^{d}z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_d end_POSTSUPERSCRIPT.\nA nearest-neighbor lookup betweenzesubscriptz𝑒\\textbf{z}_{e}z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPTand each of the prototype vectors in the codebook yields a quantized representationzq=Quantize⁢(ze)=eksubscriptz𝑞Quantizesubscriptz𝑒subscripte𝑘\\textbf{z}_{q}=\\mathrm{Quantize}(\\textbf{z}_{e})=\\textbf{e}_{k}z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT = roman_Quantize ( z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT ) = e start_POSTSUBSCRIPT italic_k end_POSTSUBSCRIPTwherek=arg⁢minj⁢‖ze−ej‖2𝑘subscriptargmin𝑗subscriptnormsubscriptz𝑒subscripte𝑗2k=\\operatorname*{arg\\,min}_{j}||\\textbf{z}_{e}-\\textbf{e}_{j}||_{2}italic_k = start_OPERATOR roman_arg roman_min end_OPERATOR start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT | | z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT - e start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT | | start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT.\nThe decoderD𝐷Ditalic_Dthen useszqsubscriptz𝑞\\textbf{z}_{q}z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPTto reconstruct the input:x^=D⁢(zq)^xDsubscriptz𝑞\\hat{\\textbf{x}}=\\text{D}(\\textbf{z}_{q})over^ start_ARG x end_ARG = D ( z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT ).\nThe model is trained with the following loss:ℒ=‖x−x^‖22⏟Reconstruction+‖s⁢g⁢[ze]−zq‖22⏟Codebook+β⁢‖ze−s⁢g⁢[zq]‖22⏟Commitment.ℒsubscript⏟superscriptsubscriptnormx^x22Reconstructionsubscript⏟superscriptsubscriptnorm𝑠𝑔delimited-[]subscriptz𝑒subscriptz𝑞22Codebook𝛽subscript⏟superscriptsubscriptnormsubscriptz𝑒𝑠𝑔delimited-[]subscriptz𝑞22Commitment\\mathcal{L}=\\underbrace{||\\textbf{x}-\\hat{\\textbf{x}}||_{2}^{2}}_{\\text{%\nReconstruction}}+\\underbrace{||sg[\\textbf{z}_{e}]-\\textbf{z}_{q}||_{2}^{2}}_{%\n\\text{Codebook}}+\\beta\\underbrace{||\\textbf{z}_{e}-sg[\\textbf{z}_{q}]||_{2}^{2%\n}}_{\\text{Commitment}}\\ .caligraphic_L = under⏟ start_ARG | | x - over^ start_ARG x end_ARG | | start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_ARG start_POSTSUBSCRIPT Reconstruction end_POSTSUBSCRIPT + under⏟ start_ARG | | italic_s italic_g [ z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT ] - z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT | | start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_ARG start_POSTSUBSCRIPT Codebook end_POSTSUBSCRIPT + italic_β under⏟ start_ARG | | z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT - italic_s italic_g [ z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPT ] | | start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT end_ARG start_POSTSUBSCRIPT Commitment end_POSTSUBSCRIPT .The first term is a reconstruction loss and is used to train the encoder and decoder.\nA straight-through estimator(Bengio et al.,2013)is used to estimate the gradients through the quantization step by copying the gradients fromzqsubscriptz𝑞\\textbf{z}_{q}z start_POSTSUBSCRIPT italic_q end_POSTSUBSCRIPTtozesubscriptz𝑒\\textbf{z}_{e}z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT.\nThe second term is the codebook loss which encourages the prototype vectors in the codebook to be close to the output of the encoder.\nThe third term, scaled by a constant hyperparameterβ𝛽\\betaitalic_β, is the commitment loss and helps to stabilize the training by encouraging the output of the encoder to not deviate too much from the chosen prototype vectors. Instead of the codebook loss, we use exponential moving average (EMA) updates on the codebook, which we found to speed up training in our experiments(Razavi et al.,2019; Dhariwal et al.,2020; Yan et al.,2021).When VQ-VAEs are applied to imagesx∈ℝH×W×Cxsuperscriptℝ𝐻𝑊𝐶\\textbf{x}\\in\\mathbb{R}^{H\\times W\\times C}x ∈ blackboard_R start_POSTSUPERSCRIPT italic_H × italic_W × italic_C end_POSTSUPERSCRIPT, the encoderE⁢(x)𝐸xE(\\textbf{x})italic_E ( x )is typically implemented as a convolution encoder, outputting a feature map of latentsze∈ℝHz×Wz×dsubscriptz𝑒superscriptℝsubscript𝐻𝑧subscript𝑊𝑧𝑑\\textbf{z}_{e}\\in\\mathbb{R}^{H_{z}\\times W_{z}\\times d}z start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_H start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT × italic_W start_POSTSUBSCRIPT italic_z end_POSTSUBSCRIPT × italic_d end_POSTSUPERSCRIPT.\nThis means that each latent corresponds to a local area represented by a convolutional feature cell and thus can only capture information in a local receptive field (Figure1a).\nHowever, images typically contain multiple objects, and the discrete factors underlying visual scenes typically correspond to different properties of the objects in the scene, such as shape, color, type, and so on.\nThe local patches from convolutional feature maps are inadequate to capture this rich structure.\n\n\n2.2Object-Centric RepresentationsThe goal of unsupervised object-centric representation learning is to decompose a scene into a set of representations each capturing a different object in the scene. It is shown that this structural decomposition, matching to the true factor structure of the world, facilitates some high-level cognition abilities such as relational reasoning(Wu et al.,2021; Yoon et al.,2023; Webb et al.,2023a;b)and out-of-distribution generalization(Dittadi et al.,2022; Yoon et al.,2023).\nWe build on top of Slot Attention(Locatello et al.,2020), a spatial attention-based object-centric representation method.Given an imagex∈ℝH×W×Cxsuperscriptℝ𝐻𝑊𝐶\\textbf{x}\\in\\mathbb{R}^{H\\times W\\times C}x ∈ blackboard_R start_POSTSUPERSCRIPT italic_H × italic_W × italic_C end_POSTSUPERSCRIPT, slot attention learns a set of slots,s={s1,…,sN}ssubscripts1…subscripts𝑁\\textbf{s}=\\{\\textbf{s}_{1},\\ldots,\\textbf{s}_{N}\\}s = { s start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , … , s start_POSTSUBSCRIPT italic_N end_POSTSUBSCRIPT }, wheresn∈ℝdssubscripts𝑛superscriptℝsubscript𝑑𝑠\\textbf{s}_{n}\\in\\mathbb{R}^{d_{s}}s start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_d start_POSTSUBSCRIPT italic_s end_POSTSUBSCRIPT end_POSTSUPERSCRIPTandN𝑁Nitalic_Nis the total number of slots.\nAn encoder is applied toxand, after adding a positional encoding, the result is flattened to anL𝐿Litalic_L-length input feature vectorF∈ℝL×dFFsuperscriptℝ𝐿subscript𝑑𝐹\\textbf{F}\\in\\mathbb{R}^{L\\times d_{F}}F ∈ blackboard_R start_POSTSUPERSCRIPT italic_L × italic_d start_POSTSUBSCRIPT italic_F end_POSTSUBSCRIPT end_POSTSUPERSCRIPT.\nThen, an iterative attention mechanism is used to spatially group the input featuresFto the slot representationss.\nFirst, the slots are randomly initialized from a Gaussian distribution with learned parameters.\nThen, in each iteration, the slots are used as queries in aninvertedversion of dot-product attention(Tsai et al.,2020)with the input featuresFas the keys and values.\nInstead of normalizing over the keys as is done in traditional dot-product attention, normalization is done over the queries (ie. slots).\nAdditionally, a weighted mean is used to aggregate the values instead of the normal weighted sum, which is shown to stabilize training.\nThe result is then used to update the slots with a per-slot GRU(Chung et al.,2014)followed by a per-slot residual MLP, both with shared parameters across the slots.The slot representations are then used in a decoder to reconstruct the image and the entire model is trained with an image reconstruction loss. The original formulation of slot attention used a spatial broadcast decoder(Watters et al.,2019b)to create masked images per slot which are then combined to form a final reconstructed image. Recently,(Singh et al.,2022a)proposed using a transformer decoder to reconstruct the image while attending to the slots with cross attention. This method was shown to scale to more complex scenes than the spatial broadcast decoder(Singh et al.,2022b)and is what we choose to use in our model.Figure 1:Comparison between VQ-VAE, Quantized Slots, and SVQ. (a) VQ-VAE quantizes the scene at a local patch level and may not capture the semantic structure of the scene. (b) Quantized Slots (QS) would quantize the scene at the slot level but require a separate code for every possible configuration of an object. (c) SVQ quantizes at the block level, representing each factor (such as color or shape) as a code. In this example, to represent all possible object configurations, SVQ requires only 10 codebook entries at the block level while QS requires 25.\n\n",
    "origin_citations": [
      "Neural discrete representation learning.",
      "Estimating or propagating gradients through stochastic neurons for conditional computation.",
      "Generating diverse high-fidelity images with VQ-VAE-2.",
      "Jukebox: A generative model for music.",
      "Videogpt: Video generation using VQ-VAE and transformers.",
      "Generative video transformer: Can objects be the words?",
      "An investigation into pre-training object-centric representations for reinforcement learning.",
      "Systematic visual reasoning through object-centric relational abstraction.",
      "Generalization and robustness implications in object-centric learning.",
      "Object-centric learning with slot attention, 2020.",
      "Capsules with inverted dot-product attention routing.",
      "Empirical evaluation of gated recurrent neural networks on sequence modeling.",
      "Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes.",
      "Illiterate DALL-E learns to compose.",
      "Simple unsupervised object-centric learning for complex and naturalistic videos."
    ],
    "citations": [
      3018,
      8375,
      7999,
      8727,
      789395,
      410408,
      490117,
      307268,
      373266,
      167570,
      2663,
      39,
      46484,
      514414,
      145949
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 92,
      "Final_Score": 87.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 92\nFinal_Score: 87"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 84,
      "Synthesis_Analysis": 85,
      "Final_Score": 84.0,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 84\nSynthesis_Analysis: 85\nFinal_Score: 84"
    },
    "overall_score": 85.24,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "HYyRwm367m",
        "forum": "HYyRwm367m",
        "number": 7823,
        "title": "Neural Language of Thought Models",
        "normalized_title": "neural language of thought models",
        "abstract": "The Language of Thought Hypothesis suggests that human cognition operates on a structured, language-like system of mental representations. While neural language models can naturally benefit from the compositional structure inherently and explicitly expressed in language data, learning such representations from non-linguistic general observations, like images, remains a challenge. In this work, we introduce the Neural Language of Thought Model (NLoTM), a novel approach for unsupervised learning of LoTH-inspired representation and generation. NLoTM comprises two key components: (1) the Semantic Vector-Quantized Variational Autoencoder, which learns hierarchical, composable discrete representations aligned with objects and their properties, and (2) the Autoregressive LoT Prior, an autoregressive transformer that learns to generate semantic concept tokens compositionally, capturing the underlying data distribution. We evaluate NLoTM on several 2D and 3D image datasets, demonstrating superior performance in downstream tasks, out-of-distribution generalization, and image generation quality compared to patch-based VQ-VAE and continuous object-centric representations. Our work presents a significant step towards creating neural networks exhibiting more human-like understanding by developing LoT-like representations and offers insights into the intersection of cognitive science and machine learning.",
        "authors": [
          "Yi-Fu Wu",
          "Minseung Lee",
          "Sungjin Ahn"
        ],
        "decision": null,
        "pdate": 1705411021330,
        "cdate": 1695487029042
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "HYyRwm367m",
          "forum": "HYyRwm367m",
          "number": 7823,
          "title": "Neural Language of Thought Models",
          "normalized_title": "neural language of thought models",
          "abstract": "The Language of Thought Hypothesis suggests that human cognition operates on a structured, language-like system of mental representations. While neural language models can naturally benefit from the compositional structure inherently and explicitly expressed in language data, learning such representations from non-linguistic general observations, like images, remains a challenge. In this work, we introduce the Neural Language of Thought Model (NLoTM), a novel approach for unsupervised learning of LoTH-inspired representation and generation. NLoTM comprises two key components: (1) the Semantic Vector-Quantized Variational Autoencoder, which learns hierarchical, composable discrete representations aligned with objects and their properties, and (2) the Autoregressive LoT Prior, an autoregressive transformer that learns to generate semantic concept tokens compositionally, capturing the underlying data distribution. We evaluate NLoTM on several 2D and 3D image datasets, demonstrating superior performance in downstream tasks, out-of-distribution generalization, and image generation quality compared to patch-based VQ-VAE and continuous object-centric representations. Our work presents a significant step towards creating neural networks exhibiting more human-like understanding by developing LoT-like representations and offers insights into the intersection of cognitive science and machine learning.",
          "authors": [
            "Yi-Fu Wu",
            "Minseung Lee",
            "Sungjin Ahn"
          ],
          "decision": null,
          "pdate": 1705411021330,
          "cdate": 1695487029042
        }
      ]
    }
  },
  {
    "paper_id": 24725,
    "title": "xRAG: Extreme Context Compression for Retrieval-augmented Generation with One Token",
    "abstract": "This paper introduces xRAG, a novel context compression method designed specifically for retrieval-augmented generation. xRAG redefines the use of document embeddings in dense retrieval—traditionally limited to retrieval purposes—by integrating them as features from the retrieval modality. Through a modality fusion approach, xRAG effectively merges these embeddings into the language model’s representation space, eliminating the need for their textual counterparts and achieving an extreme compression rate. In xRAG, the modality bridge is the only trainable component, while the retriever and language model remain frozen. This design choice allows for the reuse of offline-constructed document embeddings and preserves the plug-and-play nature of retrieval augmentation. Experimental results demonstrate that xRAG achieves an average improvement of over 10% across six knowledge-intensive tasks, compatible with various language model backbones, ranging from a dense 7B model to an 8x7B Mixture of Experts configuration. xRAG not only significantly outperforms previous context compression methods but also matches the performance of uncompressed models on several benchmarks, while reducing overall FLOPs by a factor of 3.53. This work pioneers new avenues in retrieval-augmented generation through multimodal fusion, potentially setting a groundwork for future developments in efficient and scalable retrieval systems.",
    "introduction": "Retrieval-Augmented Language Models (RALMs)[41,21,8,13,70]have shown exceptional performance in a variety of knowledge-intensive tasks. By retrieving domain-specific, long-tailed, and up-to-date knowledge from a non-parametric datastore, RALMs significantly extend the boundaries of parametric Large Language Models (LLMs). However, the integration of entire documents into prompts can significantly increase inference costs and may surpass the context limit of LLMs[28,79]. As illustrated in Figure1, while the inclusion of a relevant document enables the LLM to generate accurate responses, it does so at the expense of processing documents that expand the original query by more than tenfold.\n\n\nHow might we mitigate the costs associated with extended context while maintaining the benefits of retrieval augmentation? Recent research interest has converged on a promising direction: Context Compression. This concept is pursued through two primary strategies: soft-prompting methods, such as Gist[58], AutoCompressor[14], and ICAE[19], which compress the context into dense memory slots, and hard-prompting methods, such as LLMLingua[28]and RECOMP[79], where compression is applied on the surface form. These approaches, however, either require significant memory for storing LLM activations (e.g., 1.05 MB per token as reported by[58]) or suffer from relatively low compression rates. More critically, these methods overlook a crucial characteristic of RALMs: through large-scale contrastive learning with question-document pairs, modern dense retrieval systems already distill document content into a single high-dimensional embedding, andthis embedding reveals (almost) as much information as the text[57,43].\n\n\nIn this paper, we pioneer an innovative approach to retrieval augmentation and context compression through the lens of modality fusion. Drawing from multimodal research, where text-only language models are taught to \"perceive\" and \"listen,\" a pretrained modality encoder like CLIP[64]is typically used to extract modality features. These features are then integrated into language models using a modality fusion bridge[44,51,33]. Building on the conceptual overlap between the retriever encoder and modality encoder, we introduce xRAG. This model redefines document embeddings from dense retrieval—traditionally solely for retrieval purposes—as retrieval modality features. xRAG employs a modality fusion methodology to seamlessly integrate these embeddings into the language model’s representation space, thus obviating the need for textual counterparts and achieving significant context compression. In xRAG, the modality bridge is the only trainable component, while both the retriever and the LLM are kept frozen. This design decision facilitates the reuse of pre-constructed document embeddings and maintains the plug-and-play nature of retrieval augmentation—two essential factors for a functional RAG system.\n\n\nTo verify the effectiveness and versatility of our framework, we conducted comprehensive experiments with different LLM backbones, ranging from a dense 7B model to an 8x7B Mixture of Experts model. Our results reveal that adding just one document token could lead to over a 10% improvement across six knowledge-intensive tasks, significantly surpassing previous compression methods. xRAG also delivers results comparable to uncompressed models on several benchmarks. This is remarkable considering that the only trainable component constitutes less than 0.1% of the LLM’s parameters. In terms of efficiency, xRAG reduces total FLOPs by a factor of 3.53 compared to the uncompressed RAG model. We further provide detailed analyses of xRAG, examining various training strategies, data blends, and component selections for the retrieval system. We believe this research sets a strong foundation for the development of future efficient and scalable retrieval-augmented systems.\n\n",
    "related_work": "Retrieval-augmented GenerationEquipping a parametric language model with a non-parametric datastore has proven effective for a range of NLP tasks, including language modeling[36,56,87], open-domain question answering[24,41,70,86], domain adaptation[6]and machine translation[35,11], among others. Given the vast design space of this generation paradigm, numerous approaches with different focuses have been proposed. For instance, RETRO[8]and PlugLM[12]introduce architectural innovations for enhanced integration with the non-parametric datastore. REALM[21]pioneers an end-to-end approach for simultaneous optimization of the language model and retriever. REPLUG[70]and RA-DIT[50]improve retriever alignment using feedback from LLMs. DSP[37]and InteR[17]investigate complex interactions between the retriever and the language model. Selfmem[13]utilizes a reward model to refine retrieval and generation iteratively. Self-RAG[3]incorporates a self-reflection mechanism to enhance the quality and factuality of language model outputs. For a detailed overview, see[18,4,2]. Our contribution, xRAG, stands out by implementing a modality fusion approach to retrieval augmentation, creating an effective and efficient RAG system.\n\n\nContext CompressionContext compression, aimed at reducing the input length for LLMs while retaining essential information, has recently attracted substantial interest[46]. Gist[58]achieves a compression rate of up to 26x by modifying the attention mask and caching soft gist token activations. ICAE[19]AutoCompressor[14], and 500xCompressor[47]condense lengthy contexts into succinct, compact memory slots, which are directly utilizable by LLMs for diverse functions. LLMLingua[28,29,62]and CompAct[82]introduces a coarse-to-fine prompt compression technique based on perplexity scores and distilled token-level score. While these methods are generally applicable, others are tailored specifically for RAG systems, such as FilCo[76]and RECOMP[79]. A concurrent work directly employs passage embeddings for efficient listwise reranking[52]. For an in-depth comparison of these compression methods regarding memory efficiency, compression rates, and adaptability, refer to AppendixA.Figure 2:Overview of xRAG (a) and RAG (b). For a given query, RAG typically concatenates the retrieved document with the query, significantly extending the context length. In contrast, xRAG addresses this issue through modality fusion by directly projecting the document embedding into the LLM’s representation space. This allows for efficient retrieval-augmentation with the addition of only one token.\n\n",
    "origin_citations": [
      "Generalization through memorization: Nearest neighbor language models, 2020.",
      "Nonparametric masked language modeling, 2023.",
      "Training language models with memory augmentation, 2022.",
      "Leveraging passage retrieval with generative models for open domain question answering, 2021.",
      "Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021.",
      "Replug: Retrieval-augmented black-box language models, 2023.",
      "Funnelrag: A coarse-to-fine progressive retrieval paradigm for rag.",
      "Quebec automobile insurance question-answering with retrieval-augmented generation.",
      "Nearest neighbor machine translation, 2021.",
      "Neural machine translation with contrastive translation memories, 2022.",
      "Improving language models by retrieving from trillions of tokens, 2022.",
      "Decouple knowledge from parameters for plug-and-play language modeling, 2023.",
      "Realm: Retrieval-augmented language model pre-training, 2020.",
      "Ra-dit: Retrieval-augmented dual instruction tuning, 2023.",
      "Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp, 2023.",
      "Synergistic interplay between search and large language models for information retrieval, 2023.",
      "Lift yourself up: Retrieval-augmented text generation with self memory, 2023.",
      "Self-rag: Learning to retrieve, generate, and critique through self-reflection, 2023.",
      "Retrieval-augmented generation for large language models: A survey, 2024.",
      "Reliable, adaptable, and attributable language models with retrieval, 2024.",
      "Retrieval-based language models and applications.",
      "Prompt compression for large language models: A survey, 2024.",
      "Learning to compress prompts with gist tokens, 2024.",
      "In-context autoencoder for context compression in a large language model, 2023.",
      "Adapting language models to compress contexts, 2023.",
      "500xcompressor: Generalized prompt compression for large language models, 2024.",
      "Llmlingua: Compressing prompts for accelerated inference of large language models, 2023.",
      "Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression, 2023.",
      "Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression, 2024.",
      "Compact: Compressing retrieved documents actively for question answering, 2024.",
      "Learning to filter context for retrieval-augmented generation, 2023.",
      "Recomp: Improving retrieval-augmented lms with compression and selective augmentation, 2023.",
      "Leveraging passage embeddings for efficient listwise reranking with large language models, 2024."
    ],
    "citations": [
      452245,
      223475,
      5578,
      5579,
      7225,
      7228,
      729072,
      832037,
      357472,
      222704,
      365,
      260966,
      7790,
      377,
      381,
      114584,
      373,
      378,
      349125,
      290934,
      789825,
      54549,
      184970,
      259108,
      950614,
      321652,
      430,
      426360,
      729761,
      411,
      416,
      1087308
    ],
    "extract_rate": 0.9696969696969697,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 83,
      "Final_Score": 85.8,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 83\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 75,
      "Final_Score": 84.80000000000001,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 92\nSynthesis_Analysis: 75\nFinal_Score: 84"
    },
    "overall_score": 85.2,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "6pTlXqrO0p",
        "forum": "6pTlXqrO0p",
        "number": 9004,
        "title": "xRAG: Extreme Context Compression for Retrieval-augmented Generation with One Token",
        "normalized_title": "xrag extreme context compression for retrieval augmented generation with one token",
        "abstract": "This paper introduces xRAG, an innovative context compression method tailored for retrieval-augmented generation. xRAG reinterprets document embeddings in dense retrieval--traditionally used solely for retrieval--as features from the retrieval modality. By employing a modality fusion methodology, xRAG seamlessly integrates these embeddings into the language model representation space, effectively eliminating the need for their textual counterparts and achieving an extreme compression rate. \nIn xRAG, the only trainable component is the modality bridge, while both the retriever and the language model remain frozen. This design choice allows for the reuse of offline-constructed document embeddings and preserves the plug-and-play nature of retrieval augmentation. \nExperimental results demonstrate that xRAG achieves an average improvement of over 10% across six knowledge-intensive tasks, adaptable to various language model backbones, ranging from a dense 7B model to an 8x7B Mixture of Experts configuration. xRAG not only significantly outperforms previous context compression methods but also matches the performance of uncompressed models on several datasets, while reducing overall FLOPs by a factor of 3.53. Our work pioneers new directions in retrieval-augmented generation from the perspective of multimodality fusion, and we hope it lays the foundation for future efficient and scalable retrieval-augmented systems.",
        "authors": [
          "Xin Cheng",
          "Xun Wang",
          "Xingxing Zhang",
          "Tao Ge",
          "Si-Qing Chen",
          "Furu Wei",
          "Huishuai Zhang",
          "Dongyan Zhao"
        ],
        "decision": null,
        "pdate": 1727287897889,
        "cdate": 1715673016598
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "6pTlXqrO0p",
          "forum": "6pTlXqrO0p",
          "number": 9004,
          "title": "xRAG: Extreme Context Compression for Retrieval-augmented Generation with One Token",
          "normalized_title": "xrag extreme context compression for retrieval augmented generation with one token",
          "abstract": "This paper introduces xRAG, an innovative context compression method tailored for retrieval-augmented generation. xRAG reinterprets document embeddings in dense retrieval--traditionally used solely for retrieval--as features from the retrieval modality. By employing a modality fusion methodology, xRAG seamlessly integrates these embeddings into the language model representation space, effectively eliminating the need for their textual counterparts and achieving an extreme compression rate. \nIn xRAG, the only trainable component is the modality bridge, while both the retriever and the language model remain frozen. This design choice allows for the reuse of offline-constructed document embeddings and preserves the plug-and-play nature of retrieval augmentation. \nExperimental results demonstrate that xRAG achieves an average improvement of over 10% across six knowledge-intensive tasks, adaptable to various language model backbones, ranging from a dense 7B model to an 8x7B Mixture of Experts configuration. xRAG not only significantly outperforms previous context compression methods but also matches the performance of uncompressed models on several datasets, while reducing overall FLOPs by a factor of 3.53. Our work pioneers new directions in retrieval-augmented generation from the perspective of multimodality fusion, and we hope it lays the foundation for future efficient and scalable retrieval-augmented systems.",
          "authors": [
            "Xin Cheng",
            "Xun Wang",
            "Xingxing Zhang",
            "Tao Ge",
            "Si-Qing Chen",
            "Furu Wei",
            "Huishuai Zhang",
            "Dongyan Zhao"
          ],
          "decision": null,
          "pdate": 1727287897889,
          "cdate": 1715673016598
        }
      ]
    }
  },
  {
    "paper_id": 16982,
    "title": "Introspective Planning:\nAligning Robots’ Uncertainty with Inherent Task Ambiguity",
    "abstract": "Large language models (LLMs) exhibit advanced reasoning skills, enabling robots to comprehend natural language instructions and strategically plan high-level actions through proper grounding. However, LLM hallucination may result in robots confidently executing plans that are misaligned with user goals or even unsafe in critical scenarios. Additionally, inherent ambiguity in natural language instructions can introduce uncertainty into the LLM’s reasoning and planning processes.\nWe propose introspective planning, a systematic approach that align LLM’s uncertainty with the inherent ambiguity of the task.\nOur approach constructs a knowledge base containing introspective reasoning examples as post-hoc rationalizations of human-selected safe and compliant plans, which are retrieved during deployment. Evaluations on three tasks, including a newly introduced safe mobile manipulation benchmark, demonstrate that introspection substantially improves both compliance and safety over state-of-the-art LLM-based planning methods. Furthermore, we empirically show that introspective planning, in combination with conformal prediction, achieves tighter confidence bounds, maintaining statistical success guarantees while minimizing unnecessary user clarification requests. The webpage and code are accessible athttps://introplan.github.io.",
    "introduction": "Large Language Models (LLMs), when pre-trained on internet-scale text corpora, demonstrate emergent capabilities that extend far beyond mere text comprehension and generation as their scale increases[5,40]. Through prompting[9]and in-context learning[26], these models have shown remarkable adaptability ranging from answering complex questions and solving mathematical problems to generating computer code and engaging in sophisticated reasoning processes during inference[36]. Robots interacting with humans can leverage the capabilities of LLMs to interpret task instructions in natural language, employ common sense reasoning to understand their environment, and devise high-level action plans grounded in the capabilities and affordances of the robot[1,21].\n\n\nFigure 1:Illustration of the introspective planning pipeline.Knowledge base construction:The LLM generates knowledge entries based on human-provided instructions and valid options.Deployment:Upon receiving an instruction, the LLM formulates possible next steps, consults the knowledge base to retrieve the most relevant examples, and uses them as prompts for prediction.\n\n\nThe reliability of LLM outputs has direct implications downstream robotics tasks. Language models are prone to hallucinations[13], which cause models to generate plans that are at odds with common-sense knowledge, not executable by the robot, or incompatible with the environment constraints[48]. For example, if a human user asks a robot to bake some bread in a kitchen containing an oven and varied cookware,\nthe robot’s LLM may generate a decision to use a plastic tray without considering the risk of it melting. Furthermore, possible ambiguities in the user’s request can also introduce uncertainty into the LLM’s reasoning and planning[37]. In our example, while multiple containers are suitable for the stated task, biases inherited from training data may tilt action generation towards certain options. Therefore, the robot needs to calibrate its uncertainty quantification and seek further communication with users when ambiguities are identified.\n\n\nHuman beings assess their internal values and knowledge of their own abilities to guide domain-level reasoning processes: this is referred to as introspective reasoning[19]. In this paper, we observe that LLMs can leverage an analogous reasoning scheme to better assess underlying uncertainty when generating plans. We propose a novel method for constructing a knowledge base that utilizes LLMs to generate human-aligned introspective reasoning examples with minimal human input. During inference, this human-aligned knowledge guides LLMs to produce more reliable and interpretable plans. Unlike traditional Retrieval-Augmented Generation (RAG) approaches[10,20,29,43,51], which utilize open-source, off-the-shelf knowledge bases to enhance text generation, our approach retrieves few-shot introspective reasoning examples from the knowledge base. This enables LLMs to explicitly reason about uncertainties and formulate plans in a structured format. Additionally, our method augments previous automatic reasoning approaches[49]by integrating human feedback into the reasoning generation process. We have observed that introspective planning, when integrated with conformal prediction[2,3], refines the LLM’s uncertainty and achieves a tighter guarantee.\n\n\nStatement of contributions.To the best of our knowledge, this is the first work to integrate retrieval-augmented planning with conformal prediction, refining language agents’ uncertainty and reducing user queries while maintaining statistical guarantees. Key contributions are summarized as follows:•We propose a novelintrospective planningscheme that prompts language-enabled agents to proactively assess their own confidence regarding\ntask compliance and safety for multiple candidate plans, with a guaranteed probability\nthat the agent will either execute the actions desired by the user\nor ask an appropriate follow-up question to disambiguate the user’s intent.•We introduce a new, weakly supervised offline\nknowledge base construction method that\nguides the LLM to generate\nhuman-aligned introspective reasoning examples as post-hoc rationalizations of human-selected safe-and-compliant plans.•We create a new Safe Mobile Manipulation benchmark, which augments previous mobile manipulation datasets with safety-critical scenarios and introduces new metrics to evaluate a planner’s specification compliance, safety, and degree of conservativeness.\n\n",
    "related_work": "LLMs as reasoning engines.Through a process known as (zero-shot) chain-of-thought (CoT), LLMs can be prompted to generate multiple reasoning steps\nby instructing them to “think step by step” at inference time[18].\nThis method’s accuracy can be improved by including manually designed examples (few-shot CoT)[41].\nThe tree-of-thoughts approach[44]generalizes CoT by considering multiple reasoning paths.\nOur work is inspired by the retrieval augmentation mechanism in\nAuto-CoT[49]and Retrieval-Q-CoT[49], which first use zero-shot CoT to generate a diverse set of reasoning chains and then sample them at runtime as few-shot CoT examples. However, these pre-generated\nreasoning examples are sometimes incorrect due to LLM hallucination, leading to inference-time errors.\nAs demonstrated in Section5, our new knowledge base generation approach substantially addresses this issue by instead querying the LLM forpost-hoc rationalizationsconditioned on human-provided valid/invalid labels on candidate solutions.\n\n\nRetrieval-augmented generation.Retrieval-augmented generation (RAG) augments the input space of LMs with retrieved text passages, significantly improving performance on knowledge-intensive tasks[10,20,29]. Traditional RAG methods typically use open-source, off-the-shelf knowledge bases for text generation[16,33,25,47,43,51].\nConversely,\nour approach retrieves few-shot introspective reasoning examples from the knowledge base. This guides LLMs to explicitly reason about uncertainties and safety, formulating plan in a structured format, as shown inTab.13. In practice, we observe this strategy results in a more grounded reasoning process compared to conventional RAG, which relies on open-source knowledge bases. While existing RAG literature primarily addresses content hallucination, our approach aims to equip language agents with the capability to introspect and refine their own uncertainties. This emphasis allows for uncertainty-aware planning in robotics and achieves tighter statistical guarantees with conformal prediction.\n\n\nLLMs as planners.Emergent reasoning allows LLMs to break down a task into intermediate subgoals and generate actions as a sequence of plans[11]. Through prompting and in-context learning, LLMs can ground human instructions in natural language into executable robot actions conditioned on scene descriptions[21,35,1,12]. Recent works further enhance the reasoning and planning ability by iteratively refining actions through self-reflection during planning[45,24,28,6,34,39,23]. ReAct[45]and Reflexion[34]focus onmulti-step planningscenarios, in which robots can execute certain actions, observe the state feedback, and replan for correction. However, in safety-critical robotic applications, certain invalid actions can immediately lead to catastrophic safety failures that cannot be recovered from. Therefore, instead of relying on self-correction by trial and error, our method uses retrieval augmentation to guide the language agent to proactively reason about task compliance and safety at the planning stage. Additionally, recent work[17]has shown that LLMs cannot plan effectively through self-reflection alone but can do so when integrated with external verifiers, aligning with our method that employs LLMs to support planning by constructing an external knowledge base.\n\n\nQuantifying uncertainty in LLMs.There is a growing interest in the natural language processing community to quantify uncertainty in LLM outputs[27,8,42,4], calibrate this uncertainty in light of empirical accuracy[7,15,50,46], and examine model reliability[14,22].\nOur work is most closely related to the recently proposed KnowNo framework[31],\nwhich\ncasts\ntask-level planning as multiple choice question answering (MCQA)\nand uses conformal prediction to output a subset of LLM-generated candidate plans with a desired (marginal) probability of containing at least one valid course of action.\nUnfortunately, the statistical guarantees achieved by KnowNo come at the cost of frequent superfluous user queries (oroverasking, as defined in Section4and empirically quantified in Section5).\nIn contrast,\nour method introduces a new introspection-based approach to automatically align the robot’s uncertainty with the inherent task specification ambiguity before predicting a high-confidence subset of plans.\nThis uncertainty alignment step mitigates the need for conservativeness in the calibration stage and drastically reduces the resulting rate of overasking while maintaining the desired statistical success guarantees.\n\n",
    "origin_citations": [
      "Large language models are zero-shot reasoners.",
      "Chain-of-thought prompting elicits reasoning in large language models.",
      "Tree of thoughts: Deliberate problem solving with large language models.",
      "Automatic chain of thought prompting in large language models.",
      "Retrieval augmented language model pre-training.",
      "Retrieval-augmented generation for knowledge-intensive nlp tasks.",
      "In-context retrieval-augmented language models.",
      "Active retrieval augmented generation.",
      "Toolformer: Language models can teach themselves to use tools.",
      "When not to trust language models: Investigating effectiveness of parametric and non-parametric memories.",
      "Making retrieval-augmented language models robust to irrelevant context.",
      "Recomp: Improving retrieval-augmented lms with compression and selective augmentation.",
      "Language agent tree search unifies reasoning acting and planning in language models.",
      "Language models as zero-shot planners: Extracting actionable knowledge for embodied agents.",
      "Code as policies: Language model programs for embodied control.",
      "Progprompt: Generating situated robot task plans using large language models.",
      "Do as i can, not as i say: Grounding language in robotic affordances.",
      "Inner monologue: Embodied reasoning through planning with language models.",
      "React: Synergizing reasoning and acting in language models.",
      "Self-refine: Iterative refinement with self-feedback.",
      "Refiner: Reasoning feedback on intermediate representations.",
      "Prompting large language models with the socratic method.",
      "Reflexion: an autonomous agent with dynamic memory and self-reflection.",
      "Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents.",
      "Reflect: Summarizing robot experiences for failure explanation and correction.",
      "Llms can’t plan, but can help planning in llm-modulo frameworks.",
      "Analyzing uncertainty in neural machine translation.",
      "Unsupervised quality estimation for neural machine translation.",
      "Uncertainty quantification with pre-trained language models: A large-scale empirical analysis.",
      "Uncertainty in natural language generation: From theory to applications.",
      "Calibration of pre-trained transformers.",
      "How can we know when language models know? on the calibration of language models for question answering.",
      "Calibrate before use: Improving few-shot performance of language models.",
      "Do large language models know what they don’t know?",
      "Survey of hallucination in natural language generation.",
      "Truthfulqa: Measuring how models mimic human falsehoods.",
      "Robots that ask for help: Uncertainty alignment for large language model planners."
    ],
    "citations": [
      7094,
      1578,
      7086,
      2467,
      7225,
      729779,
      379,
      431,
      139706,
      415,
      416,
      322399,
      5587,
      3031,
      177711,
      7658,
      2997,
      2954,
      7654,
      20329,
      227690,
      1087923,
      491445,
      47779,
      101336,
      775932,
      60090,
      459841,
      467569,
      253954,
      111165,
      1594,
      2429,
      3111,
      2217,
      187219
    ],
    "extract_rate": 0.972972972972973,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 81,
      "Final_Score": 85.2,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 81\nFinal_Score: 85"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 85.08000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "4TlUE0ufiz",
        "forum": "4TlUE0ufiz",
        "number": 2620,
        "title": "Introspective Planning: Aligning Robots' Uncertainty with Inherent Task Ambiguity",
        "normalized_title": "introspective planning aligning robots uncertainty with inherent task ambiguity",
        "abstract": "Large language models (LLMs) exhibit advanced reasoning skills, enabling robots to comprehend natural language instructions and strategically plan high-level actions through proper grounding. However, LLM hallucination may result in robots confidently executing plans that are misaligned with user goals or even unsafe in critical scenarios. Additionally, inherent ambiguity in natural language instructions can introduce uncertainty into the LLM's reasoning and planning. We propose introspective planning, a systematic approach that guides LLMs to refine their own uncertainty in alignment with inherent task ambiguity. Our approach constructs a knowledge base containing introspective reasoning examples as post-hoc rationalizations of human-selected safe and compliant plans, which are retrieved during deployment. Evaluations on three tasks, including a new safe mobile manipulation benchmark, indicate that introspection substantially improves both compliance and safety over state-of-the-art LLM-based planning methods. Additionally, we empirically show that introspective planning, in combination with conformal prediction, achieves tighter confidence bounds, maintaining statistical success guarantees while minimizing unnecessary user clarification requests.",
        "authors": [
          "Kaiqu Liang",
          "Zixu Zhang",
          "Jaime Fernández Fisac"
        ],
        "decision": null,
        "pdate": 1727287695731,
        "cdate": 1715128809666
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "4TlUE0ufiz",
          "forum": "4TlUE0ufiz",
          "number": 2620,
          "title": "Introspective Planning: Aligning Robots' Uncertainty with Inherent Task Ambiguity",
          "normalized_title": "introspective planning aligning robots uncertainty with inherent task ambiguity",
          "abstract": "Large language models (LLMs) exhibit advanced reasoning skills, enabling robots to comprehend natural language instructions and strategically plan high-level actions through proper grounding. However, LLM hallucination may result in robots confidently executing plans that are misaligned with user goals or even unsafe in critical scenarios. Additionally, inherent ambiguity in natural language instructions can introduce uncertainty into the LLM's reasoning and planning. We propose introspective planning, a systematic approach that guides LLMs to refine their own uncertainty in alignment with inherent task ambiguity. Our approach constructs a knowledge base containing introspective reasoning examples as post-hoc rationalizations of human-selected safe and compliant plans, which are retrieved during deployment. Evaluations on three tasks, including a new safe mobile manipulation benchmark, indicate that introspection substantially improves both compliance and safety over state-of-the-art LLM-based planning methods. Additionally, we empirically show that introspective planning, in combination with conformal prediction, achieves tighter confidence bounds, maintaining statistical success guarantees while minimizing unnecessary user clarification requests.",
          "authors": [
            "Kaiqu Liang",
            "Zixu Zhang",
            "Jaime Fernández Fisac"
          ],
          "decision": null,
          "pdate": 1727287695731,
          "cdate": 1715128809666
        }
      ]
    }
  },
  {
    "paper_id": 22954,
    "title": "Graph as Point Set",
    "abstract": "Graph is a fundamental data structure to model interconnections between entities. Set, on the contrary, stores independent elements. To learn graph representations, current Graph Neural Networks (GNNs) primarily use message passing to encode the interconnections. In contrast, this paper introduces a novel graph-to-set conversion method that bijectively transforms interconnected nodes into a set of independent points and then uses a set encoder to learn the graph representation. This conversion method holds dual significance. Firstly, it enables using set encoders to learn from graphs, thereby significantly expanding the design space of GNNs. Secondly, for Transformer, a specific set encoder, we provide a novel and principled approach to inject graph information losslessly, different from all the heuristic structural/positional encoding methods adopted in previous graph transformers. To demonstrate the effectiveness of our approach, we introduce Point Set Transformer (PST), a transformer architecture that accepts a point set converted from a graph as input. Theoretically, PST exhibits superior expressivity for both short-range substructure counting and long-range shortest path distance tasks compared to existing GNNs. Extensive experiments further validate PST’s outstanding real-world performance. Besides Transformer, we also devise a Deepset-based set encoder, which achieves performance comparable to representative GNNs, affirming the versatility of our graph-to-set method.",
    "introduction": "Graph, composed of interconnected nodes, has a wide range of applications and has been extensively studied. In graph machine learning, a central focus is to effectively leverage node connections. Various architectures have arisen for graph tasks, exhibiting significant divergence in their approaches to utilizing adjacency information.\n\n\nTwo primary paradigms have evolved for encoding adjacency information. The first paradigm involves message passing between nodes via edges. Notable methods in this category include Message Passing Neural Network (MPNN)(Gilmer et al.,2017), a foundational framework for GNNs such as GCN(Kipf & Welling,2017), GIN(Xu et al.,2019a), and GraphSAGE(Hamilton et al.,2017). Subgraph-based GNNs(Zhang & Li,2021; Huang et al.,2023b; Bevilacqua et al.,2022; Qian et al.,2022; Frasca et al.,2022; Zhao et al.,2022; Zhang et al.,2023a)select subgraphs from the whole graph and run MPNN within each subgraph. These models aggregate messages from neighbors to update the central nodes’ representations. Additionally, Graph Transformers (GTs) integrate adjacency information into the attention matrix(Mialon et al.,2021; Kreuzer et al.,2021; Wu et al.,2021; Dwivedi & Bresson,2020; Ying et al.,2021; Shirzad et al.,2023)(note that some early GTs have options to not use adjacency matrix by using only positional encodings, but the performance is significantly worse(Dwivedi & Bresson,2020)). Some recent GTs even directly incorporate message-passing layers into their architectures(Rampásek et al.,2022; Kim et al.,2021). In summary, this paradigm relies on adjacency relationships to facilitate information exchange among nodes.\n\n\nThe second paradigm designs permutation-equivariant neural networks that directly take adjacency matrices as input. This category includes high-order Weisfeiler-Leman tests(Maron et al.,2019a), invariant graph networks(Maron et al.,2019b), and relational pooling(Chen et al.,2020). Additionally, various studies have explored manual feature extraction from the adjacency matrix, including random walk structural encoding(Dwivedi et al.,2022a; Li et al.,2020), Laplacian matrix eigenvectors(Wang et al.,2022; Lim et al.,2023; Huang et al.,2023a), and shortest path distances(Li et al.,2020). However, these approaches typically serve as data augmentation steps for other models, rather than constituting an independent paradigm.\n\n\nBoth paradigms heavily rely on adjacency information in graph encoding. In contrast, this paper explores whether we can give up adjacency matrix in graph models while achieving competitive performance. As shown in Figure1, our innovative graph-to-set method converts interconnected nodes into independent points, subsequently encoded by a set encoder like Transformer. Leveraging oursymmetric rank decomposition, we break down the augmented adjacency matrixA+D𝐴𝐷A+Ditalic_A + italic_DintoQ⁢QT𝑄superscript𝑄𝑇QQ^{T}italic_Q italic_Q start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT, whereinQ𝑄Qitalic_Qis constituted by column-full-rank rows, each denoting a nodecoordinate. This representation enables us to express the presence of edges as inner products of coordinate vectors (Qisubscript𝑄𝑖Q_{i}italic_Q start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTandQjsubscript𝑄𝑗Q_{j}italic_Q start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT). Consequently, interlinked nodes can be transformed into independent points and supplementary coordinates without information loss. Theoretically, two graphs are isomorphic iff the two converted point sets are equal up to anorthogonal transformation(because for anyQ⁢QT=A+D𝑄superscript𝑄𝑇𝐴𝐷QQ^{T}=A+Ditalic_Q italic_Q start_POSTSUPERSCRIPT italic_T end_POSTSUPERSCRIPT = italic_A + italic_D,Q⁢R𝑄𝑅QRitalic_Q italic_Ris also a solution whereR𝑅Ritalic_Ris any orthogonal matrix). This equivalence empowers us to encode the set with coordinates in an orthogonal-transformation-equivariant manner, akin to E(3)-equivariant models designed for 3D geometric deep learning. Importantly, our approach is versatile, allowing for using any equivariant set encoder, thereby significantly expanding the design space of GNNs. Furthermore, for Transformer, a specific set encoder, our method offers a novel and principled way to inject graph information losslessly. In AppendixD, we additionally show that it unifies various heuristic structural/positional encodings in previous GTs, including random walk(Li et al.,2020; Dwivedi et al.,2023; Rampásek et al.,2022), heat kernel(Mialon et al.,2021), and resistance distance(Zhang et al.,2023b).\n\n\nFigure 1:Our method converts the input graph to a point set first and encoding it with a set encoder.O⁢(r)𝑂𝑟O(r)italic_O ( italic_r )denotes the set ofr𝑟ritalic_r-dimension orthogonal transformations.\n\n\nTo instantiate our method, we introduce an orthogonal-transformation-equivariant Transformer, namely Point Set Transformer (PST), to encode the point set. PST provably surpasses existing models in long-range and short-range expressivity. Extensive experiments verify these claims across synthetic datasets, graph property prediction datasets, and long-range graph benchmarks. Specifically, PST outperforms all baselines on QM9(Wu et al.,2017)dataset. Moreover, our graph-to-set method is not constrained to only one specific set encoder. We also propose a Deepset(Segol & Lipman,2020)-based model, which outperforms comparable to GIN(Xu et al.,2019b)on our datasets.\n\n\nDifferences from eigendecomposition.Note that our graph-to-set method is distinct from previous approaches that decompose adjacency matrices for positional encodings(Dwivedi et al.,2023; Wang et al.,2022; Lim et al.,2023; Bo et al.,2023). The key differences root in that previous methods primarily relied on eigendecomposition (EVD), whereas our method is based on symmetric rank decomposition (SRD). Their differences are as follows:•SRD enables a practical conversion of graph problems into set problems. SRD of a matrix is unique up to a single orthogonal transformation, while EVD is unique up to a combination of orthogonal transformations within each eigenspace. This difference allows SRD-based models to easily maintain symmetry, ensuring consistent predictions for isomorphic graphs, while EVD-based methods(Lim et al.,2023)struggle because they need to deal with each eigenspace individually, making them less suitable for graph-level tasks where eigenspaces vary between graphs.•Due to the advantage of SRD, we can utilize set encoder with coordinates to capture graph structure, thus expanding the design space of GNN. Moreover, our method provides a principled way to add graph information to Transformers. Note that previous GTs usually require multiple heuristic encodings together. Besides node positional encodings, they also use adjacency matrices: Grit(Ma et al.,2023)and graphit(Mialon et al.,2021)use random walk matrix (normalized adjacency) as relative positional encoding (RPE). Graph Transformer(Dwivedi & Bresson,2020), Graphormer(Ying et al.,2021), and SAN(Kreuzer et al.,2021)use adjacency matrix as RPE.Dwivedi & Bresson (2020)’s ablation shows that adjacency is crucial. GPS(Rampásek et al.,2022), Exphormer(Shirzad et al.,2023), higher-order Transformer(Kim et al.,2021), and GraphVit/MLP-Mixer(He et al.,2023)even directly incorporate message passing blocks which use adjacency matrix to guide message passing between nodes.In summary, this paper introduces a novel approach to graph representation learning by converting interconnected graphs into independent points and subsequently encoding them using an orthogonal-transformation-equivariant set encoder like our Point Set Transformer. This innovative approach outperforms existing methods in both long- and short-range tasks, as validated by comprehensive experiments.\n\n",
    "related_work": "Graph Neural Network with Eigen-Decomposition.Our approach employs coordinates derived from the symmetric rank decomposition (SRD) of adjacency or related matrices, differing from prior studies that primarily rely on eigendecomposition (EVD). While both approaches have similarities, SRD transforms the graph isomorphism problem into a set problembijectively, which is challenging for EVD, because SRD of a matrix is unique up to a single orthogonal transformation, while EVD is unique up to multiple orthogonal transformations in different eigenspaces. This key theoretical difference has profound implications for model design. Early efforts, likeDwivedi et al. (2023), introduce eigenvectors into MPNNs’ input node feature(Gilmer et al.,2017), and subsequent works, such as Graph Transformers (GTs)(Dwivedi & Bresson,2020; Kreuzer et al.,2021), incorporate eigenvectors as node positional encodings. However, due to the non-uniqueness of eigenvectors, these models produce varying predictions for isomorphic graphs, limiting their generalization.Lim et al. (2023)partially solve the non-uniqueness problem. However, their solutions are limited to cases with constant eigenvalue multiplicity in graph tasks due to the property of EVD. On the other hand, approaches likeWang et al. (2022),Bo et al. (2023), andHuang et al. (2024)completely solve non-uniqueness and even apply permutation-equivariant functions to eigenvalues, similar to our PSRD. However, these methods aim to enhance existing MPNNs and GTs with heuristic features. In contrast, we perfectly align graph-level tasks with set-level tasks through SRD, allowing us to convert orthogonal-transformation-equivariant set encoders to graph encoders and to inject graph structure information into Transformers in a principled ways.\n\n\nTable 1:Normalized MAE (↓↓\\downarrow↓) on substructure counting tasks. FollowingHuang et al. (2023b), models can count the structure if the test loss≤\\leq≤10 units (yellow cellin the table), measured using a scale of10−3superscript10310^{-3}10 start_POSTSUPERSCRIPT - 3 end_POSTSUPERSCRIPT. TT: Tailed Triangle. CC: Chordal Cycle, TR: Triangle-Rectangle.Method2-Path3-Path4-Path5-path6-path3-Cycle4-Cycle5-Cycle6-Cycle7-cycleTTCCTRMPNN1.067.3159.2235.3321.5351.5274.2208.8155.5169.8363.1311.4297.9IDGNN1.91.827.368.678.30.62.24949.549.9105.345.462.8NGNN1.52.124.475.482.60.31.340.243.952.2104.439.272.9GNNAK4.540.77.547.948.80.44.113.323.879.84.311.2131.1I2-GNN1.52.64.154.463.80.31.62.88.239.91.11.01.3PPGN0.31.74.115.121.70.30.93.67.127.12.61.514.4PSDS2.2±0.1subscript2.2plus-or-minus0.12.2_{\\pm 0.1}2.2 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT2.6±0.4subscript2.6plus-or-minus0.42.6_{\\pm 0.4}2.6 start_POSTSUBSCRIPT ± 0.4 end_POSTSUBSCRIPT4.9±0.8subscript4.9plus-or-minus0.84.9_{\\pm 0.8}4.9 start_POSTSUBSCRIPT ± 0.8 end_POSTSUBSCRIPT9.9±0.5subscript9.9plus-or-minus0.59.9_{\\pm 0.5}9.9 start_POSTSUBSCRIPT ± 0.5 end_POSTSUBSCRIPT15.8±0.2subscript15.8plus-or-minus0.215.8_{\\pm 0.2}15.8 start_POSTSUBSCRIPT ± 0.2 end_POSTSUBSCRIPT0.6±0.7subscript0.6plus-or-minus0.70.6_{\\pm 0.7}0.6 start_POSTSUBSCRIPT ± 0.7 end_POSTSUBSCRIPT2.2±0.3subscript2.2plus-or-minus0.32.2_{\\pm 0.3}2.2 start_POSTSUBSCRIPT ± 0.3 end_POSTSUBSCRIPT5.8±0.6subscript5.8plus-or-minus0.65.8_{\\pm 0.6}5.8 start_POSTSUBSCRIPT ± 0.6 end_POSTSUBSCRIPT25.1±0.7subscript25.1plus-or-minus0.725.1_{\\pm 0.7}25.1 start_POSTSUBSCRIPT ± 0.7 end_POSTSUBSCRIPT57.7±0.3subscript57.7plus-or-minus0.357.7_{\\pm 0.3}57.7 start_POSTSUBSCRIPT ± 0.3 end_POSTSUBSCRIPT6.0±1.3subscript6.0plus-or-minus1.36.0_{\\pm 1.3}6.0 start_POSTSUBSCRIPT ± 1.3 end_POSTSUBSCRIPT29.8±3.0subscript29.8plus-or-minus3.029.8_{\\pm 3.0}29.8 start_POSTSUBSCRIPT ± 3.0 end_POSTSUBSCRIPT56.4±4.7subscript56.4plus-or-minus4.756.4_{\\pm 4.7}56.4 start_POSTSUBSCRIPT ± 4.7 end_POSTSUBSCRIPTPST0.7±0.1subscript0.7plus-or-minus0.10.7_{\\pm 0.1}0.7 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT1.1±0.1subscript1.1plus-or-minus0.11.1_{\\pm 0.1}1.1 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT1.5±0.1subscript1.5plus-or-minus0.11.5_{\\pm 0.1}1.5 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT2.2±0.1subscript2.2plus-or-minus0.12.2_{\\pm 0.1}2.2 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT3.3±0.3subscript3.3plus-or-minus0.33.3_{\\pm 0.3}3.3 start_POSTSUBSCRIPT ± 0.3 end_POSTSUBSCRIPT0.8±0.1subscript0.8plus-or-minus0.10.8_{\\pm 0.1}0.8 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT1.9±0.2subscript1.9plus-or-minus0.21.9_{\\pm 0.2}1.9 start_POSTSUBSCRIPT ± 0.2 end_POSTSUBSCRIPT3.1±0.3subscript3.1plus-or-minus0.33.1_{\\pm 0.3}3.1 start_POSTSUBSCRIPT ± 0.3 end_POSTSUBSCRIPT4.9±0.3subscript4.9plus-or-minus0.34.9_{\\pm 0.3}4.9 start_POSTSUBSCRIPT ± 0.3 end_POSTSUBSCRIPT8.6±0.5subscript8.6plus-or-minus0.58.6_{\\pm 0.5}8.6 start_POSTSUBSCRIPT ± 0.5 end_POSTSUBSCRIPT3.0±0.1subscript3.0plus-or-minus0.13.0_{\\pm 0.1}3.0 start_POSTSUBSCRIPT ± 0.1 end_POSTSUBSCRIPT4.0±0.7subscript4.0plus-or-minus0.74.0_{\\pm 0.7}4.0 start_POSTSUBSCRIPT ± 0.7 end_POSTSUBSCRIPT9.2±0.9subscript9.2plus-or-minus0.99.2_{\\pm 0.9}9.2 start_POSTSUBSCRIPT ± 0.9 end_POSTSUBSCRIPT\n\n\nEquivariant Point Cloud and 3-D Molecule Neural Networks.Equivariant point cloud and 3-D molecule tasks share resemblances: both involve unordered sets of 3-D coordinate points as input and require models to produce predictions invariant/equivariant to orthogonal transformations and translations of coordinates. Several works(Chen et al.,2021; Winkels & Cohen,2018; Cohen et al.,2018; Gasteiger et al.,2021)introduce specialized equivariant convolution operators to preserve prediction symmetry, yet are later surpassed by models that learn both invariant and equivariant representations for each point, transmitting these representations between nodes. Notably, certain models(Satorras et al.,2021; Schütt et al.,2021; Deng et al.,2021; Wang & Zhang,2022)directly utilize vectors mirroring input coordinate changes as equivariant features, while others(Thomas et al.,2018; Batzner et al.,2022; Fuchs et al.,2020; Hutchinson et al.,2021; Worrall et al.,2017; Weiler et al.,2018)incorporate high-order irreducible representations of the orthogonal group, achieving proven universal expressivity(Dym & Maron,2021). Our Point Set Transformer (PST) similarly learns both invariant and equivariant point representations. However, due to the specific conversion of point sets from graphs, PST’s architecture varies from existing models. While translation invariance characterizes point clouds and molecules, graph properties are sensitive to coordinate translations in our method. Hence, we adopt inner products of coordinates. Additionally, these prior works center on 3D point spaces, whereas our coordinates exist in high-dimensional space, rendering existing models and theoretical expressivity results based on high-order irreducible representations incompatible with our framework.\n\n",
    "origin_citations": [
      "Benchmarking graph neural networks.",
      "Neural message passing for quantum chemistry.",
      "A generalization of transformer networks to graphs, 2020.",
      "Rethinking graph transformers with spectral attention.",
      "Sign and basis invariant networks for spectral graph representation learning.",
      "Equivariant and stable positional encoding for more powerful graph neural networks.",
      "Specformer: Spectral graph neural networks meet transformers.",
      "On the stability of expressive positional encodings for graph neural networks.",
      "Boosting the cycle counting power of graph neural networks with i$^2$-gnns.",
      "Equivariant point network for 3d point cloud analysis.",
      "3d g-cnns for pulmonary nodule detection, 2018.",
      "Spherical cnns.",
      "Gemnet: Universal directional graph neural networks for molecules.",
      "E (n) equivariant graph neural networks.",
      "Equivariant message passing for the prediction of tensorial properties and molecular spectra.",
      "Vector neurons: A general framework for so(3)-equivariant networks.",
      "Graph neural network with local frame for molecular potential energy surface.",
      "Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds.",
      "E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.",
      "Se(3)-transformers: 3d roto-translation equivariant attention networks.",
      "Lietransformer: Equivariant self-attention for lie groups.",
      "Harmonic networks: Deep translation and rotation equivariance.",
      "3d steerable cnns: Learning rotationally equivariant features in volumetric data.",
      "On the universality of rotation equivariant point cloud networks."
    ],
    "citations": [
      8251,
      216,
      5273,
      195038,
      173256,
      519639,
      78925,
      335312,
      507004,
      35507,
      7394,
      196366,
      6403,
      119706,
      163974,
      236341,
      7395,
      39240,
      169368,
      106179,
      6397,
      43905,
      356627
    ],
    "extract_rate": 0.9583333333333334,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 74,
      "Citation_Quality": 91,
      "Synthesis_Analysis": 85,
      "Final_Score": 84.1,
      "raw_text": "Content_Coherence: 74\nCitation_Quality: 91\nSynthesis_Analysis: 85\nFinal_Score: 84"
    },
    "overall_score": 85.06,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "b6yHkQpSwZ",
        "forum": "b6yHkQpSwZ",
        "number": 1718,
        "title": "Graph As Point Set",
        "normalized_title": "graph as point set",
        "abstract": "Graph is a fundamental data structure to model interconnections between entities. Set, on the contrary, stores independent elements. To learn graph representations, current Graph Neural Networks (GNNs) primarily use message passing to encode the interconnections. In contrast, this paper introduces a novel graph-to-set conversion method that bijectively transforms interconnected nodes into a set of independent points and then uses a set encoder to learn the graph representation. This conversion method holds dual significance. Firstly, it enables using set encoders to learn from graphs, thereby significantly expanding the design space of GNNs. Secondly, for Transformer, a specific set encoder, we provide a novel and principled approach to inject graph information losslessly, different from all the heuristic structural/positional encoding methods adopted in previous graph transformers. To demonstrate the effectiveness of our approach, we introduce Point Set Transformer (PST), a transformer architecture that accepts a point set converted from a graph as input. Theoretically, PST exhibits superior expressivity for both short-range substructure counting and long-range shortest path distance tasks compared to existing GNNs. Extensive experiments further validate PST's outstanding real-world performance. Besides Transformer, we also devise a Deepset-based set encoder, which achieves performance comparable to representative GNNs, affirming the versatility of our graph-to-set method.",
        "authors": [
          "Xiyuan Wang",
          "Pan Li",
          "Muhan Zhang"
        ],
        "decision": null,
        "pdate": 1714610294590,
        "cdate": 1706123039574
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "b6yHkQpSwZ",
          "forum": "b6yHkQpSwZ",
          "number": 1718,
          "title": "Graph As Point Set",
          "normalized_title": "graph as point set",
          "abstract": "Graph is a fundamental data structure to model interconnections between entities. Set, on the contrary, stores independent elements. To learn graph representations, current Graph Neural Networks (GNNs) primarily use message passing to encode the interconnections. In contrast, this paper introduces a novel graph-to-set conversion method that bijectively transforms interconnected nodes into a set of independent points and then uses a set encoder to learn the graph representation. This conversion method holds dual significance. Firstly, it enables using set encoders to learn from graphs, thereby significantly expanding the design space of GNNs. Secondly, for Transformer, a specific set encoder, we provide a novel and principled approach to inject graph information losslessly, different from all the heuristic structural/positional encoding methods adopted in previous graph transformers. To demonstrate the effectiveness of our approach, we introduce Point Set Transformer (PST), a transformer architecture that accepts a point set converted from a graph as input. Theoretically, PST exhibits superior expressivity for both short-range substructure counting and long-range shortest path distance tasks compared to existing GNNs. Extensive experiments further validate PST's outstanding real-world performance. Besides Transformer, we also devise a Deepset-based set encoder, which achieves performance comparable to representative GNNs, affirming the versatility of our graph-to-set method.",
          "authors": [
            "Xiyuan Wang",
            "Pan Li",
            "Muhan Zhang"
          ],
          "decision": null,
          "pdate": 1714610294590,
          "cdate": 1706123039574
        }
      ]
    }
  },
  {
    "paper_id": 43503,
    "title": "X-Drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios",
    "abstract": "Recent advancements have exploited diffusion models for the synthesis of either LiDAR point clouds or camera image data in driving scenarios.\nDespite their success in modeling single-modality data marginal distribution, there is an under-exploration in the mutual reliance between different modalities to describe complex driving scenes. To fill in this gap, we propose a novel framework,X-Drive, to model the joint distribution of point clouds and multi-view images via a dual-branch latent diffusion model architecture.\nConsidering the distinct geometrical spaces of the two modalities,X-Driveconditions the synthesis of each modality on the corresponding local regions from the other modality, ensuring better alignment and realism.\nTo further handle the spatial ambiguity during denoising, we design the cross-modality condition module based on epipolar lines to adaptively learn the cross-modality local correspondence.\nBesides,X-Driveallows for controllable generation through multi-level input conditions, including text, bounding box, image, and point clouds. Extensive results demonstrate the high-fidelity synthetic results ofX-Drivefor both point clouds and multi-view images, adhering to input conditions while ensuring reliable cross-modality consistency. Our code will be made publicly available athttps://github.com/yichen928/X-Drive.",
    "introduction": "Autonomous driving vehicles perceive the world with multiple sensors of different kinds, where LiDAR and cameras play crucial roles by capturing point clouds and multi-view images.\nThey provide complementary geometric measurements and semantic information about the surrounding environment, significantly benefiting tasks such as object detection(Liu et al.,2023b; Xie et al.,2023), motion planning(Sobh et al.,2018), scene reconstruction(Huang et al.,2024; Zhou et al.,2024), and self-supervised representation learning(Yang et al.,2024a; Xie et al.,2024).\nHowever, these advancements hinge on access tolarge amounts of aligned multi-modality data, specifically well-calibrated LiDAR and multi-view camera inputs that describe the same scene.\n\n\nScaling up the collection of such high-quality multi-modality data is costly and a non-trivial effort.\nThe high-quality sensors are expensive, and the calibration process demands intensive human efforts.\nAdditionally, real-world driving data suffers from the severe long-tailed distribution problem, making it an obstacle for corner case collection such as in extreme weather conditions.\nThis raises a natural question:Can we use a controllable way to synthesize aligned multi-modality data?\n\n\nGiven the success in other fields(Rombach et al.,2022; Blattmann et al.,2023; Liu et al.,2023a), generative models offer a promising solution.\nCurrent research focuses either on synthesizing point clouds(Zyrianov et al.,2022; Hu et al.,2024a; Ran et al.,2024)or multi-view images(Gao et al.,2023; Wang et al.,2023; Wen et al.,2024), with limited attention paid to generating multi-modality data.\nA simple combination of these single-modality algorithms results in serious cross-modality mismatches in the synthetic scenes (Fig.1(a)).\nSuch inconsistencies create ambiguities and even contradictory inputs or supervision signals, thus hindering the performance of downstream tasks.\n\n\nCross-modality consistencyserves as the key desiderata of multi-modality data generation. However, there are several challenges in the generation of consistent LiDAR and camera data.\nFirstly, synthetic point clouds and multi-view images must be spatially aligned in all the local regions since they describe the same driving scene,i.e.the shapes and layouts of both foregrounds and backgrounds must be matched.\nSecondly, unlike those 2D pixel-level tasks(Zhang et al.,2023), point clouds and multi-view images have distinct geometrical spaces and data formats. Multi-view images are represented by RGB values in camera perspective views, while point clouds are XYZ coordinates in the 3D space.\nThirdly, the 3D spatial information is ambiguous during generation for both point clouds and multi-view images without reliable point location or pixel depth in the denoising process.\n\n\nIn this paper, we fill in this gap by proposingX-Drive, a novel framework for the joint generation of LiDAR and camera data, as demonstrated in Fig.1(b).\nWe design a dual-branch architecture with two latent diffusion models separately dedicated to the synthesis of point clouds and multi-view images, while a key cross-modality condition module enhancescross-modality consistencybetween them.\nTo model the joint distribution and ensure local spatial alignment, we perform the cross-modality conditions locally with only corresponding regions from the other modality considered.\nAn explicit transform bridges the two different geometrical spaces, converting the cross-modality condition to match each other’s noisy latent space.\nTo handle the positional ambiguity, we resort to a 3D-aware design based on epipolar lines on range images and multi-view images, allowing the cross-attention module to adaptively determine the cross-modality correspondence without explicit 3D positions.\nIn consequence,X-Driveis able to exploit existing single-modality data as conditions to seamlessly generate data in the other modality.\nAdditionally,X-Driveenhances controllability by introducing 3D bounding boxes for geometrical layout control and text prompts for attribute control (e.g.weather and lighting), enabling more flexible, fine-grained, and precise control over both modalities.\n\n\nExtensive experiments demonstrate the great ability ofX-Drivein generating realistic multi-modality sensor data. It notably outperforms previous specialized single-modality algorithms in the quality of both synthetic point clouds and multi-view images.\nMore importantly, for the first time, it demonstrates reliable cross-modality consistency in synthetic scenes with comprehensive and flexible conditions.\nOur contributions are summarized as follows.•We introduceX-Drive, a dual-branch multi-modality latent diffusion framework that, for the first time, enables controllable and reliable synthesis of aligned LiDAR and multi-view camera data.•Our cross-modality epipolar condition module bridges the geometrical gap under spatial ambiguity between point clouds and multi-view images, significantly enhancing modality consistency.•Extensive experimental results demonstrate the effectiveness ofX-Drive, with notable MMD (for point cloud generation) and FID (for multi-view generation) improvements, establishing a new state-of-the-art for remarkable cross-modality consistency.\n\n",
    "related_work": "Conditional generation with diffusion models.Diffusion models(Ho et al.,2020; Dhariwal & Nichol,2021; Podell et al.,2023; Peebles & Xie,2023)exhibit remarkable ability in generating diverse images by learning a progressive denoising process. They yield state-of-the-art results in various tasks such as text-to-image generation(Rombach et al.,2022; Nichol et al.,2021), text-to-video generation(Singer et al.,2022; Yu et al.,2022), and instructional image editing(Brooks et al.,2023; Meng et al.,2022). Beyond text condition, several works emerges with the competence in managing additional forms of control signals.Zhang et al. (2023)integrates spatial conditions to a pretrained text-to-image diffusion model via efficient finetuning.Li et al. (2023); Zheng et al. (2023)condition the image synthesis on the fine-grained geometrical annotations to facilitate downstream tasks like 2D object detection.Liu et al. (2023a); Sargent et al. (2024)introduce 3D-aware diffusion models for novel view synthesis based on single-view image inputs. Unlike these prior work for image domain, our framework simultaneously generates point clouds and multi-view images conditioned on text prompts and 3D bounding boxes.\n\n\nCross-modality data generation.Beyond image domain, generative models are extended to the synthesis of multi-modality data.Hu et al. (2024b); Shao et al. (2024)leverage video diffusion model to estimate sequential depths.Bai et al. (2024)formulates various vision tasks as next token prediction by extracting diverse visual information as unified visual tokens. The cross-modality synthesis between videos and audios also draws great attention. Efforts are devoted to the video-to-audio(Zhu et al.,2022), audio-to-video(Chatterjee & Cherian,2020), bi-directional(Chen et al.,2017; Hao et al.,2018), and joint multi-modality(Ruan et al.,2023)generation. However, in all the above cases, there exists clear cross-modality alignments, such as the pixel-to-pixel spatial correspondence for depth estimation and frame-to-frame temporal alignment between audios and videos. In contrast, point clouds and multi-view images are in different geometrical spaces without clear one-to-one local correlation between noisy multi-modality samples.\n\n\nGeneration of vehicle sensor data.Panoramic driving environment is usually perceived by the multi-view cameras and LiDARs on the ego-vehicle. Alternative to the laborious collection and annotation process, many works resort to the fast-growing generative models for the synthesis of either camera(Swerdlow et al.,2024; Yang et al.,2023; Gao et al.,2023)or LiDAR(Zyrianov et al.,2022; Hu et al.,2024a; Zyrianov et al.,2024)sensor data. For multi-view images(Yang et al.,2023; Wang et al.,2023; Gao et al.,2023)or videos(Wen et al.,2024; Lu et al.,2023), previous work leverages pretrained image diffusion models(Rombach et al.,2022)by incorporating extra inter-view modules to ensure the multi-view consistency. For point clouds,Zyrianov et al. (2022); Ran et al. (2024); Hu et al. (2024a)adapt latent diffusion models for the generation of range images(Fan et al.,2021; Li et al.,2016)since this range-view representation shares a similar format with RGB images. However, all above methods concentrate on the single-modality data, either LiDAR or camera sensor. Despite various control signals, there exists no guarantee for the cross-modality consistency between point clouds and multi-view images generated by independent single-modality models. To this end, we propose to synthesize consistent multi-modality data in a joint manner.\n\n",
    "origin_citations": [
      "Denoising diffusion probabilistic models.",
      "Diffusion models beat gans on image synthesis.",
      "Sdxl: Improving latent diffusion models for high-resolution image synthesis.",
      "Scalable diffusion models with transformers.",
      "High-resolution image synthesis with latent diffusion models.",
      "Glide: Towards photorealistic image generation and editing with text-guided diffusion models.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Generating videos with dynamics-aware implicit generative adversarial networks.",
      "Instructpix2pix: Learning to follow image editing instructions.",
      "Sdedit: Guided image synthesis and editing with stochastic differential equations.",
      "Adding conditional control to text-to-image diffusion models.",
      "Gligen: Open-set grounded text-to-image generation.",
      "Layoutdiffusion: Controllable diffusion model for layout-to-image generation.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Zeronvs: Zero-shot 360-degree view synthesis from a single image.",
      "Depthcrafter: Generating consistent long depth sequences for open-world videos.",
      "Learning temporally consistent video depth from video diffusion priors.",
      "Sequential modeling enables scalable learning for large vision models.",
      "Discrete contrastive diffusion for cross-modal music and image generation.",
      "Sound2sight: Generating visual dynamics from sound and context.",
      "Deep cross-modal audio-visual generation.",
      "Cmcgan: A uniform framework for cross-modal visual-audio mutual generation.",
      "Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation.",
      "Street-view image generation from a bird’s-eye view layout.",
      "Bevcontrol: Accurately controlling street-view elements with multi-perspective consistency via bev sketch layout.",
      "Magicdrive: Street view generation with diverse 3d geometry control.",
      "Learning to generate realistic lidar point clouds.",
      "Rangeldm: Fast realistic lidar point cloud generation.",
      "Lidardm: Generative lidar simulation in a generated world.",
      "Drivedreamer: Towards real-world-driven world models for autonomous driving.",
      "Panacea: Panoramic and controllable video generation for autonomous driving.",
      "Wovogen: World volume-aware diffusion for controllable multi-camera driving scene generation.",
      "Towards realistic scene generation with lidar diffusion models.",
      "Rangedet: In defense of range view for lidar-based 3d object detection.",
      "Vehicle detection from 3d lidar using fully convolutional network."
    ],
    "citations": [
      4827,
      165,
      187204,
      140016,
      2240,
      2559,
      176197,
      174321,
      2235,
      96204,
      489610,
      214623,
      492337,
      495077,
      312104,
      743479,
      1045878,
      274494,
      553813,
      521006,
      210622,
      398913,
      140115,
      302457,
      182079,
      323290,
      229461,
      920783,
      448895,
      481042,
      22634,
      273413,
      918682,
      508741,
      7412
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 80,
      "Final_Score": 83.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 80\nFinal_Score: 83.5"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 89,
      "Synthesis_Analysis": 85,
      "Final_Score": 86.0,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 89\nSynthesis_Analysis: 85\nFinal_Score: 86"
    },
    "overall_score": 85.0,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "IEMmEd5Jgm",
        "forum": "IEMmEd5Jgm",
        "number": 2234,
        "title": "X-Drive: Cross-modality Consistent Multi-Sensor Data Synthesis for Driving Scenarios",
        "normalized_title": "x drive cross modality consistent multi sensor data synthesis for driving scenarios",
        "abstract": "Recent advancements have exploited diffusion models for the synthesis of either LiDAR point clouds or camera image data in driving scenarios. Despite their success in modeling single-modality data marginal distribution, there is an under- exploration in the mutual reliance between different modalities to describe com- plex driving scenes. To fill in this gap, we propose a novel framework, X-DRIVE, to model the joint distribution of point clouds and multi-view images via a dual- branch latent diffusion model architecture. Considering the distinct geometrical spaces of the two modalities, X-DRIVE conditions the synthesis of each modality on the corresponding local regions from the other modality, ensuring better alignment and realism. To further handle the spatial ambiguity during denoising, we design the cross-modality condition module based on epipolar lines to adaptively learn the cross-modality local correspondence. Besides, X-DRIVE allows for controllable generation through multi-level input conditions, including text, bounding box, image, and point clouds. Extensive results demonstrate the high-fidelity synthetic results of X-DRIVE for both point clouds and multi-view images, adhering to input conditions while ensuring reliable cross-modality consistency. Our code will be made publicly available at https://github.com/yichen928/X-Drive.",
        "authors": [
          "Yichen Xie",
          "Chenfeng Xu",
          "Chensheng Peng",
          "Shuqi Zhao",
          "Nhat Ho",
          "Alexander T. Pham",
          "Mingyu Ding",
          "Masayoshi Tomizuka",
          "Wei Zhan"
        ],
        "decision": null,
        "pdate": 1737562264237,
        "cdate": 1726864930986
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "IEMmEd5Jgm",
          "forum": "IEMmEd5Jgm",
          "number": 2234,
          "title": "X-Drive: Cross-modality Consistent Multi-Sensor Data Synthesis for Driving Scenarios",
          "normalized_title": "x drive cross modality consistent multi sensor data synthesis for driving scenarios",
          "abstract": "Recent advancements have exploited diffusion models for the synthesis of either LiDAR point clouds or camera image data in driving scenarios. Despite their success in modeling single-modality data marginal distribution, there is an under- exploration in the mutual reliance between different modalities to describe com- plex driving scenes. To fill in this gap, we propose a novel framework, X-DRIVE, to model the joint distribution of point clouds and multi-view images via a dual- branch latent diffusion model architecture. Considering the distinct geometrical spaces of the two modalities, X-DRIVE conditions the synthesis of each modality on the corresponding local regions from the other modality, ensuring better alignment and realism. To further handle the spatial ambiguity during denoising, we design the cross-modality condition module based on epipolar lines to adaptively learn the cross-modality local correspondence. Besides, X-DRIVE allows for controllable generation through multi-level input conditions, including text, bounding box, image, and point clouds. Extensive results demonstrate the high-fidelity synthetic results of X-DRIVE for both point clouds and multi-view images, adhering to input conditions while ensuring reliable cross-modality consistency. Our code will be made publicly available at https://github.com/yichen928/X-Drive.",
          "authors": [
            "Yichen Xie",
            "Chenfeng Xu",
            "Chensheng Peng",
            "Shuqi Zhao",
            "Nhat Ho",
            "Alexander T. Pham",
            "Mingyu Ding",
            "Masayoshi Tomizuka",
            "Wei Zhan"
          ],
          "decision": null,
          "pdate": 1737562264237,
          "cdate": 1726864930986
        }
      ]
    }
  },
  {
    "paper_id": 11262,
    "title": "Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling",
    "abstract": "In the realm of music AI, arranging rich and structured multi-track accompaniments from a simple lead sheet presents significant challenges. Such challenges include maintaining track cohesion, ensuring long-term coherence, and optimizing computational efficiency. In this paper, we introduce a novel system that leveragesprior modelling over disentangled style factorsto address these challenges. Our method presents a two-stage process: initially, a piano arrangement is derived from the lead sheet by retrievingpiano texturestyles; subsequently, a multi-track orchestration is generated by infusingorchestral functionstyles into the piano arrangement. Our key design is the use of vector quantization and a unique multi-stream Transformer to model thelong-termflow of the orchestration style, which enables flexible, controllable, and structured music generation. Experiments show that by factorizing the arrangement task into interpretable sub-stages, our approach enhances generative capacity while improving efficiency. Additionally, our system supports a variety of music genres and provides style control at different composition hierarchies. We further show that our system achieves superior coherence, structure, and overall arrangement quality compared to existing baselines.",
    "introduction": "Representation learning techniques have enabled new possibilities for controllable generative modelling. By learningimplicitstyle representations, which are often hard to explicitly label (e.g., timbre of music audio[21], texture of music composition[39], and artistic style in paintings[20]), new music and artworks can be created via style transfer and latent space sampling. These learned style factors can also serve as external controls for downstream generative models, including Transformers[18,36]and diffusion models[42]. However, applying style factors tolong-termsequence generation remains a challenging task. Existing approaches rely on style templates specified manually or by heuristic rules[36,42,51], which are impractical for long-term generation. Moreover, when structural constraints are imposed, misaligned style factors can result in incoherent outputs.\n\n\nTo address these challenges, we aim to develop a novel sequence generation framework leveraging a global style planner, orprior, which models the conditional distribution ofstyle factorsgiven the model input’scontent factors. Both style and content factors are sequences of compact, structurally aligned latent codes over a disentangled representation space. By infusing the style back to the content, we can recover the observational target with globally coherent style patterns.\n\n\nIn this paper, we studystyle prior modellingthrough the task ofmulti-track accompaniment arrangement, a typical scenario for long-term conditional sequence generation. We assume the input of a piano accompaniment score, which typically carries a verse-chorus structure. Our target is to generate corresponding multi-track arrangements featuring band orchestration. We start by disentangling a band score at timet𝑡titalic_tintopiano reduction𝐜tsubscript𝐜𝑡\\mathbf{c}_{t}bold_c start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT(content factor) andorchestral function𝐬tksuperscriptsubscript𝐬𝑡𝑘\\mathbf{s}_{t}^{k}bold_s start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_k end_POSTSUPERSCRIPT(style factors for individual tracksk=1,2,⋯,K𝑘12⋯𝐾k=1,2,\\cdots,Kitalic_k = 1 , 2 , ⋯ , italic_K). On top of this, we model the prior offinding appropriate functions to orchestrate a given piano score, or formallyp⁢(𝐬1:T1:K∣𝐜1:T)𝑝conditionalsuperscriptsubscript𝐬:1𝑇:1𝐾subscript𝐜:1𝑇p(\\mathbf{s}_{1:T}^{1:K}\\mid\\mathbf{c}_{1:T})italic_p ( bold_s start_POSTSUBSCRIPT 1 : italic_T end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 1 : italic_K end_POSTSUPERSCRIPT ∣ bold_c start_POSTSUBSCRIPT 1 : italic_T end_POSTSUBSCRIPT ). To model dependencies in both time (T𝑇Titalic_T) and track (K𝐾Kitalic_K) directions, we develop a multi-stream Transformer with interleaved time-wise and track-wise layers. The track-wise layer allows for flexible control over the choice of instruments and the number of tracks, while the time-wise layer ensures structural alignment through cross-attention to thepiano reduction. Decoding the inferred𝐬1:T1:Ksuperscriptsubscript𝐬:1𝑇:1𝐾\\mathbf{s}_{1:T}^{1:K}bold_s start_POSTSUBSCRIPT 1 : italic_T end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 1 : italic_K end_POSTSUPERSCRIPTwith𝐜1:Tsubscript𝐜:1𝑇\\mathbf{c}_{1:T}bold_c start_POSTSUBSCRIPT 1 : italic_T end_POSTSUBSCRIPT, we can address accompaniment arrangement in a flexiblemulti-trackform with extendedwhole-songstructure.\n\n\nExperiments show that our method outperforms existing sequential token prediction approaches and provides better multi-track cohesion, structural coherence, and computational efficiency. Additionally, compared to existing designs of multi-stream language models, our model handles flexible stream combinations more effectively with enhanced generative capacity.\n\n\nTo summarize, our contributions in this paper are three-folded:•We proposestyle prior modelling, a hierarchical generative methodologyaddressing both long-term structure (via style prior at high level) and fine-grained condition/control (via representation disentanglement at low level). Our approach moves beyond the limitation of manual specification of style factors, providing a flexible, efficient, and self-supervised solution for long-term sequence prediction and generation tasks.•We propose a novellayer interleaving architecturefor multi-stream language modelling. In our case, it models parallel music tracks with a flexible track number, controllable instruments, and manageable computation. To our knowledge, it is the first multi-stream language model with tractable generalization to flexible stream combinations.•Integrating our previous study onpiano texturestyle transfer[39,50], we present acomplete music automation systemarranging an input lead sheet (a basic music form with melody and chord only) via piano accompaniment to multi-track arrangement. The entire system is interpretable at two composition hierarchies: 1)piano textureand 2)orchestral function, and demonstrates state-of-the-art arrangement performance for varied genres of music.111Demo and more resources:https://zhaojw1998.github.io/structured-arrangement/\n\n",
    "related_work": "In this section, we overview three topics related to our study. Section2.1reviews existing studies on representation disentanglement. Section2.2summarizes prior modelling methods in music generation. Section2.3reviews the current progress with the task of accompaniment arrangement.\n\n\n2.1Content-Style Disentanglement via Representation LearningRepresentation disentanglement is a popular technique in deep generative modelling[49,16,3,48]. In the music domain, this approach has proven valuable by learning compositional factors related to musicstyleandcontent. By manipulating these factors through interpolation[32], swapping[39], and prior sampling[46], it provides a self-supervised and controllable pathway for various music automation tasks. Recent works leverage disentangled style factors as control signals for long-term music generation[36,42]. However, these approaches typically treat style representations as fixed condition sequences during training, requiring manual specification or additional algorithms for control during inference. In contrast, we model the prior ofthe style to applyconditional on the given music content, which is a more generalized and flexible approach.\n\n\n2.2Music Generation with Latent PriorIn sequence generation tasks (e.g., music and audio), learning a prior sampler over a compact, latent representation space is often more efficient and effective. Jukebox[7]models the latent codes encoded by VQ-VAEs[34]as music priors, which can further reconstruct minutes of music audio. More recently, MusicLM[2]and MusicGen[4]learn multi-modal priors for generating music from text prompts. While prior modelling facilitates long-term generation, the latent codes in these works are not interpretable, thus lacking a precise control by music content-based signals (e.g., music structure). Such controls are essential for conditional generation tasks, including accompaniment arrangement. In this paper, we model astyle priorconditional on the disentangled music content, which allows for structured long-term music generation, enhancing both interpretability and controllability.\n\n\n2.3Accompaniment ArrangementAccompaniment arrangement aims to compose the accompaniment part given a lead sheet, which is a difficult conditional generation task involving structural constraints. Existing methods mainly train a conditional language model based on sequential note-level tokenization[15,30,14,33], which often suffer from slow inference speed, truncated structural context, and/or simplified instrumentation. Recent attempts with diffusion models show higher sample quality with faster inference[26,27,23], but still consider limited instruments or tracks. AccoMontage[50,47]maintains a whole-song structure by manipulating high-level composition factors, but is limited to piano arrangement alone. Our paper presents a two-stage approach: from lead sheet to piano accompaniment, and from piano to multi-track, both leveraging prior modelling of high-level style factors. This approach offers modularity[11]and enables high-qualitywhole-songandmulti-trackaccompaniment arrangement.\n\n",
    "origin_citations": [
      "Improving zero-shot voice style transfer via disentangled representation learning.",
      "Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech.",
      "Everybody dance now.",
      "Dance style transfer with cross-modal transformer.",
      "Music fadernets: Controllable music generation based on high-level features via low-level feature modelling.",
      "Learning interpretable representation for controllable polyphonic music generation.",
      "Deep music analogy via latent representation disentanglement.",
      "Figaro: Generating symbolic music with fine-grained artistic control.",
      "Whole-song hierarchical generation of symbolic music using cascaded diffusion models.",
      "Jukebox: A generative model for music.",
      "Neural discrete representation learning.",
      "Musiclm: Generating music from text.",
      "Simple and controllable music generation.",
      "Music transformer: Generating music with long-term structure.",
      "Popmag: Pop music accompaniment generation.",
      "Compound word transformer: Learning to compose full-song music over dynamic directed hypergraphs.",
      "Anticipatory music transformer.",
      "Polyffusion: A diffusion model for polyphonic score generation with internal and external controls.",
      "Discrete diffusion probabilistic models for symbolic music generation.",
      "Getmusic: Generating any music tracks with a unified representation and diffusion framework.",
      "Accomontage: Accompaniment arrangement via phrase selection and style transfer.",
      "Accomontage2: A complete harmonization and accompaniment arrangement system.",
      "Enabling factorized piano music modeling and generation with the MAESTRO dataset."
    ],
    "citations": [
      5690,
      8618,
      5138,
      68492,
      519953,
      238767,
      268913,
      56383,
      977560,
      8727,
      3018,
      8033,
      305828,
      238572,
      39503,
      304226,
      470137,
      262066,
      261379,
      453161,
      230894,
      385978
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 84,
      "Final_Score": 83.8,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 84\nFinal_Score: 84"
    },
    "overall_score": 84.88,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "M75dBr10dZ",
        "forum": "M75dBr10dZ",
        "number": 17467,
        "title": "Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling",
        "normalized_title": "structured multi track accompaniment arrangement via style prior modelling",
        "abstract": "In the realm of music AI, arranging rich and structured multi-track accompaniments from a simple lead sheet presents significant challenges. Such challenges include maintaining track cohesion, ensuring long-term coherence, and optimizing computational efficiency. In this paper, we introduce a novel system that leverages prior modelling over disentangled style factors to address these challenges. Our method presents a two-stage process: initially, a piano arrangement is derived from the lead sheet by retrieving piano texture styles; subsequently, a multi-track orchestration is generated by infusing orchestral function styles into the piano arrangement. Our key design is the use of vector quantization and a unique multi-stream Transformer to model the long-term flow of the orchestration style, which enables flexible, controllable, and structured music generation. Experiments show that by factorizing the arrangement task into interpretable sub-stages, our approach enhances generative capacity while improving efficiency. Additionally, our system supports a variety of music genres and provides style control at different composition hierarchies. We further show that our system achieves superior coherence, structure, and overall arrangement quality compared to existing baselines.",
        "authors": [
          "Jingwei Zhao",
          "Gus Xia",
          "Ziyu Wang",
          "Ye Wang"
        ],
        "decision": null,
        "pdate": 1727288154293,
        "cdate": 1715779730248
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "M75dBr10dZ",
          "forum": "M75dBr10dZ",
          "number": 17467,
          "title": "Structured Multi-Track Accompaniment Arrangement via Style Prior Modelling",
          "normalized_title": "structured multi track accompaniment arrangement via style prior modelling",
          "abstract": "In the realm of music AI, arranging rich and structured multi-track accompaniments from a simple lead sheet presents significant challenges. Such challenges include maintaining track cohesion, ensuring long-term coherence, and optimizing computational efficiency. In this paper, we introduce a novel system that leverages prior modelling over disentangled style factors to address these challenges. Our method presents a two-stage process: initially, a piano arrangement is derived from the lead sheet by retrieving piano texture styles; subsequently, a multi-track orchestration is generated by infusing orchestral function styles into the piano arrangement. Our key design is the use of vector quantization and a unique multi-stream Transformer to model the long-term flow of the orchestration style, which enables flexible, controllable, and structured music generation. Experiments show that by factorizing the arrangement task into interpretable sub-stages, our approach enhances generative capacity while improving efficiency. Additionally, our system supports a variety of music genres and provides style control at different composition hierarchies. We further show that our system achieves superior coherence, structure, and overall arrangement quality compared to existing baselines.",
          "authors": [
            "Jingwei Zhao",
            "Gus Xia",
            "Ziyu Wang",
            "Ye Wang"
          ],
          "decision": null,
          "pdate": 1727288154293,
          "cdate": 1715779730248
        }
      ]
    }
  },
  {
    "paper_id": 2764,
    "title": "On the Variance of Neural Network Training with respect to Test Sets and Distributions",
    "abstract": "Typical neural network trainings have substantial variance in test-set performance between repeated runs, impeding hyperparameter comparison and training reproducibility.\nIn this work we present the following results towards understanding this variation.\n(1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have little variance in performance on the underlying test-distributionsfrom which their test-sets are sampled.\n(2) We show that these trainings make approximately independent errors on their test-sets.\nThat is, the event that a trained network makes an error on one particular example does not affect its chances of making errors on other examples, relative to their average rates over repeated runs of training with the same hyperparameters.\n(3) We prove that the variance of neural network trainings on their test-sets is a downstream consequence of the class-calibration property discovered byJiang et al. (2021). Our analysis yields a simple formula which accurately predicts variance for the binary classification case.\n(4) We conduct preliminary studies of data augmentation, learning rate, finetuning instability and distribution-shift through the lens of variance between runs.",
    "introduction": "Modern neural networks(Krizhevsky et al.,2012; He et al.,2016; Vaswani et al.,2017)are trained using stochastic gradient-based algorithms, involving randomized weight initialization, data ordering, and data augmentations. Because of this stochasticity, each independent run of training produces a different network with better or worse performance than average.\n\n\nThis variance between such independent runs is often substantial.Picard (2021)shows that for standard CIFAR-10(Krizhevsky et al.,2009)training configurations, there exist random seeds which differ by 1.3% in terms of test-set accuracy. In comparison, the gap between the top two methods competing for state-of-the-art on CIFAR-10 has been less than 1% throughout the majority of the benchmark’s lifetime111https://paperswithcode.com/sota/image-classification-on-cifar-10.\nPrior works therefore view this variance as an obstacle which impedes comparisons between training configurations(Bouthillier et al.,2021; Picard,2021)and reproducibility(Bhojanapalli et al.,2021; Zhuang et al.,2022). To mitigate stochasticity,Zhuang et al. (2022)study deterministic tooling,Bhojanapalli et al. (2021)develop regularization methods, and many recent works(Wightman et al.,2021; Liu et al.,2022)report the average of validation metrics across multiple runs when comparing training configurations.\n\n\nIn this work we contribute new results towards understanding the variance between runs of neural network trainings. We use an empirical approach involving hundreds of thousands of trained models in order to answer the following questions.\n\n\n1.Repeatedly running a standard training yields a series of models with often substantial variance in test-set accuracy. Do such models have genuine differences in underlying quality? Or is this variance just a form of finite-sample noise due to the limited size of the test-set?2.Does the empirical distribution of test-set accuracy across repeated runs of training possess any definable structure?3.Is there any way to estimate the variance of a given training configuration a priori,i.e., without the need to empirically measure it across many repeated runs?Both the first and last of these questions have immediate practical consequences.\nFor the first question, if the answer is that every run of training yields a model with the same underlying performance on held-out batches of test data, then practitioners can confidently execute just a single run per configuration. Otherwise, a superior strategy would be to take the best model from multiple trials.\nFor the third question, if a method for estimating variance a priori does exist, then this provides a useful tool by which practitioners can confidently estimate the statistical significance of hyperparameter comparisons, without the need for many runs of training.\nTo answer all three questions, we contribute the following results.•Random seeds which are “lucky” with respect to one set of test data perform no better than average with respect to a second set. (Section3.1)•Over repeated runs of training, the distribution of errors made by trained networks can be approximately explained via the framework ofindependent errors. (Section3.2)•Although standard trainings have substantial variance on their test-sets, they have little variance in performance on their underlying test-distributions. (Section3.3)•Variable test-set performance is a downstream effect of the class-calibration property(Jiang et al.,2021)of neural network trainings. For the binary classification case, this property implies a simple formula which accurately predicts variance a priori. (Section3.4)\n\n\nOur experiments show that these results hold true for standard training configurations across both CIFAR-10 and ImageNet(Deng et al.,2009). As a limitation, we show that they donothold true for two exceptional scenarios:\ntrainings with pathological instability (Section4.1), and trainings where there is a shift between the training and test distributions (Section4.4, SectionC). Both of these cases have a large distribution-wise variance, differing from our results in the standard cases.\n\n\nTo complete our study of variance,\nwe additionally conduct preliminary investigations regarding the effect of learning rate (Section4.3) and data augmentation (Section4.2). We find that when increasing the learning rate, accuracy begins to decline at the same point at which significant distribution-wise variance appears. And we find that data augmentation reduces variance, although the mechanism by which this happens is not yet clear.\n\n\n1.1Related workA number of prior works investigate which sources of stochasticity are most responsible for the variation between runs of training.Fort et al. (2019)observe that when using a below-optimal learning rate, randomized data ordering has a smaller impact than model initialization on the churn of predictions between runs.Bhojanapalli et al. (2021)similarly find that fixing the data ordering has no effect, while fixing the model initialization reduces churn. On the other hand,Bouthillier et al. (2021)report that data ordering has a larger impact than model initialization. And finally,Summers & Dinneen (2021)find instead that most variation can be attributed to the high sensitivity of the training process to initial conditions, by showing that a single bit difference in starting parameters leads to the full quantity of prediction churn between runs. We replicate the results ofSummers & Dinneen (2021)in SectionD, although we find that the result depends upon the training duration.Dodge et al. (2020)study variation between runs of BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPTfinetuning, and achieve substantial gains in validation performance via the strategy of re-running finetuning many times and taking the best-performing result.\nWe demonstrate (Section4.1) that for the case of BERTBASEBASE{}_{\\text{BASE}}start_FLOATSUBSCRIPT BASE end_FLOATSUBSCRIPT, the low amount of genuine distribution-wise variance between runs indicates that any performance gains yielded by this strategy would only amount to overfitting the validation set. On the other hand, for BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPTwe demonstrate that there is genuinely significant distribution-wise variance, supporting the use of multiple runs of training as suggested.Mosbach et al. (2020)also study the finetuning instability of BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPT, and suggest that it can be mitigated by warming up the learning rate, training for longer with a smaller learning rate, and using bias correction for Adam(Kingma & Ba,2014).Earlier works on neural network ensembles have found that they are more well-calibrated than individual networks(Lakshminarayanan et al.,2017; Nixon et al.,2020). AndMukhoti et al. (2021)observed that the usefulness of an ensemble’s uncertainty scores depends upon variance between the individual networks.\nOur theoretical results in Section3.4draw upon the relatedclass-wise calibration propertyof neural network trainings. This property was discovered byJiang et al. (2021), who used it to obtain a theoretical proof of the empirical phenomenon that the disagreement rate between two independently trained networks is approximately equal to their error rates(Nakkiran & Bansal,2020; Jiang et al.,2021).Figure 1:Accuracy distributions.The test-set accuracy distributions across our four training durations, displayed as unsmoothed histograms for 60,000 repeated runs of training each. The differences between the “luckiest” and most unlucky run (max minus min accuracy) are 13.2%, 6.6%, 1.7%, and 1.4% for the 0, 4, 16, and 64-epoch training durations, respectively. The standard deviations are 1.87%, 0.56%, 0.19%, and 0.15%.Several prior works(Baldock et al.,2021; Ilyas et al.,2022; Lin et al.,2022)study the effect of randomly varying the data used to train a neural network, across a large number of training runs.\nOur study differs from these works in that we consider the simpler scenario of running a single training algorithm many times without varying anything except the random seed.Broadly, our work is related to research aiming to understand the relationship between pairs of neural networks produced by repeated runs of training. This topic is of both theoretical and practical interest, and has been studied from a variety of angles, including the similarity of internal representations(Li et al.,2015; Kornblith et al.,2019), degree of correlation between predictions(Fort et al.,2019; Jiang et al.,2021), similarity of decision boundaries(Somepalli et al.,2022), path-connectivity in weight-space(Draxler et al.,2018; Garipov et al.,2018), and linear mode connectivity(Frankle et al.,2020; Tatro et al.,2020; Entezari et al.,2021).\n\n",
    "related_work": "A number of prior works investigate which sources of stochasticity are most responsible for the variation between runs of training.Fort et al. (2019)observe that when using a below-optimal learning rate, randomized data ordering has a smaller impact than model initialization on the churn of predictions between runs.Bhojanapalli et al. (2021)similarly find that fixing the data ordering has no effect, while fixing the model initialization reduces churn. On the other hand,Bouthillier et al. (2021)report that data ordering has a larger impact than model initialization. And finally,Summers & Dinneen (2021)find instead that most variation can be attributed to the high sensitivity of the training process to initial conditions, by showing that a single bit difference in starting parameters leads to the full quantity of prediction churn between runs. We replicate the results ofSummers & Dinneen (2021)in SectionD, although we find that the result depends upon the training duration.\n\n\nDodge et al. (2020)study variation between runs of BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPTfinetuning, and achieve substantial gains in validation performance via the strategy of re-running finetuning many times and taking the best-performing result.\nWe demonstrate (Section4.1) that for the case of BERTBASEBASE{}_{\\text{BASE}}start_FLOATSUBSCRIPT BASE end_FLOATSUBSCRIPT, the low amount of genuine distribution-wise variance between runs indicates that any performance gains yielded by this strategy would only amount to overfitting the validation set. On the other hand, for BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPTwe demonstrate that there is genuinely significant distribution-wise variance, supporting the use of multiple runs of training as suggested.Mosbach et al. (2020)also study the finetuning instability of BERTLARGELARGE{}_{\\text{LARGE}}start_FLOATSUBSCRIPT LARGE end_FLOATSUBSCRIPT, and suggest that it can be mitigated by warming up the learning rate, training for longer with a smaller learning rate, and using bias correction for Adam(Kingma & Ba,2014).\n\n\nEarlier works on neural network ensembles have found that they are more well-calibrated than individual networks(Lakshminarayanan et al.,2017; Nixon et al.,2020). AndMukhoti et al. (2021)observed that the usefulness of an ensemble’s uncertainty scores depends upon variance between the individual networks.\nOur theoretical results in Section3.4draw upon the relatedclass-wise calibration propertyof neural network trainings. This property was discovered byJiang et al. (2021), who used it to obtain a theoretical proof of the empirical phenomenon that the disagreement rate between two independently trained networks is approximately equal to their error rates(Nakkiran & Bansal,2020; Jiang et al.,2021).\n\n\nFigure 1:Accuracy distributions.The test-set accuracy distributions across our four training durations, displayed as unsmoothed histograms for 60,000 repeated runs of training each. The differences between the “luckiest” and most unlucky run (max minus min accuracy) are 13.2%, 6.6%, 1.7%, and 1.4% for the 0, 4, 16, and 64-epoch training durations, respectively. The standard deviations are 1.87%, 0.56%, 0.19%, and 0.15%.\n\n\nSeveral prior works(Baldock et al.,2021; Ilyas et al.,2022; Lin et al.,2022)study the effect of randomly varying the data used to train a neural network, across a large number of training runs.\nOur study differs from these works in that we consider the simpler scenario of running a single training algorithm many times without varying anything except the random seed.\n\n\nBroadly, our work is related to research aiming to understand the relationship between pairs of neural networks produced by repeated runs of training. This topic is of both theoretical and practical interest, and has been studied from a variety of angles, including the similarity of internal representations(Li et al.,2015; Kornblith et al.,2019), degree of correlation between predictions(Fort et al.,2019; Jiang et al.,2021), similarity of decision boundaries(Somepalli et al.,2022), path-connectivity in weight-space(Draxler et al.,2018; Garipov et al.,2018), and linear mode connectivity(Frankle et al.,2020; Tatro et al.,2020; Entezari et al.,2021).\n\n",
    "origin_citations": [
      "Deep ensembles: A loss landscape perspective.",
      "On the reproducibility of neural network predictions.",
      "Accounting for variance in machine learning benchmarks.",
      "Nondeterminism and instability in neural network optimization.",
      "Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping.",
      "On the stability of fine-tuning bert: Misconceptions, explanations, and strong baselines.",
      "Adam: A method for stochastic optimization.",
      "Simple and scalable predictive uncertainty estimation using deep ensembles.",
      "Why are bootstrapped deep ensembles not better?",
      "Deep deterministic uncertainty: A simple baseline.",
      "Assessing generalization of sgd via disagreement.",
      "Distributional generalization: A new kind of generalization.",
      "Deep learning through the lens of example difficulty.",
      "Datamodels: Predicting predictions from training data.",
      "Measuring the effect of training data on deep learning predictions via randomized experiments.",
      "Convergent learning: Do different neural networks learn the same representations?",
      "Similarity of neural network representations revisited.",
      "Can neural nets learn the same model twice? investigating reproducibility and double descent from the decision boundary perspective.",
      "Essentially no barriers in neural network energy landscape.",
      "Loss surfaces, mode connectivity, and fast ensembling of dnns.",
      "Linear mode connectivity and the lottery ticket hypothesis.",
      "Optimizing mode connectivity via neuron alignment.",
      "The role of permutation invariance in linear mode connectivity of neural networks."
    ],
    "citations": [
      4671,
      119615,
      218419,
      217017,
      8329,
      351017,
      462387,
      3288,
      473360,
      392702,
      310810,
      394539,
      91039,
      81388,
      407603,
      529910,
      816714,
      93751,
      388070,
      203215,
      189622,
      341188
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 88,
      "Final_Score": 83.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 85,
      "Final_Score": 85.3,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 85\nFinal_Score: 85"
    },
    "overall_score": 84.7,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "pEGSdJu52I",
        "forum": "pEGSdJu52I",
        "number": 8145,
        "title": "On the Variance of Neural Network Training with respect to Test Sets and Distributions",
        "normalized_title": "on the variance of neural network training with respect to test sets and distributions",
        "abstract": "Neural network trainings are stochastic, causing the performance of trained networks to vary across repeated runs of training.\nWe contribute the following results towards understanding this variation.\n(1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have little variance in their performance on the test-distributions from which their test-sets are sampled.\n(2) We introduce the independent errors assumption and show that it suffices to recover the structure and variance of the empirical accuracy distribution across repeated runs of training.\n(3) We prove that test-set variance is unavoidable given the observation that ensembles of identically trained networks are calibrated (Jiang et al., 2021), and demonstrate that the variance of binary classification trainings closely follows a simple formula based on the error rate and number of test examples.\n(4) We conduct preliminary studies of data augmentation, learning rate, finetuning instability and distribution-shift through the lens of variance between runs.",
        "authors": [
          "Keller Jordan"
        ],
        "decision": null,
        "pdate": 1705411032458,
        "cdate": 1695498608181
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "pEGSdJu52I",
          "forum": "pEGSdJu52I",
          "number": 8145,
          "title": "On the Variance of Neural Network Training with respect to Test Sets and Distributions",
          "normalized_title": "on the variance of neural network training with respect to test sets and distributions",
          "abstract": "Neural network trainings are stochastic, causing the performance of trained networks to vary across repeated runs of training.\nWe contribute the following results towards understanding this variation.\n(1) Despite having significant variance on their test-sets, we demonstrate that standard CIFAR-10 and ImageNet trainings have little variance in their performance on the test-distributions from which their test-sets are sampled.\n(2) We introduce the independent errors assumption and show that it suffices to recover the structure and variance of the empirical accuracy distribution across repeated runs of training.\n(3) We prove that test-set variance is unavoidable given the observation that ensembles of identically trained networks are calibrated (Jiang et al., 2021), and demonstrate that the variance of binary classification trainings closely follows a simple formula based on the error rate and number of test examples.\n(4) We conduct preliminary studies of data augmentation, learning rate, finetuning instability and distribution-shift through the lens of variance between runs.",
          "authors": [
            "Keller Jordan"
          ],
          "decision": null,
          "pdate": 1705411032458,
          "cdate": 1695498608181
        }
      ]
    }
  },
  {
    "paper_id": 22717,
    "title": "StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation",
    "abstract": "For recent diffusion-based generative models, maintaining consistent content across a series of generated images, especially those containing subjects and complex details, presents a significant challenge.\nIn this paper, we propose a new way of self-attention calculation, termed Consistent Self-Attention, that significantly boosts the consistency between the generated images and augments prevalent pre-trained diffusion-based text-to-image models in a zero-shot manner.\nTo extend our method to long-range video generation, we further introduce a novel semantic space temporal motion prediction module, named Semantic Motion Predictor. It is trained to estimate the motion conditions between two provided images in the semantic spaces.\nThis module converts the generated sequence of images into videos with smooth transitions and consistent subjects that are significantly more stable than the modules based on latent spaces only, especially in the context of long video generation.\nBy merging these two novel components, our framework, referred to as StoryDiffusion, can describe a text-based story with consistent images or videos encompassing a rich variety of contents.\nThe proposed StoryDiffusion encompasses pioneering explorations in visual story generation with the presentation of images and videos, which we hope could inspire more research from the aspect of architectural modifications.",
    "introduction": "Recently, diffusion models have been developed rapidly and demonstrated extraordinary potential for content generation, such as images(Rombach et al.,2022; Peebles & Xie,2022; Podell et al.,2023), 3D objects(Zeng et al.,2022; Zhou et al.,2021)and videos(Ho et al.,2022b; Wang et al.,2023b).\nWith extensive pre-training and advanced architectures, diffusion models show superior performance in generating very high-quality images and videos over previous generative-adversarial network (GAN) based methods(Brock et al.,2018).\nHowever, generating subject-consistent (e.g. characters with consistent identity and attire) images and videos to describe a story is still challenging for existing models.\nThe commonly used IP-Adapter(Ye et al.,2023)taking an image as a reference could be used to guide the diffusion process to generate images similar to it.\nHowever, due to the strong guidance, the controllability over the generated content of the text prompts is reduced.\nOn the other hand, recent state-of-the-art identity preservation methods, such as InstantID(Wang et al.,2024)and PhotoMaker(Li et al.,2023a), focus on identity controllability but the consistency of the attires and the scenarios cannot be guaranteed.\nHence, in this paper, we aim to find a method that can generate images and videos with consistent characters in terms of both identity and attire while maximizing the controllability of the user via text prompts.\n\n\nConsistent images generated by StoryDiffusion“Jungle Adventure”“The Moon Exploration by Lecun”(a)(b)Transition Videos generated by StoryDiffusion“Video Clips”“Long-Range Video”\\animategraphics[width=]8./videos/teaser/ironman1/015\\animategraphics[width=]8./videos/teaser/kiss/015\\animategraphics[width=]8./videos/teaser/bear/177\\animategraphics[width=]15./videos/teaser/man_long/60212(c)Figure 1:Images and videos generated by our StoryDiffusion. (a) Comic generated by StoryDiffusion telling the story of a man who discovers a treasure while exploring the jungle. (b) Comic generated by StoryDiffusion describing the expedition to the moon by Lecun, with a reference image control(Li et al.,2023a)same as Fig.6(b). (c) Videos generated by our StoryDiffusion. Click the image to play the video. Best viewed withAcrobat Reader. More generated videos can be found in our home project:https://StoryDiffusion.github.io.\n\n\nA common approach to preserve the consistency between different images (or frames in the context of video generation) is to use a temporal module(Ho et al.,2022a; Blattmann et al.,2023b).\nHowever, this requires extensive computational resources and data. Differently, we target to explore a lightweight method with minimum data and computational cost, or even in a zero-shot manner.\n\n\nAs evidenced by previous works(Tian et al.,2023; Hong et al.,2023), self-attention is one of the most important modules for modeling the overall structure of the generated visual content.\nOur main motivation is that if we could use a reference image\nto guide the self-attention calculation, the consistency between the two images is supposed to be improved significantly.\nAs the self-attention weights are input-dependent, model training or fine-tuning might not be required.\nFollowing this idea, we propose Consistent Self-Attention, the core of our StoryDiffusion, which can be inserted into the diffusion backbone to replace the original self-attention in a zero-shot manner.\n\n\nDifferent from the standard self-attention that operates on the tokens representing a single image (as shown in Fig.2(d)), Consistent Self-Attention\nincorporates sampled reference tokens from the reference images during the token similarity matrix calculation and token merging.\nThe sampled tokens share the same set ofQ𝑄Qitalic_Q-K𝐾Kitalic_K-V𝑉Vitalic_Vweights and thus no extra training is required.\nAs shown in Fig.1, the generated images using Consistent Self-Attention successfully preserve the consistency in both identity and attire, which is vital for storytelling.\nIntuitively, Consistent Self-Attention builds correlations across images in the batch, generating consistent character images in terms of identity and attire, such as clothes.\nThis enables us to generate subject-consistent images for storytelling.\n\n\nFor any given story text, we begin by dividing it into several prompts, with each prompt corresponding to an individual image.\nThen our method could generate highly consistent images that effectively narrate a story.\nTo support long story generation, we also implement Consistent Self-Attention together with a sliding window along the temporal dimension. This removes the peak memory consumption’s dependency on the input text length, making it possible to generate long stories.\nTo stream the generated story frames into videos, we further propose Semantic Motion Predictor that can predict transitions between two images in the semantic spaces. We empirically found that predicting motions in the semantic space generates more stable results than the predictions in the image latent spaces. Combined with the pre-trained motion module(Guo et al.,2024), Semantic Motion Predictor can generate smooth video frames that are significantly better than recent conditional video generation methods, such as SEINE(Chen et al.,2023)and SparseCtrl(Guo et al.,2023).\n\n\nOur contributions are summarized below:•We propose a training-free and hot-pluggable attention module, termed Consistent Self-Attention. It can maintain the consistency of characters in a sequence of generated images for storytelling with high text controllability.•We propose a new motion prediction module that can predict transitions between two images in the semantic space, termed Semantic Motion Predictor. It can generate significantly more stable long-range video frames that can be easily upscaled to minutes than recent popular image conditioning methods, such as SEINE(Chen et al.,2023)and SparseCtrl(Guo et al.,2023).•We demonstrate that our approach could generate long image sequences or videos based on a pre-defined text-based story with the proposed Consistent Self-Attention and Semantic Motion Predictor with motions specified by text prompts. We term the new framework as StoryDiffusion.\n\n",
    "related_work": "2.1Diffusion ModelsDiffusion models have rapidly demonstrated their stunning capabilities in generating realistic images and this also enables them to dominate the field of generative modeling in recent years(Rombach et al.,2022; Saharia et al.,2022; Ho et al.,2020; Ramesh et al.,2022; Peebles & Xie,2022).\nBy utilizing a deep denoising network(Ronneberger et al.,2015), diffusion models establish a connection between the noise distribution and the real image distribution through iterative noise addition and denoising.\nEarly works(Ho et al.,2020; Song et al.,2021; Sohl-Dickstein et al.,2015)establish the theoretical foundation of diffusion models mainly focusing on unconditional image generation.\nLater, various efforts are made to enhance the efficiency and performance of diffusion models.\nTypical examples should be efficient sampling methods(Song et al.,2020; Zhang & Chen,2023; Lu et al.,2022), denoising in the latent space(Rombach et al.,2022), controllability(Feng et al.,2023a; Zhou et al.,2023b), diffusion backbones(Ramesh et al.,2022; Peebles & Xie,2022).\nConcurrently with the exploration of the foundational theory, diffusion models gradually gain popularity and demonstrate strong performance across various domains, such as image generation(Zhou et al.,2023b), video generation(Zhou et al.,2023a), 3D generation(Zeng et al.,2022; Zhou et al.,2021), image segmentation(Li et al.,2023b; Amit et al.,2021)and low-level vision tasks(Wang et al.,2023a; Lugmayr et al.,2022; Xie et al.,2022; Saharia et al.,2023).\n\n\n2.2Controllable Text-to-Image GenerationAs an important sub-field of diffusion model applications, text-to-image generation, represented by Latent Diffusion(Rombach et al.,2022), DiT(Peebles & Xie,2022), and Stable XL(Podell et al.,2023), have attracted considerable attention recently.\nIn addition, to enhance the controllability of text-to-image generation, a multitude of methods emerged as well.\nAmong them, ControlNet(Zhang et al.,2023b)and T2I-Adapter(Mou et al.,2023)introduce control conditions, such as depth maps, pose images, or sketches, to direct the generation of images.\nMaskDiffusion(Zhou et al.,2023b)and StructureDiffusion(Feng et al.,2023b)focus on enhancing the text controllability.\nThere are also some works(Mao & Wang,2023; Ma et al.,2023a)controlling the layout of generated images.ID-Preservation, which is expected to generate images with a specified ID, is also a hot topic. According to whether test-time fine-tuning is required, these works can be divided into two major categories.\nThe first one only requires fine-tuning a part of the model with a given image, such as Textual Inversion(Gal et al.,2022), DreamBooth(Ruiz et al.,2023), and Custom Diffusion(Kumari et al.,2023).\nThe other one, exemplified by IPAdapter(Ye et al.,2023)and PhotoMaker(Li et al.,2023a), leverages models that have undergone pre-training on large datasets, allowing the direct use of a given image to control image generation.\nDifferent from both of the two types, we focus on maintaining the subject consistency in multiple images, to narrate a story.\nOur Consistent Self-Attention is training-free and pluggable and can build connections across images within a batch to generate multiple subject-consistent images.\n\n\n2.3Video GenerationDue to the success of diffusion models in the field of image generation(Rombach et al.,2022; Ho et al.,2020), the exploration in the domain of video generation is also becoming popular.\nAs text is the most intuitive descriptor users can specify, text-based video generation has attracted the most attention(Guo et al.,2024; Jiang et al.,2023; Singer et al.,2022; Wang et al.,2023d; Yang et al.,2023).\nVDM(Ho et al.,2022a)is among the first that extend the 2D U-Net from image diffusion models to a 3D U-Net to achieve video generation.\nDue to the significant increase of computational cost for video generation, later works, such as MagicVideo(Zhou et al.,2023a)and Mindscope(Wang et al.,2023b), introduce 1D temporal attention mechanisms, reducing computations by building upon latent diffusion models.\nFollowing Imagen, Imagen Video(Ho et al.,2022b)employs a cascaded sampling pipeline that generates videos through multiple stages.\nShow-1(Zhang et al.,2023a)also proposes a multi-stage approach to balance the generation quality and efficiency.In addition to traditional end-to-end text-to-video (T2V) generation,\nvideo generation using other conditions is also an important direction.\nThis type of methods generates videos with other auxiliary controls, such as depth maps(Guo et al.,2023; He et al.,2023), pose maps(Xu et al.,2023; Hu et al.,2023; Wang et al.,2023c; Ma et al.,2023b), RGB images(Blattmann et al.,2023a; Chen et al.,2023; Ni et al.,2023), or other guided motion videos(Zhao et al.,2023; Wu et al.,2023).\nDifferent from the ambiguity of the text prompt, introducing this conditional information enhances the controllability of video generation.Our video generation method focuses on transition video generation, which is expected to generate videos with a given start frame and an end frame.\nTypical related works are SEINE(Chen et al.,2023)and SparseCtrl(Chen et al.,2023).\nSEINE randomly masks video sequences as the initial input of the video diffusion models in training to enable the predictions of the transition between two frames.\nSparseCtrl introduces a sparse control network to synthesize the corresponding control information for each frame using sparse control data, thereby directing the generation of videos.\nHowever, the aforementioned transition video generation methods rely solely on temporal networks in image latent space for the predictions of intermediate content.\nThus, these methods often perform poorly on complex transitions, such as large-scale movements of characters.\nOur StoryDiffusion aims to perform predictions in image semantic spaces to achieve better performance and can handle larger movements, which we will show in our experiment section.\n\n",
    "origin_citations": [
      "High-resolution image synthesis with latent diffusion models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Denoising diffusion probabilistic models.",
      "Hierarchical text-conditional image generation with clip latents.",
      "Scalable diffusion models with transformers.",
      "U-net: Convolutional networks for biomedical image segmentation.",
      "Score-based generative modeling through stochastic differential equations.",
      "Deep unsupervised learning using nonequilibrium thermodynamics.",
      "Denoising diffusion implicit models.",
      "Fast sampling of diffusion models with exponential integrator.",
      "Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.",
      "Training-free structured diffusion guidance for compositional text-to-image synthesis.",
      "Maskdiffusion: Boosting text-to-image consistency with conditional mask.",
      "Magicvideo: Efficient video generation with latent diffusion models.",
      "Lion: Latent point diffusion models for 3d shape generation.",
      "3d shape generation and completion through point-voxel diffusion.",
      "Open-vocabulary object segmentation with diffusion models.",
      "Segdiff: Image segmentation with diffusion probabilistic models.",
      "Exploiting diffusion prior for real-world image super-resolution.",
      "Repaint: Inpainting using denoising diffusion probabilistic models.",
      "Smartbrush: Text and shape guided object inpainting with diffusion model.",
      "Image super-resolution via iterative refinement.",
      "Sdxl: Improving latent diffusion models for high-resolution image synthesis.",
      "Adding conditional control to text-to-image diffusion models.",
      "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.",
      "Training-free location-aware text-to-image synthesis.",
      "Directed diffusion: Direct control of object placement through attention guidance.",
      "An image is worth one word: Personalizing text-to-image generation using textual inversion.",
      "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation.",
      "Multi-concept customization of text-to-image diffusion.",
      "Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.",
      "Photomaker: Customizing realistic human photos via stacked id embedding.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Text2performer: Text-driven human video generation.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Lavie: High-quality video generation with cascaded latent diffusion models.",
      "Probabilistic adaptation of text-to-video models.",
      "Video diffusion models. arxiv 2022.",
      "Modelscope text-to-video technical report.",
      "Imagen video: High definition video generation with diffusion models.",
      "Show-1: Marrying pixel and latent diffusion models for text-to-video generation.",
      "Sparsectrl: Adding sparse controls to text-to-video diffusion models.",
      "Animate-a-story: Storytelling with retrieval-augmented video generation.",
      "Magicanimate: Temporally consistent human image animation using diffusion model.",
      "Animate anyone: Consistent and controllable image-to-video synthesis for character animation.",
      "Disco: Disentangled control for referring human dance generation in real world.",
      "Follow your pose: Pose-guided text-to-video generation using pose-free videos.",
      "Stable video diffusion: Scaling latent video diffusion models to large datasets.",
      "Seine: Short-to-long video diffusion model for generative transition and prediction.",
      "Conditional image-to-video generation with latent flow diffusion models.",
      "Motiondirector: Motion customization of text-to-video diffusion models.",
      "Lamp: Learn a motion pattern for few-shot-based video generation."
    ],
    "citations": [
      2240,
      5926,
      4827,
      2558,
      140016,
      825,
      431228,
      5918,
      356595,
      367777,
      413881,
      221877,
      128856,
      244720,
      459199,
      28779,
      302242,
      153340,
      114784,
      8039,
      221874,
      894646,
      187204,
      489610,
      228018,
      52295,
      225874,
      236013,
      67241,
      222136,
      179737,
      272723,
      185878,
      54542,
      176197,
      478403,
      307620,
      179813,
      461035,
      478059,
      22570,
      184973,
      22953,
      22671,
      20659,
      23455,
      1040620,
      494021,
      320357,
      315738
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 75,
      "Final_Score": 83.9,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 92\nSynthesis_Analysis: 75\nFinal_Score: 84"
    },
    "overall_score": 84.7,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "VFqzxhINFU",
        "forum": "VFqzxhINFU",
        "number": 1165,
        "title": "StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation",
        "normalized_title": "storydiffusion consistent self attention for long range image and video generation",
        "abstract": "For recent diffusion-based generative models, maintaining consistent content across a series of generated images, especially those containing subjects and complex details, presents a significant challenge. In this paper, we propose a simple but effective self-attention mechanism, termed Consistent Self-Attention, that boosts the consistency between the generated images. It can be used to augment pre-trained diffusion-based text-to-image models in a zero-shot manner. Based on the images with consistent content, we further show that our method can be extended to long range video generation by introducing a semantic space temporal motion prediction module, named Semantic Motion Predictor. It is trained to estimate the motion conditions between two provided images in the semantic spaces. This module converts the generated sequence of images into videos with smooth transitions and consistent subjects that are more stable than the modules based on latent spaces only, especially in the context of long video generation. By merging these two novel components, our framework, referred to as StoryDiffusion, can describe a text-based story with consistent images or videos encompassing a rich variety of contents. The proposed StoryDiffusion encompasses pioneering explorations in visual story generation with the presentation of images and videos, which we hope could inspire more research from the aspect of architectural modifications.",
        "authors": [
          "Yupeng Zhou",
          "Daquan Zhou",
          "Ming-Ming Cheng",
          "Jiashi Feng",
          "Qibin Hou"
        ],
        "decision": null,
        "pdate": 1727287653780,
        "cdate": 1714449082695
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "VFqzxhINFU",
          "forum": "VFqzxhINFU",
          "number": 1165,
          "title": "StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation",
          "normalized_title": "storydiffusion consistent self attention for long range image and video generation",
          "abstract": "For recent diffusion-based generative models, maintaining consistent content across a series of generated images, especially those containing subjects and complex details, presents a significant challenge. In this paper, we propose a simple but effective self-attention mechanism, termed Consistent Self-Attention, that boosts the consistency between the generated images. It can be used to augment pre-trained diffusion-based text-to-image models in a zero-shot manner. Based on the images with consistent content, we further show that our method can be extended to long range video generation by introducing a semantic space temporal motion prediction module, named Semantic Motion Predictor. It is trained to estimate the motion conditions between two provided images in the semantic spaces. This module converts the generated sequence of images into videos with smooth transitions and consistent subjects that are more stable than the modules based on latent spaces only, especially in the context of long video generation. By merging these two novel components, our framework, referred to as StoryDiffusion, can describe a text-based story with consistent images or videos encompassing a rich variety of contents. The proposed StoryDiffusion encompasses pioneering explorations in visual story generation with the presentation of images and videos, which we hope could inspire more research from the aspect of architectural modifications.",
          "authors": [
            "Yupeng Zhou",
            "Daquan Zhou",
            "Ming-Ming Cheng",
            "Jiashi Feng",
            "Qibin Hou"
          ],
          "decision": null,
          "pdate": 1727287653780,
          "cdate": 1714449082695
        }
      ]
    }
  },
  {
    "paper_id": 42634,
    "title": "IPO: Interpretable Prompt Optimization for Vision-Language Models",
    "abstract": "Pre-trained vision-language models like CLIP have remarkably adapted to various downstream tasks. Nonetheless, their performance heavily depends on the specificity of the input text prompts, which requires skillful prompt template engineering. Instead, current approaches to prompt optimization learn the prompts through gradient descent, where the prompts are treated as adjustable parameters. However, these methods tend to lead to overfitting of the base classes seen during training and produce prompts that are no longer understandable by humans.\nThis paper introduces a simple but interpretable prompt optimizer (IPO), that utilizes large language models (LLMs) to generate textual prompts dynamically.\nWe introduce a Prompt Optimization Prompt that not only guides LLMs in creating effective prompts but also stores past prompts with their performance metrics, providing rich in-context information. Additionally, we incorporate a large multimodal model (LMM) to condition on visual content by generating image descriptions, which enhance the interaction between textual and visual modalities. This allows for the creation of dataset-specific prompts that improve generalization performance, while maintaining human comprehension.\nExtensive testing across 11 datasets reveals that IPO not only improves the accuracy of existing gradient-descent-based prompt learning methods but also considerably enhances the interpretability of the generated prompts. By leveraging the strengths of LLMs, our approach ensures that the prompts remain human-understandable, thereby facilitating better transparency and oversight for vision-language models.",
    "introduction": "Vision-language models, trained on a diverse array of image-text pairs encapsulating a broad vocabulary of real-world concepts[1,2,3], have demonstrated notable adaptability across various downstream tasks[4,5,6,7]. These models perform zero-shot image classification by filling in a predefined prompt template (e.g., “a photo of a [CLASS]”) with specific class names for the text encoder. Despite their effective generalization to new tasks, the performance can be influenced by minor changes in the wording of prompt templates[8].\nInstead of manually creating hand-crafted prompts, recent developments in natural language processing[9,10]and computer vision[8,11,12,13]have proposed methods to learn a set of soft prompts with minimal labeled data.\nDespite the strides made in learning prompts, the current state of the art remains limited by its lack of interpretability and the overfitting problems on the base classes, which can be prohibitive in diverse and dynamic application environments. These limitations underscore the need for a more adaptable and user-friendly approach to prompt optimization in vision-language models.\n\n\nDrawing from recent advancements in using large language models (LLMs) as optimization tools[14], our paper, for the first time, incorporates these capabilities into vision-language modeling. Unlike gradient descent-based methods[8,11,13], which often fail to provide explanations for the generated prompts and tend to overfit on base classes, natural language-based methods enable LLMs to develop and refine solutions through continuous feedback iteratively. This approach improves interpretability in complex tasks like prompt optimization for vision-language models, making it easier for humans to understand the generated prompts. However, existing research on these methods[14,15,16]primarily addresses language tasks and has not yet explored their potential for integrating prompt optimization with an LLM in vision-language models.\n\n\nFigure 1:Comparison between traditional gradient-based prompt optimization (a) and our interpretable prompt optimization (b) for vision-language models. Traditional gradient descent-based prompt learning methods[8,11]treat the text prompt as learnable parametersV𝑉Vitalic_V. By minimizing the loss through gradient descent on the training set, an optimized promptV^^𝑉\\hat{V}over^ start_ARG italic_V end_ARGis obtained afterI𝐼Iitalic_Iiterations, which is not interpretable by humans. In contrast, our interpretable prompt optimization leverages an LLM as optimizer to optimize the loss and accuracy. AfterI𝐼Iitalic_Iiterations, the resulting optimized top prompt is effective and human-readable.\n\n\nTo address these challenges, this paper proposes an interpretable prompt optimizer (IPO) for vision-language models that leverages the capabilities of LLMs to generate and refine text prompts dynamically.\nFirst, we design a Prompt Optimization Prompt to prompt LLMs to generate more effective prompts that improve the accuracy of CLIP and reduce the loss in base classes.\nOur Prompt Optimization Prompt also stores past prompts along with their corresponding accuracy and loss as episodic memory, thereby providing richer in-context information to enable LLMs to generate more effective prompts.\nSecond, to incorporate image information within the Prompt Optimization Prompt, we propose using a large multimodal model (LMM) to generate descriptions of images in base classes that can be added to the Prompt Optimization Prompt.\nThis integration facilitates a more intuitive interaction between the textual and visual modalities, which\nallows the Prompt Optimization Prompt to utilize image information, thereby generating dataset-specific prompts to enhance the generalization performance of CLIP. The framework of our IPO, illustrated in Figure1, showcases the comparison between traditional gradient-based prompt optimization and the proposed interpretable prompt optimization.\nThird, the prompts generated by our optimizer are human-interpretable. For example, on the Food101 dataset[17], the initial prompt evolves from [“a photo of a [CLASS]”] to [“Categorize the image depicting a delicious and appetizing <CLASS> with remarkable visual qualities.”]. Our generated prompts perform 10.29% better in novel classes than the gradient-based method CoOP[8], reducing overfitting while maintaining interpretability.\n\n\nWe validated our IPO across 11 different datasets, demonstrating that it surpasses traditional gradient-based state-of-the-art methods in accuracy and excels in interpretability. Our approach generates human-comprehensible prompts that can be seamlessly integrated into existing vision-language models to enhance performance. We conducted rigorous comparative experiments to quantify the interpretability between gradient-based prompt learning and our method. We demonstrated the importance of specific keywords in our generated prompts and revealed that not all tokens learned through traditional prompt learning methods are essential.\n\n",
    "related_work": "Prompt learning in vision-language models.Prompt learning, originally introduced in the natural language processing community[18,19,20], involves applying a fixed function to input tokens to provide task instructions to the model. In the computer vision community, prompt learning has been explored in various forms, including textual prompt tuning[8,11,21,22,23,24], and prefix tuning[13,12,25,26,27,28,29].\n1) Prompt tuning mainly involves treating text prompts as learnable parameters, using a small amount of data to fine-tune these parameters. As pioneered by CoOp[8]and CoCoOp[11], which both fine-tune a CLIP vision-language model[30]for few-shot transfer by optimizing a continuous set of prompt vectors within its language branch. Bayesian prompt learning[21]formulated prompt learning as a variational inference problem and demonstrated its ability for unseen class generalization.\n2) Prefix tuning primarily involves adding learnable tokens to the text encoder[31], vision encoder[12,25], or both encoders[13,27,28,32]. These tokens are fine-tuned using a small amount of data. Note that these methods do not optimize the initial text prompts. Instead, they focus on enhancing the model’s understanding capabilities by integrating these additional, trainable tokens.\nOur method belongs to prompt tuning, but unlike previous approaches that use gradient descent to optimize prompts, we propose using LLMs to optimize prompts. Our method leverages the natural language capabilities of LLMs to iteratively refine feedback-based prompts, aiming to enhance both the effectiveness and the explainability of the prompts.\n\n\nLLMs as prompt optimizers.Several recent works explore the role of LLMs as prompt optimizers for NLP tasks. Some use LLMs to directly optimize the task instruction for in-context learning[33,34,14]. Other studies use LLMs to mutate prompts for evolutionary algorithms[35,36].\nHowever, to the best of our knowledge, no existing studies have investigated how LLMs could be used to optimize text prompts within vision-language models. This approach could potentially open up new avenues for integrating and enhancing the capabilities of vision-language models through more effective and contextually appropriate text prompts.\n\n\nMeta-prompting.Suzgun and Kalai[37]introduce meta-prompting to transform a single LLM into a versatile “conductor” capable of managing and integrating multiple independent LLM queries. By using high-level instructions, meta-prompting guides the LLM in decomposing complex tasks into smaller subtasks.\nThe core of OPRO[14]involves designing a meta-prompt for LLMs to optimize prompts for each task. This meta-prompt includes two key pieces of information: previously generated prompts with their corresponding training accuracies and a description of the optimization problem.\nSelf-select[38]leverages meta-prompting to optimize instruction selection. It considers a set of provided templates and chooses the most suitable template.\nMeta-prompting is related to instruction tuning[39]as both techniques provide high-level guidance to improve the performance and adaptability of LLMs. However, while instruction tuning focuses on fine-tuning models with a variety of tasks to improve generalization, meta-prompting offers the advantage of dynamically guiding the model to decompose and manage complex tasks in real-time.\nLiu et al.[40]proposes a method that utilizes LLMs as black-box optimizers for vision-language models, iteratively refining prompts based on in-context examples. Their approach focuses on leveraging ChatGPT to improve prompt templates for visual classification tasks.\nMirza et al.[41]explores a different aspect of prompt optimization by focusing on zero-shot vision-language models.\nOur Prompt Optimization Prompt is akin to meta-prompting, it stores past prompts along with their corresponding accuracy and loss, thereby providing richer in-context information to enable LLMs to generate more effective prompts. Different from prior meta-prompting, our Prompt Optimization Prompt generates prompts beyond LLMs for vision-language models.\n\n",
    "origin_citations": [
      "Autoprompt: Eliciting knowledge from language models with automatically generated prompts.",
      "How can we know what language models know?",
      "Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.",
      "Learning to prompt for vision-language models.",
      "Conditional prompt learning for vision-language models.",
      "Bayesian prompt learning for image-language model generalization.",
      "Prompt distribution learning.",
      "Prompt-aligned gradient for prompt tuning.",
      "Visual-language prompt tuning with knowledge-guided context optimization.",
      "Maple: Multi-modal prompt learning.",
      "Visual prompt tuning.",
      "Exploring visual prompts for adapting large-scale models.",
      "Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection.",
      "Promptkd: Unsupervised prompt distillation for vision-language models.",
      "Self-regulating prompts: Foundational model adaptation without forgetting.",
      "Any-shift prompting for generalization over distributions.",
      "Learning transferable visual models from natural language supervision.",
      "Learning to prompt with text only supervision for vision-language models.",
      "Consistency-guided prompt learning for vision-language models.",
      "Large language models are human-level prompt engineers.",
      "Automatic prompt optimization with\" gradient descent\" and beam search.",
      "Large language models as optimizers.",
      "Connecting large language models with evolutionary algorithms yields powerful prompt optimizers.",
      "Promptbreeder: Self-referential self-improvement via prompt evolution.",
      "Meta-prompting: Enhancing language models with task-agnostic scaffolding.",
      "Self-select: Optimizing instruction selection for large language models.",
      "Instruction tuning for large language models: A survey.",
      "Language models as black-box optimizers for vision-language models.",
      "Meta-prompting for automating zero-shot visual recognition with llms."
    ],
    "citations": [
      7787,
      2976,
      408717,
      3306,
      3303,
      460999,
      378511,
      145526,
      494202,
      460646,
      3304,
      1038295,
      311643,
      291131,
      184968,
      907528,
      1639,
      485313,
      307953,
      3380,
      116775,
      129325,
      704408,
      477645,
      124897,
      405844,
      128125,
      427035
    ],
    "extract_rate": 0.9655172413793104,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 79,
      "Final_Score": 83.5,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 79\nFinal_Score: 84"
    },
    "overall_score": 84.7,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "WPPC7FHtaM",
        "forum": "WPPC7FHtaM",
        "number": 2578,
        "title": "IPO: Interpretable Prompt Optimization for Vision-Language Models",
        "normalized_title": "ipo interpretable prompt optimization for vision language models",
        "abstract": "Pre-trained vision-language models like CLIP have remarkably adapted to various downstream tasks. Nonetheless, their performance heavily depends on the specificity of the input text prompts, which requires skillful prompt template engineering. Instead, current approaches to prompt optimization learn the prompts through gradient descent, where the prompts are treated as adjustable parameters. However, these methods tend to lead to overfitting of the base classes seen during training and produce prompts that are no longer understandable by humans. This paper introduces a simple but interpretable prompt optimizer (IPO), that utilizes large language models (LLMs) to generate textual prompts dynamically. We introduce a Prompt Optimization Prompt that not only guides LLMs in creating effective prompts but also stores past prompts with their performance metrics, providing rich in-context information. Additionally, we incorporate a large multimodal model (LMM) to condition on visual content by generating image descriptions, which enhance the interaction between textual and visual modalities. This allows for the creation of dataset-specific prompts that improve generalization performance, while maintaining human comprehension. Extensive testing across 11 datasets reveals that IPO not only improves the accuracy of existing gradient-descent-based prompt learning methods but also considerably enhances the interpretability of the generated prompts. By leveraging the strengths of LLMs, our approach ensures that the prompts remain human-understandable, thereby facilitating better transparency and oversight for vision-language models.",
        "authors": [
          "Yingjun Du",
          "Wenfang Sun",
          "Cees G. M. Snoek"
        ],
        "decision": null,
        "pdate": 1727287694333,
        "cdate": 1715106416243
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "WPPC7FHtaM",
          "forum": "WPPC7FHtaM",
          "number": 2578,
          "title": "IPO: Interpretable Prompt Optimization for Vision-Language Models",
          "normalized_title": "ipo interpretable prompt optimization for vision language models",
          "abstract": "Pre-trained vision-language models like CLIP have remarkably adapted to various downstream tasks. Nonetheless, their performance heavily depends on the specificity of the input text prompts, which requires skillful prompt template engineering. Instead, current approaches to prompt optimization learn the prompts through gradient descent, where the prompts are treated as adjustable parameters. However, these methods tend to lead to overfitting of the base classes seen during training and produce prompts that are no longer understandable by humans. This paper introduces a simple but interpretable prompt optimizer (IPO), that utilizes large language models (LLMs) to generate textual prompts dynamically. We introduce a Prompt Optimization Prompt that not only guides LLMs in creating effective prompts but also stores past prompts with their performance metrics, providing rich in-context information. Additionally, we incorporate a large multimodal model (LMM) to condition on visual content by generating image descriptions, which enhance the interaction between textual and visual modalities. This allows for the creation of dataset-specific prompts that improve generalization performance, while maintaining human comprehension. Extensive testing across 11 datasets reveals that IPO not only improves the accuracy of existing gradient-descent-based prompt learning methods but also considerably enhances the interpretability of the generated prompts. By leveraging the strengths of LLMs, our approach ensures that the prompts remain human-understandable, thereby facilitating better transparency and oversight for vision-language models.",
          "authors": [
            "Yingjun Du",
            "Wenfang Sun",
            "Cees G. M. Snoek"
          ],
          "decision": null,
          "pdate": 1727287694333,
          "cdate": 1715106416243
        }
      ]
    }
  },
  {
    "paper_id": 25967,
    "title": "Zero-to-Hero: Enhancing Zero-Shot Novel View Synthesis via Attention Map Filtering",
    "abstract": "Generating realistic images from arbitrary views based on a single source image remains a significant challenge in computer vision, with broad applications ranging from e-commerce to immersive virtual experiences. Recent advancements in diffusion models, particularly the Zero-1-to-3 model, have been widely adopted for generating plausible views, videos, and 3D models. However, these models still struggle with inconsistencies and implausibility in new views generation, especially for challenging changes in viewpoint. In this work, we propose Zero-to-Hero, a novel test-time approach that enhances view synthesis by manipulating attention maps during the denoising process of Zero-1-to-3. By drawing an analogy between the denoising process and stochastic gradient descent (SGD), we implement a filtering mechanism that aggregates attention maps, enhancing generation reliability and authenticity. This process improves geometric consistency without requiring retraining or significant computational resources. Additionally, we modify the self-attention mechanism to integrate information from the source view, reducing shape distortions. These processes are further supported by a specialized sampling schedule. Experimental results demonstrate substantial improvements in fidelity and consistency, validated on a diverse set of out-of-distribution objects. Additionally, we demonstrate the general applicability and effectiveness of Zero-to-Hero in multi-view, and image generation conditioned on semantic maps and pose.",
    "introduction": "The pursuit of realistic image synthesis at arbitrary views, given only a single source image, has long been a cornerstone challenge in computer vision and graphics. This technology can cater to countless applications, such as interactive product inspection, robot-scene interaction, and immersive virtual experiences. In this work, we aim to advance this important line of research by improving the generation of novel views that are plausible and faithful to the input image. A recent, promising approach, Zero-1-to-3[27]has developed a foundation model to synthesize novel views based on a single source image and a target view angle. By leveraging a pre-trained, image-conditioned stable diffusion model backbone[3], fine-tuned with target camera poses, and trained on paired source and target views from a vast collection of 3D models[10,9], Zero-1-to-3 can generalize beyond its training set and generate plausible novel views. As a result, this model has quickly gained popularity, inspiring subsequent work in 3D and 4D scene generation[8,24,39,28,26,25,57,33,45,35,20].\n\n\nDespite its remarkable ability, Zero-1-to-3 has been observed to generate views that are implausible, or inconsistent with the input object in terms of shape and appearance[24,8]. Previous works have tried to mitigate these issues by retraining diffusion models with more data[9]or to generate multiple views[39,25,26,24,8,28,56]. Despite substantial improvement, both approaches are resource-intensive due to the required re-training on large-scale 3D datasets. Another line of work attempts to consolidate inconsistencies across multiple generated views through a 3D representation like NeRF[14,30]. However, direct aggregation often results in blurry outputs, as observed by[24]. Instead, ViVid-1-to-3[24]employed a multiview representation that naturally supports the use of a video foundation model. Nevertheless, this approach requires generating the entire trajectory from the source to the target view, adding significant complexity and computational overhead. Notably, the denoising process in Zero-1-to-3 remains unchanged.\n\n\nIn this work, we propose Zero-to-Hero, a novel test-time technique that addresses view synthesis artifacts through attention map manipulation. Recognizing attention maps as crucial for latent predictions, we hypothesize that enhancing robustness in attention maps predictions can significantly reduce generation misalignment. To achieve this, we draw an analogy between the denoising process in diffusion models and stochastic gradient descent (SGD) optimization of neural networks. Specifically, we relate network weights that predict local gradients at each optimization step, based on sampled training examples and labels, to the denoising network’s attention maps that predict latent representations from sampled noise at each denoising step. In this work, we view the generation (denoising) process as an unrolled optimization, with attention maps as parameters of a score prediction model. Inspired by gradient aggregation and weight-averaging techniques that improve prediction robustness (e.g., consistency training[47]), we propose a filtering mechanism to enhance attention map reliability. This mechanism comprises iterative map aggregation within denoising steps and map averaging across denoising steps. The result is more reliable maps, particularly during the early denoising stages when coarse output shapes are determined, leading to more plausible and realistic views.\n\n\nTo further promote consistency with the input, we modify the self-attention operation by running a parallel generation branch using the identity pose, incorporating its keys and values into the attention layer of the target view. Unlike previous applications of this technique[29,5,1]), we find it beneficial in view synthesis to limit its use to the early denoising stages, preventing shape distortions. Our unique denoising procedure is further complemented by a novel sampling schedule that emphasizes early and late denoising stages, maximizing performance. Our main contributions are as follows:•To address the main limitations of the Zero-1-to-3 model, we perform an in-depth analysis and identify self-attention maps as the main candidate for correcting generation artifacts.•We establish a conceptual analogy between model weights in stochastic gradient descent-based network training and the role of attentions map updates during generation of a denoising diffusion model. Based on this, we propose a simple yet powerful attention map filtering process resulting in enhanced target shape generation. We supplement our filtering technique with identity view information injection and a specialized sampling schedule.•Our method requires no additional training, and it avoids the overhead of external models or generating multiple views.\n\n\nThrough comprehensive experiments on out-of-distribution objects, we demonstrate that our technique robustifies Zero-1-to-3 and its extended version, Zero123-XL, leading to views that are more faithful to both the input image and desired camera transformation. Our results show significant and consistent improvements across both appearance and shape evaluation metrics. Additionally, we find that Zero-to-Hero naturally generalizes to additional tasks including multi-view, and image generation conditioned on semantic maps and pose. In all cases, we observed significant improvement in condition following and visual quality.\n\n",
    "related_work": "2.1Novel View Synthesis with Diffusion ModelsDiffusion models have dominated various generative applications[17,11,38,36,41]. Particularly, novel-view synthesis, as a core of applications like augmented reality and simulations, naturally enjoys the benefits of high-fidelity zero-shot synthesis with diffusion models. One line of works[51,27,44,22,54,24]is to generate a novel-view image given a source image (i.e.,image-to-image). These approaches typically involve training a diffusion model conditioned on both an arbitrary camera pose and the source view. For instance, the representative work, Zero-1-to-3, fine-tunes a pre-trained Stable Diffusion model[36]by replacing the text prompt with camera pose and CLIP features[34].\nMoreover, another research trajectory[32,21,14,7,46]proposes generating a 3D representation from a single image (i.e., image-to-3D), which allows for sampling desired views from these 3D models. Our method, Zero-to-Hero, builds on the first approach (specifically Zero-1-to-3 and Zero123-XL) and distinguishes itself by eliminating the need for extensive training. Instead, it offers a test-time, plug-and-play approach that significantly enhances visual quality and consistency.\n\n\n2.2Test-Time Refinement in Diffusion-Based GenerationA common test-time refinement strategy in diffusion generation is leveraging guidance[29,2,18]to direct the sampling process with additional conditions. For example, Repaint[29]utilizes a mask-then-renoise strategy to refine the generation results. Repaint also introduces a per-step resampling technique, where given a horizon-size h, a latentztsubscript𝑧𝑡z_{t}italic_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTis re-noised tozt+hsubscript𝑧𝑡ℎz_{t+h}italic_z start_POSTSUBSCRIPT italic_t + italic_h end_POSTSUBSCRIPTand then denoised again toztsubscript𝑧𝑡z_{t}italic_z start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTmultiple times. They observe that resampling helps to generate more harmonized outputs, given an external guidance or condition.\nRestart[55]offers a sampling algorithm based on a variation of resampling within a chosen interval of steps.\nOur method is inspired by the strategy of per-step resampling. We show that it serves as a powerful correction mechanism throughout the generation process, even when no external guidance or condition is provided.\n\n\n2.3Attention Map Manipulation in Diffusion ModelsStable Diffusion[36]utilizes attention to enforce the condition information onto the generated results. Previous works demonstrate that manipulating the attention operation can achieve new capabilities[49,5,1,57]. For example, MasaCtrl[5]uses Mutual Self-Attention where source and target images are generated jointly while sharing information, by injecting source image keys and values to the target through self-attention.\nHere we employ Mutual Self-Attention in the context of novel view synthesis. Differently to prior works we find it beneficial to limit its use to the early denoising stages.\n\n",
    "origin_citations": [
      "Denoising diffusion probabilistic models.",
      "Diffusion models beat gans on image synthesis.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "High-resolution image synthesis with latent diffusion models.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Novel view synthesis with diffusion models.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Viewfusion: Learning composable diffusion models for novel view synthesis.",
      "invs: Repurposing diffusion inpainters for novel view synthesis.",
      "3difftection: 3d object detection with geometry-aware diffusion features.",
      "Vivid-1-to-3: Novel view synthesis with video diffusion models.",
      "Learning transferable visual models from natural language supervision.",
      "Point-e: A system for generating 3d point clouds from complex prompts.",
      "Shap-e: Generating conditional 3d implicit functions.",
      "Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion.",
      "Generative novel view synthesis with 3d-aware diffusion models.",
      "Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior.",
      "Repaint: Inpainting using denoising diffusion probabilistic models.",
      "Universal guidance for diffusion models.",
      "Classifier-free diffusion guidance.",
      "Restart sampling for improving generative processes.",
      "Plug-and-play diffusion features for text-driven image-to-image translation.",
      "Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing.",
      "Cross-image attention for zero-shot appearance transfer.",
      "Repaint123: Fast and high-quality one image to 3d generation with progressive controllable 2d repainting."
    ],
    "citations": [
      4827,
      165,
      5926,
      2240,
      176197,
      460675,
      495077,
      100519,
      313014,
      295040,
      274229,
      1639,
      8038,
      117011,
      227198,
      20037,
      493794,
      8039,
      488851,
      5922,
      48098,
      243990,
      54551,
      295511,
      348085
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 86,
      "Publication_Potential": 89,
      "Final_Score": 86.89999999999999,
      "raw_text": "Content_Quality: 86\nPublication_Potential: 89\nFinal_Score: 87"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 78,
      "Final_Score": 83.19999999999999,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 78\nFinal_Score: 83"
    },
    "overall_score": 84.67999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "3uQtNWNTwz",
        "forum": "3uQtNWNTwz",
        "number": 375,
        "title": "Zero-to-Hero: Enhancing Zero-Shot Novel View Synthesis via Attention Map Filtering",
        "normalized_title": "zero to hero enhancing zero shot novel view synthesis via attention map filtering",
        "abstract": "Generating realistic images from arbitrary views based on a single source image remains a significant challenge in computer vision, with broad applications ranging from e-commerce to immersive virtual experiences. Recent advancements in diffusion models, particularly the Zero-1-to-3 model, have been widely adopted for generating plausible views, videos, and 3D models. However, these models still struggle with inconsistencies and implausibility in new views generation, especially for challenging changes in viewpoint. In this work, we propose Zero-to-Hero, a novel test-time approach that enhances view synthesis by manipulating attention maps during the denoising process of Zero-1-to-3. By drawing an analogy between the denoising process and stochastic gradient descent (SGD), we implement a filtering mechanism that aggregates attention maps, enhancing generation reliability and authenticity. This process improves geometric consistency without requiring retraining or significant computational resources. Additionally, we modify the self-attention mechanism to integrate information from the source view, reducing shape distortions. These processes are further supported by a specialized sampling schedule. Experimental results demonstrate substantial improvements in fidelity and consistency, validated on a diverse set of out-of-distribution objects. Additionally, we demonstrate the general applicability and effectiveness of Zero-to-Hero in multi-view, and image generation conditioned on semantic maps and pose.",
        "authors": [
          "Ido Sobol",
          "Chenfeng Xu",
          "Or Litany"
        ],
        "decision": null,
        "pdate": 1727287636376,
        "cdate": 1713878864104
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "3uQtNWNTwz",
          "forum": "3uQtNWNTwz",
          "number": 375,
          "title": "Zero-to-Hero: Enhancing Zero-Shot Novel View Synthesis via Attention Map Filtering",
          "normalized_title": "zero to hero enhancing zero shot novel view synthesis via attention map filtering",
          "abstract": "Generating realistic images from arbitrary views based on a single source image remains a significant challenge in computer vision, with broad applications ranging from e-commerce to immersive virtual experiences. Recent advancements in diffusion models, particularly the Zero-1-to-3 model, have been widely adopted for generating plausible views, videos, and 3D models. However, these models still struggle with inconsistencies and implausibility in new views generation, especially for challenging changes in viewpoint. In this work, we propose Zero-to-Hero, a novel test-time approach that enhances view synthesis by manipulating attention maps during the denoising process of Zero-1-to-3. By drawing an analogy between the denoising process and stochastic gradient descent (SGD), we implement a filtering mechanism that aggregates attention maps, enhancing generation reliability and authenticity. This process improves geometric consistency without requiring retraining or significant computational resources. Additionally, we modify the self-attention mechanism to integrate information from the source view, reducing shape distortions. These processes are further supported by a specialized sampling schedule. Experimental results demonstrate substantial improvements in fidelity and consistency, validated on a diverse set of out-of-distribution objects. Additionally, we demonstrate the general applicability and effectiveness of Zero-to-Hero in multi-view, and image generation conditioned on semantic maps and pose.",
          "authors": [
            "Ido Sobol",
            "Chenfeng Xu",
            "Or Litany"
          ],
          "decision": null,
          "pdate": 1727287636376,
          "cdate": 1713878864104
        }
      ]
    }
  },
  {
    "paper_id": 28220,
    "title": "Ctrl-X: Controlling Structure and Appearance for Text-To-Image Generation Without Guidance",
    "abstract": "Recent controllable generation approaches such as FreeControl[24]and Diffusion Self-Guidance[7]bring fine-grained spatial and appearance control to text-to-image (T2I) diffusion models without training auxiliary modules. However, these methods optimize the latent embedding for each type of score function with longer diffusion steps, making the generation process time-consuming and limiting their flexibility and use. This work presentsCtrl-X, a simple framework for T2I diffusion controlling structure and appearance without additional training or guidance.Ctrl-Xdesigns feed-forward structure control to enable the structure alignment with a structure image and semantic-aware appearance transfer to facilitate the appearance transfer from a user-input image. Extensive qualitative and quantitative experiments illustrate the superior performance ofCtrl-Xon various condition inputs and model checkpoints. In particular,Ctrl-Xsupports novel structure and appearance control with arbitrary condition images of any modality, exhibits superior image quality and appearance transfer compared to existing works, and provides instant plug-and-play functionality to any T2I and text-to-video (T2V) diffusion model.",
    "introduction": "The rapid advancement of large text-to-image (T2I) generative models has made it possible to generate high-quality images with just one text prompt. However, it remains challenging to specify the exact concepts that can accurately reflect human intents using only textual descriptions. Recent approaches like ControlNet[44]and IP-Adapter[43]have enabled controllable image generation upon pretrained T2I diffusion models regarding structure and appearance, respectively. Despite the impressive results in controllable generation, these approaches[44,25,46,20]require fine-tuning the entire generative model or training auxiliary modules on large amounts of paired data.\n\n\nTraining-free approaches[7,24,4]have been proposed to address the high overhead associated with additional training stages. These methods optimize the latent embedding across diffusion steps using specially designed score functions to achieve finer-grained control than text alone with a process called guidance. Although training-free approaches avoid the training cost, they significantly increase computing time and required GPU memory in the inference stage due to the additional backpropagation over the diffusion network. They also require sampling steps that are2222–20202020times longer. Furthermore, as the expected latent distribution of each time step is predefined for each diffusion model, it is critical to tune the guidance weight delicately for each score function; Otherwise, the latent might be out-of-distribution and lead to artifacts and reduced image quality.\n\n\nTo tackle these limitations, we presentCtrl-X, a simpletraining-freeandguidance-freeframework for T2I diffusion with structure and appearance control. We name our method “Ctrl-X” because we reformulate the controllable generation problem by ‘cutting’ (and ‘pasting’) two tasks together: spatial structure preservation and semantic-aware stylization. Our insight is that diffusion feature maps capture rich spatial structure and high-level appearance from early diffusion steps sufficient for structure and appearance control without guidance. To this end,Ctrl-Xemploys feature injection and spatially-aware normalization in the attention layers to facilitate structure and appearance alignment with user-provided images. By being guidance-free,Ctrl-Xeliminates additional optimization overhead and sampling steps, resulting in a35353535-fold increase in inference speed compared to guidance-based methods. Figure1shows sample generation results. Moreover,Ctrl-Xsupports arbitrary structure conditions beyond natural images and can be applied to any T2I and even text-to-video (T2V) diffusion models. Extensive quantitative and qualitative experiments, along with a user study, demonstrate the superior image quality and appearance alignment of our method over prior works.\n\n\nWe summarize our contributions as follows:\n\n\n1.We presentCtrl-X, a simple plug-and-play method that builds on pretrained text-to-image diffusion models to provide disentangled and zero-shot control of structure and appearance during the generation process requiring no additional training or guidance.2.Ctrl-Xpresents the first universal guidance-free solution that supports multiple conditional signals (structure and appearance) and model architectures (e.g.text-to-image and text-to-video).3.Our method demonstrates superior results in comparison to previous training-based and guidance-based baselines (e.g.ControlNet + IP-Adapter[44,43]and FreeControl[24]) in terms of condition alignment, text-image alignment, and image quality.\n\n",
    "related_work": "Diffusion structure control.Previous spatial structure control methods can be categorized into two types (training-basedvs.training-free) based on whether they require training on paired data.\n\n\nTraining-based structure control methodsrequire paired condition-image data to train additional modules or fine-tune the entire diffusion network to facilitate generation from spatial conditions[44,25,20,46,42,3,47,38,49]. While pixel-level spatial control can be achieved with this approach, a significant drawback is needing a large number of condition-image pairs as training data. Although some condition data can be generated from pretrained annotators (e.g.depth and segmentation maps), other condition data is difficult to obtain from given images (e.g.3D mesh, point cloud), making these conditions challenging to follow. Compared to these training-based methods,Ctrl-Xsupports conditions where paired data is challenging to obtain, making it a more flexible and effective solution.\n\n\nTraining-free structure control methodstypically focus on specific conditions. For example, R&B[40]facilitates bounding-box guided control with region-aware guidance, and DenseDiffusion[17]generates images with sparse segmentation map conditions by manipulating the attention weights. Universal Guidance[4]employs various pretrained classifiers to support multiple types of condition signals. FreeControl[24]analyzes semantic correspondence in the subspace of diffusion features and harnesses it to support spatial control from any visual condition. While these approaches do not require training data, they usually need to compute the gradient of the latent to lower an auxiliary loss, which requires substantial computing time and GPU memory. In contrast,Ctrl-Xrequires no guidance at the inference stage and controls structure via direct feature injections, enabling faster and more robust image generation with spatial control.\n\n\nDiffusion appearance control.Existing appearance control methods that build upon pretrained diffusion models can also similarly be categorized into two types (training-basedvs.training-free).\n\n\nTraining-based appearance control methodscan be divided into two categories: Those trained to handle any image prompt and those overfitting to a single instance. The first category[44,25,43,38]trains additional image encoders or adapters to align the generated process with the structure or appearance from the reference image. The second category[30,14,8,2,26,31]is typically applied to customized visual content creation by finetuning a pretrained text-to-image model on a small set of images or binding special tokens to each instance. The main limitation of these methods is that the additional training required makes them unscalable. However,Ctrl-Xoffers a scalable solution to transfer appearance from any instance without training data.\n\n\nTraining-free appearance control methodsgenerally follow two approaches: One approach[1,5,41]manipulates self-attention features using pixel-level dense correspondence between the generated image and the target appearance, and the other[7,24]extracts appearance embeddings from the diffusion network and transfers the appearance by guiding the diffusion process towards the target appearance embedding. A key limitation of these approaches is that a single text-controlled target cannot fully capture the details of the target image, and the latter methods require additional optimization steps. By contrast, our method exploits the spatial correspondence of self-attention layers to achieve semantically-aware appearance transfer without targeting specific subjects.\n\n\nFigure 2:Visualizing early diffusion features.Using20202020real, generated, and condition images of animals, we extract Stable Diffusion XL[27]features right after decoder layer00convolution. We visualize the top three principal components computed for each time step across all images.t=961𝑡961t=961italic_t = 961to881881881881correspond to inference steps1111to5555of the DDIM scheduler with50505050time steps. We obtain𝐱tsubscript𝐱𝑡\\mathbf{x}_{t}bold_x start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTby directly adding Gaussian noise to each clean image𝐱0subscript𝐱0\\mathbf{x}_{0}bold_x start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPTvia the diffusion forward process.\n\n",
    "origin_citations": [
      "Adding conditional control to text-to-image diffusion models.",
      "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.",
      "Gligen: Open-set grounded text-to-image generation.",
      "Uni-controlnet: All-in-one control to text-to-image diffusion models.",
      "Reco: Region-controlled text-to-image generation.",
      "Spatext: Spatio-textual representation for controllable image generation.",
      "Layoutdiffusion: Controllable diffusion model for layout-to-image generation.",
      "Instancediffusion: Instance-level control for image generation, 2024.",
      "Migc: Multi-instance generation controller for text-to-image synthesis, 2024.",
      "R&b: Region and boundary aware zero-shot grounded text-to-image generation.",
      "Dense text-to-image generation with attention modulation.",
      "Universal guidance for diffusion models.",
      "Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition.",
      "Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.",
      "Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation.",
      "LoRA: Low-rank adaptation of large language models.",
      "An image is worth one word: Personalizing text-to-image generation using textual inversion.",
      "Break-a-scene: Extracting multiple concepts from a single image.",
      "Orthogonal adaptation for modular customization of diffusion models, 2023.",
      "Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models, 2023b.",
      "Cross-image attention for zero-shot appearance transfer.",
      "Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing.",
      "Inversion-free image editing with natural language.",
      "Diffusion self-guidance for controllable image generation.",
      "SDXL: Improving latent diffusion models for high-resolution image synthesis."
    ],
    "citations": [
      489610,
      228018,
      214623,
      258208,
      243711,
      243181,
      492337,
      100292,
      99218,
      320168,
      404850,
      488851,
      271106,
      179737,
      67241,
      1589,
      236013,
      258217,
      273677,
      295511,
      54551,
      272681,
      308042,
      187204
    ],
    "extract_rate": 0.96,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 84.64,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "ZulWEWQOp9",
        "forum": "ZulWEWQOp9",
        "number": 3423,
        "title": "Ctrl-X: Controlling Structure and Appearance for Text-To-Image Generation Without Guidance",
        "normalized_title": "ctrl x controlling structure and appearance for text to image generation without guidance",
        "abstract": "Recent controllable generation approaches such as FreeControl and Diffusion Self-Guidance bring fine-grained spatial and appearance control to text-to-image (T2I) diffusion models without training auxiliary modules. However, these methods optimize the latent embedding for each type of score function with longer diffusion steps, making the generation process time-consuming and limiting their flexibility and use. This work presents *Ctrl-X*, a simple framework for T2I diffusion controlling structure and appearance without additional training or guidance. Ctrl-X designs feed-forward structure control to enable the structure alignment with a structure image and semantic-aware appearance transfer to facilitate the appearance transfer from a user-input image. Extensive qualitative and quantitative experiments illustrate the superior performance of Ctrl-X on various condition inputs and model checkpoints. In particular, Ctrl-X supports novel structure and appearance control with arbitrary condition images of any modality, exhibits superior image quality and appearance transfer compared to existing works, and provides instant plug-and-play functionality to any T2I and text-to-video (T2V) diffusion model. See our project page for the code and an overview of the results: https://genforce.github.io/ctrl-x",
        "authors": [
          "Kuan Heng Lin",
          "Sicheng Mo",
          "Ben Klingher",
          "Fangzhou Mu",
          "Bolei Zhou"
        ],
        "decision": null,
        "pdate": 1727287719824,
        "cdate": 1715273558747
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "ZulWEWQOp9",
          "forum": "ZulWEWQOp9",
          "number": 3423,
          "title": "Ctrl-X: Controlling Structure and Appearance for Text-To-Image Generation Without Guidance",
          "normalized_title": "ctrl x controlling structure and appearance for text to image generation without guidance",
          "abstract": "Recent controllable generation approaches such as FreeControl and Diffusion Self-Guidance bring fine-grained spatial and appearance control to text-to-image (T2I) diffusion models without training auxiliary modules. However, these methods optimize the latent embedding for each type of score function with longer diffusion steps, making the generation process time-consuming and limiting their flexibility and use. This work presents *Ctrl-X*, a simple framework for T2I diffusion controlling structure and appearance without additional training or guidance. Ctrl-X designs feed-forward structure control to enable the structure alignment with a structure image and semantic-aware appearance transfer to facilitate the appearance transfer from a user-input image. Extensive qualitative and quantitative experiments illustrate the superior performance of Ctrl-X on various condition inputs and model checkpoints. In particular, Ctrl-X supports novel structure and appearance control with arbitrary condition images of any modality, exhibits superior image quality and appearance transfer compared to existing works, and provides instant plug-and-play functionality to any T2I and text-to-video (T2V) diffusion model. See our project page for the code and an overview of the results: https://genforce.github.io/ctrl-x",
          "authors": [
            "Kuan Heng Lin",
            "Sicheng Mo",
            "Ben Klingher",
            "Fangzhou Mu",
            "Bolei Zhou"
          ],
          "decision": null,
          "pdate": 1727287719824,
          "cdate": 1715273558747
        }
      ]
    }
  },
  {
    "paper_id": 41638,
    "title": "Context-Parametric Inversion:Why Instruction Finetuning May Not Actually Improve Context Reliance",
    "abstract": "standard practice when using large language models is for users to supplement their instruction with an input context containing new information for the model to process. However, models struggle to reliably follow the input context, especially when it conflicts with their parametric knowledge from pretraining.\nIn-principle, one would expect models to adapt to the user context better after instruction finetuning, particularly when handling knowledge conflicts.\nHowever, we observe a surprising failure mode: during instruction tuning, the context reliance under knowledge conflicts initially increases as expected, but thengradually decreases as instruction finetuning progresses. This happens while the performance on standard benchmarks keeps on increasing far after this drop. We call this phenomenoncontext-parametric inversionand observe it across multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across different model families like Llama, Mistral, and Pythia. We perform various controlled studies and theoretical analysis to show that context-parametric inversion occurs due to examples in the instruction finetuning data where the input context provides information that aligns with model’s parametric knowledge.\nOur analysis suggests some natural mitigation strategies with limited but insightful gains, and serves as a useful starting point in addressing this deficiency in instruction finetuning.",
    "introduction": "Large language models (LLMs) are widely used for a variety of tasks, many of which require effectively balancing the parametric knowledge embedded in their weights with the information provided through the input context. A persistent challenge, however, is their tendency to over-rely on this parametric knowledge, even when it contradicts the contextual information. This over-reliance not only hinders the ability to update model facts or mitigate hallucinations with augmented contexts, but also exacerbates issues like biases and failures to reliably follow user instructions—particularly when they involve the generation of atypical or unconventional text(McKenzie et al.,2024; Qiu et al.,2023; Adlakha et al.,2024).\n\n\nThis tension between the context and the model’s parametric knowledge has been commonly studied under the moniker ofknowledge conflicts, with existing works exploring various strategies like contrastive decoding or augmented finetuning to encourage models to adhere to the context(Shi et al.,2023; Yuan et al.,2024; Longpre et al.,2022; Chen et al.,2022). However, this challenge still persists (see §B) and is unlikely to go away just with scale(McKenzie et al.,2024). We have limited understanding of the underlying dynamics that drives models to rely heavily on parametric knowledge despite contradictory contextual information.\n\n\nIn this work, we study the effect of instruction finetuning (IFT)—a staple part of the LLM pipeline—on the ability to override pretrained knowledge through the context. IFT seeks to enhance the model’s ability to assist with user queries or instructions. Oftentimes, these user queries and instructions are context-based where the context provides critical information needed to complete the task. For instance, an instruction “What is the total price of my trip to Hawaii?” operates on a context “Context: [Itinerary List]”, and an instruction “Rank these famous soccer players based on these scores” could contain a context like: “[Scores Table].” In these circumstances, instruction tuned models are responsible for appropriately leveraging the input context to answer the user query, instead of purely relying on parametric knowledge. In principle, instruction tuning should improve the ability of models to use context over parametric knowledge. However, we make an intriguing observation where in the presence of a conflict between parametric and context information,the model’s reliance on context initially increases as expected but surprisingly starts decreasing as IFT progresses.\n\n\n(a)(b)Figure 1:(a)Context-Parametric InversionIn the presence of conflicts between parametric and context information, a model’s reliance on context first increases and then decreases during the process of instruction finetuning. (b) Instruction tuning data includes both context-critical examples, where the context is essential, and non-context-critical examples, where responses can be generated using either the context or the model’s parametric knowledge. It is this latter group of non-context-critical examples that leads to a decline in context reliance (§4.3).\n\n\nWe measure the context reliance by designing inputs contexts that suggest a fictional answer to a user query different from facts memorized from the pretraining corpus (§3.2). We evaluate context reliance across the IFT trajectory on multiple IFT datasets—TULU, Alpaca or UltraChat and across multiple model families—Llama, Pythia and Mistral. Across all these settings, we see that context reliance initially increases and then decreases, a phenomenon we callcontext-parametric inversion. In fact, this drop begins in early timesteps of IFT, while the performance on standard benchmarks (e.g., MMLU, GSM8k, SQuAD) keeps on increasing far after this drop.\nFor example, as shown in Figure1(a), the context reliance of Llama2-7B (as measured on knowledge conflict datasets (§3.2)) increases from 30% to 60% initially with IFT. However, it start dropping as the finetuning progresses further, dipping to around35%percent3535\\%35 %.\n\n\nWhy do we observe context-parametric inversion with instruction tuning? The initial increase is apriori expected, as instruction tuning datasets typically contains datapoints that require utilizing context to complete the instruction, as discussed above. We perform controlled experiments to understand the subsequent decrease which is unexpected and detrimental. First, we observe that reliance on parametric knowledge over context extends to facts beyond those potentially “memorized” during the instruction finetuning process. Hence, this drop in context reliance is not simply due to the model acquiring new facts. Common instruction tuning datasets typically contain some datapoints that are purely about recall of pretrained knowledge, and do not involve context-dependent instructions. Could the drop be attributed to the presence of such points? We curate the datasets to only include context-dependent points butstillsee a drop in context reliance after an initial increase.\n\n\nWe analyze this phenomenon theoretically in a simple setting and uncover the optimization dynamic that explains context-parametric inversion. We can partition a generic dataset containing context-dependent datapoints into two categories: (i)context-criticaldatapoints where context provides key information needed to answer a user query that the model does not know beforehand (Fig.1(b)), and (ii)non-context-criticaldatapoints where the context is approximately redundant with model’s parametric knowledge. We formally define these categories in §4.3.\n\n\nIn the early stages of training, context-critical points tend to have higher loss and therefore dominate the gradient signal, driving the model to focus on the context. However, as training progresses, the loss on context-critical points decreases, and the non-context-critical points dominate the gradient. At later stages, the model leverages its parametric knowledge to further reduce the loss on these non-context-critical points, thus reducing the context reliance. We formalize this dynamic in a one-layer transformer under natural simplifying assumptions when finetuning on a data mixture consisting of context-critical and non-context-critical points.\n\n\nFinally, our analysis naturally leads us to some mitigation strategies by data curation, data augmentation, and regularization. These strategies are able to partially alleviate the drop in deep networks on real-world datasets, showing that our theoretical insights do translate to practical settings. However, as we discuss in §6, these mitigation strategies each have fundamental limitations and tradeoffs.\n\n\nOverall, in this work, we uncover a broad failure in IFT, where the model begins to rely more on the parametric knowledge than the input context when they conflict, despite an initial jump in context reliance. To the best of our knowledge, we are the first to point out this deficiency with instruction tuned models. We also provide empirical and theoretical understanding of this observation along with some initial mitigation strategies that we hope would serve as a useful starting point to address the fundamental challenge of improving context reliance in language models.\n\n",
    "related_work": "Knowledge Conflicts in LLMs:Language models are often exposed to user input instructions and accompanying context, which at times gives information or requests a behavior at odds with model’s prior from pretraining.\nWhile various studies under the umbrella of “knowledge conflicts” have tried to understand model’s behavior under these circumstances, i.e. whether to prefer context or parametric knowledge,\nthere has been limited analysis on how instruction finetuning (IFT) itself affects this, despite IFT being a staple part of current LLM training pipeline.\nExisting works focus mainly on improving context reliance using inference time or augmentation like approaches.For example,\nCAD(Shi et al.,2023), COIECD(Yuan et al.,2024)and AutoCAD(Wang et al.,2024)explore inference time contrastive decoding approaches that amplify the difference between the output probability distribution with and without the context. These methods provide limited gains, especially in instruction finetuned models(Wang et al.,2024).Zhou et al. (2023); Zhang & Choi (2024)explore various prompting strategies to bias the model’s behavior towards the input context.Jin et al. (2024b)tries to build a mechanistic interpretation.\nOn the other hand,Longpre et al. (2022); Fang et al. (2024); Neeman et al. (2022); Li et al. (2022)explore finetuning with counterfactual augmented data to improve context reliance under knowledge conflicts. However, in §6, we show that\ncounterfactual data augmentation cannot fix all types of context-parametric conflicts (e.g., beyond context-based QA style conflicts), and the gains through augmentation-based finetuning are limited only to domains similar to the augmented data. Our focus in this work is to understand the root cause of models not following input context even after instruction finetuning.\n\n\nRAG and Knowledge Conflicts:Understanding the effect of instruction finetuning on knowledge conflicts is of high relevance for retrieval augmented generation (RAG), an important practical usecase of LLMs. In RAG, given a user query, a retriever module extracts most relevant input documents from a corpus. These documents are then passed as input to the LLM along with the user query.\nRAG has many scenarios of conflicts, both between the various external documents or between external documents and parametric knowledge.Guu et al. (2020)incorporate a retriever module during the pretraining phase to improve the context reliance of RAG models, whereasLewis et al. (2021)incorporate a retriever during finetuning.\nIn the case of conflicts between external documents,Jin et al. (2024a); Kortukov et al. (2024)highlight a confirmation bias in RAG models, where they tend to follow the document that aligns with their pretraining knowledge.\nSome works in fact even suggest that context reliance may not always be desirable, especially when the input context is noisy and irrelevant. For example,Zhang et al. (2024)propose a training procedure to instead increase the model’s tendency to answer using parametric knowledge when the input context might be noisy.\n\n\nInstruction Tuning:Instruction tuning is done to improve models ability to comprehend user input and instructions(Ding et al.,2023b). Lately, IFT has also been used to instill additional capabilities or skills into pretrained language models by finetuning on datasets curated accordingly(Wang et al.,2023).Biderman et al. (2024); Wang et al. (2022); Kotha et al. (2024); Luo et al. (2023)highlight forgetting or worsening of performance on orthogonal (out of distribution) tasks, when finetuning LLM for specific skills, similar to the classic phenomenon of forgetting when finetuning on new distributions(Kemker et al.,2017; Goodfellow et al.,2015).\nIn contrast, in this work we show an unexpected drop in context reliance with instruction tuning, afteran expected initial increase. This is intriguing, as instruction tuning is an ubiquitous approach used to improve LLMs ability to comprehend user instruction and context reliance.\n\n",
    "origin_citations": [
      "Trusting your evidence: Hallucinate less with context-aware decoding, 2023.",
      "Discerning and resolving knowledge conflicts through adaptive decoding with contextual information-entropy constraint, 2024.",
      "Adacad: Adaptively decoding to balance conflicts between contextual and parametric knowledge, 2024.",
      "Context-faithful prompting for large language models, 2023.",
      "Mitigating temporal misalignment by discarding outdated facts, 2024.",
      "Cutting off the head ends the conflict: A mechanism for interpreting and mitigating knowledge conflicts in language models.",
      "Entity-based knowledge conflicts in question answering, 2022.",
      "Getting sick after seeing a doctor? diagnosing and mitigating knowledge conflicts in event temporal reasoning, 2024.",
      "Disentqa: Disentangling parametric and contextual knowledge with counterfactual question answering, 2022.",
      "Large language models with controllable working memory, 2022.",
      "Realm: Retrieval-augmented language model pre-training, 2020.",
      "Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021.",
      "Tug-of-war between knowledge: Exploring and resolving knowledge conflicts in retrieval-augmented language models.",
      "Studying large language model behaviors under realistic knowledge conflicts, 2024.",
      "Raft: Adapting language model to domain specific rag, 2024.",
      "Enhancing chat language models by scaling high-quality instructional conversations, 2023b.",
      "How far can camels go? exploring the state of instruction tuning on open resources, 2023.",
      "Lora learns less and forgets less, 2024.",
      "Two-stage llm fine-tuning with less specialization and more generalization.",
      "Understanding catastrophic forgetting in language models via implicit inference, 2024.",
      "An empirical study of catastrophic forgetting in large language models during continual fine-tuning.",
      "Measuring catastrophic forgetting in neural networks.",
      "An empirical investigation of catastrophic forgetting in gradient-based neural networks, 2015."
    ],
    "citations": [
      2449,
      466413,
      812391,
      495085,
      259080,
      463067,
      7092,
      258983,
      32917,
      33134,
      7790,
      7225,
      465045,
      133939,
      8356,
      2213,
      890118,
      35070,
      480909,
      406995,
      2979,
      143905
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 90,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 85\nSynthesis_Analysis: 90\nFinal_Score: 85"
    },
    "overall_score": 84.64,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "SPS6HzVzyt",
        "forum": "SPS6HzVzyt",
        "number": 12499,
        "title": "Context-Parametric Inversion: Why Instruction Finetuning May Not Actually Improve Context Reliance",
        "normalized_title": "context parametric inversion why instruction finetuning may not actually improve context reliance",
        "abstract": "Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\\textit{despite an initial expected increase}$. We call this phenomenon as the $\\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions.  We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia.  We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning. \nWe tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains.",
        "authors": [
          "Sachin Goyal",
          "Christina Baek",
          "J Zico Kolter",
          "Aditi Raghunathan"
        ],
        "decision": null,
        "pdate": 1737562892211,
        "cdate": 1727479627580
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "SPS6HzVzyt",
          "forum": "SPS6HzVzyt",
          "number": 12499,
          "title": "Context-Parametric Inversion: Why Instruction Finetuning May Not Actually Improve Context Reliance",
          "normalized_title": "context parametric inversion why instruction finetuning may not actually improve context reliance",
          "abstract": "Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\\textit{despite an initial expected increase}$. We call this phenomenon as the $\\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions.  We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia.  We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning. \nWe tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains.",
          "authors": [
            "Sachin Goyal",
            "Christina Baek",
            "J Zico Kolter",
            "Aditi Raghunathan"
          ],
          "decision": null,
          "pdate": 1737562892211,
          "cdate": 1727479627580
        }
      ]
    }
  },
  {
    "paper_id": 6419,
    "title": "Image Background Serves as Good Proxy for Out-of-distribution Data",
    "abstract": "Out-of-distribution (OOD) detection empowers the model trained on the closed image set to identify unknown data in the open world. Though many prior techniques have yielded considerable improvements in this research direction, two crucial obstacles still remain. Firstly, a unified perspective has yet to be presented to view the developed arts with individual designs, which is vital for providing insights into future work. Secondly, we expect sufficient natural OOD supervision to promote the generation of compact boundaries between the in-distribution (ID) and OOD data without collecting explicit OOD samples. To tackle these issues, we propose a general probabilistic framework to interpret many existing methods and an OOD-data-free model, namelySelf-supervisedSampling forOODDetection (SSOD). SSOD efficiently exploits natural OOD signals from the ID data based on the local property of convolution. With these supervisions, it jointly optimizes the OOD detection and conventional ID classification in an end-to-end manner. Extensive experiments reveal that SSOD establishes competitive state-of-the-art performance on many large-scale benchmarks, outperforming the best previous method by a large margin,e.g., reporting-6.28%FPR95 and+0.77%AUROC on ImageNet,-19.01%FPR95 and+3.04%AUROC on CIFAR-10, and top-ranked performance on hard OOD datasets,i.e., ImageNet-O and OpenImage-O.",
    "introduction": "Figure 1:Feature visualization of the ID and OOD images. The green/orange dots are surrogate ID/OOD features generated by SSOD. The blue/gray dots are natural ID/OOD features of ImageNet/iNaturalist.\n\n\nIdentifying the out-of-distribution (OOD) samples is vital for practical applications since the deployed systems can not handle these unknown data in a human-like fashion,e.g., performing rejection. Based on research in animal behavior, such asTierney & Jane (1986)andShettleworth (2001), these abilities demonstrated by humans are rooted in experience, which means we earn them from interactions with environments. For instance, the adults will not reach their fingers towards the flames in case of burning, while the children may take this risky action due to their limited cognitive ability. This principle reveals that the failure encountered by the well-trained classification models in detecting OOD examples results from the lack of interactions with OOD data. Consequently, acquiring sufficient and diversified OOD signals from the in-distribution (ID) data has become a promising direction, which helps mitigate the gap between classification and OOD detection. Fig.1provides an intuitive illustration of natural supervision’s importance in OOD detection. In particular, we employ the pre-trained ResNet-50(He et al.,2016)as the backbone and ImageNet(Russakovsky et al.,2015)as the ID data. The OOD samples are from iNaturalist(Horn et al.,2018). The blue points are pooled features from ImageNet,i.e., the natural ID features. The green/orange points indicate the ID/OOD signals generated by our SSOD,i.e., the ID/OOD proxy. We can see that the OOD signals sampled by SSOD build a defensive wall between the real ID and OOD data, which can serve as the proxy of authentic OOD images to supervise the training procedure.\n\n\nRecall other existing OOD detection methods that primarily rely on the perspective of statistical difference,i.e., observing distinctions of the pre-trained features between ID and OOD samples. These methods use heuristic rules to filter out OOD data in a two-stage manner,i.e., pre-training and post-processing, which suffer from the following drawbacks. Firstly, the frozen model weights are obtained on the ID classification task with limited OOD supervision. Therefore, the extracted features inherently carry bias, which is not distinguishable enough for identifying OOD data (cf. Fig.4). Secondly, the two-stage design yields poor scalability and efficiency since it is unsuitable for scenarios without pre-trained models,e.g., given insufficient training data, a two-stage model may fail to obtain high-quality weights, thus encountering the performance drop in detecting OOD data.\n\n\nTo tackle the issues above, this paper interprets the OOD detection task with a unified probabilistic framework, which can widely include many previous individual designs. Concretely, our framework starts from Bayes’ rule and divides the robust classification problem into two tasks: conventional ID classification and OOD detection. According to our theoretical analysis, the deficiency encountered by traditional neural networks in identifying OOD data arises from the absence of a critical component,i.e., an OOD factor that estimates the likelihood of images belonging to the in-distribution. Furthermore, sailing from this general foundation, we presentSelf-supervisedSampling forOODDetection (SSOD), an end-to-end trainable frameworkw/oresorting to explicit OOD annotations. In contrast to the observed paths that hug synthetic OOD features, SSOD directly samples natural OOD supervision from the background of ID images,i.e., self-supervised, getting rid of the constraints resulting from the lack of labeled OOD data and the deviation introduced within the OOD feature syntheses stage. Extensive experiments demonstrate that the joint end-to-end training significantly improves the OOD detection performance and guides the model to focus more on the object-discriminative characters instead of the meaningless background information (cf. Fig.2). The major contributions of this paper are summarized as follows:\n\n\n•We establish a general probabilistic framework to interpret the OOD detection, where various OOD methods can be analyzed, with main differences and key limitations clearly identified.•To promote ID/OOD features being more distinguishable, we design an end-to-end trainable model, namelySelf-supervisedSampling forOODDetection (SSOD), to sample natural OOD signals from the ID images. SSOD avoids the labor-intensive work of labeling sufficient OOD images.•SSOD is evaluated across various benchmarks and model architectures for OOD detection, where it outperforms current state-of-the-art approaches by a large margin,e.g., improving KNN(Sun et al.,2022)w/ and w/o contrastive learning with-20.23%and-38.17%FPR95 on Places(Zhou et al.,2018), and Energy(Liu et al.,2020)with-30.74%FPR95 and+8.44%AUROC on SUN(Xiao et al.,2010), to name a few. The scalability and superiority of SSOD promise its potential to be a starting point for solving the OOD detection problem.\n\n\n(a)Locality of ResNet-50.(b)Locality of SSOD.(c)Samplers of ResNet-50.(a)Locality of ResNet-50.(b)Locality of SSOD.(c)Samplers of ResNet-50.(d)Samplers of SSOD.Figure 2:(a): The image patches in blue are recognized aspenguinswith over 40% confidence by the vanilla ResNet-50.(b): The SSOD counterpart of (a). We can see that SSOD focuses on more discriminative characters, reducing the false negatives.(c): The image patches in green/red are sampled as ID/OOD supervision with over 95% confidence by ResNet-50.(d): The SSOD counterpart of (c), we notice that SSOD identifies more background as OOD data. SSOD is motivated by the local property of conventional networks as illustrated in (a), and the results depicted in (b) reveal that the joint training enhances the local property. The confidence is manually selected.\n\n",
    "related_work": "Score-based posterior calibration.This line of research aims to find differences between the ID and OOD data, thus designing model-specific discriminative functions to identify the OOD samples. The related work includes ODIN(Liang et al.,2018), LogitNorm(Wei et al.,2022), GradNorm(Huang et al.,2021), ReAct(Sun et al.,2021), Energy(Liu et al.,2020), and CIDER(Ming et al.,2023), to name a few. Generally, these methods are usually pre- or post-processing schemes that demonstrate no need for retraining the neural networks. Although these methods above report considerable performance improvements and sometimes are training efficient, they do not necessarily lead to significant generalization ability. For example, ReAct(Sun et al.,2021)investigates the distinct behaviors of ID and OOD data afterReLUReLU\\mathrm{ReLU}roman_ReLUfunction, and therefore, it fails to perform on architectures adopting other activations, such asGELUGELU\\mathrm{GELU}roman_GELU,SigmoidSigmoid\\mathrm{Sigmoid}roman_Sigmoid, andTanhTanh\\mathrm{Tanh}roman_Tanh,etc. Similarly, the ODIN(Liang et al.,2018)investigates post-processing schemes specially designed forSoftmaxSoftmax\\mathrm{Softmax}roman_Softmax,i.e., the temperature scaling. These specific designs promote OOD detection but limit the model’s scalability. In contrast, our SSOD doesn’t suffer from this limitation as it addresses the OOD detection directed by Bayes’ theorem, which widely holds in general scenarios.\n\n\nAuxiliary supervision from synthetic OOD data.The lack of OOD supervision is a critical factor leading to unsatisfactory performance in OOD detection. Thus, significant interest has been raised in generating synthetic OOD data. Existing approaches tackling this issue can be roughly divided into two manners, which are feature and image generation. The former samples OOD features from the ID boundary, such as VOS(Du et al.,2022), or generates them using GAN, such as BAL(Pei et al.,2022). In contrast, the image generation yields more training tax since it directly generates the OOD images, such as Conf(Lee et al.,2018a), SBO(Möller et al.,2021), MG-GAN(Dendorfer et al.,2021), NAS-OOD(Bai et al.,2021), CODEs(Tang et al.,2021), and VITA(Chen et al.,2022b). In summary, these methods either introduce bias as they only consider the approximated feature space or are costly due to the image generation. Unlike the above methods, SSOD avoids feature bias and training tax by utilizing the universal local property of neural networks, extracting realistic OOD supervision from the ID images without generation cost.\n\n",
    "origin_citations": [
      "Enhancing the reliability of out-of-distribution image detection in neural networks.",
      "Mitigating neural network overconfidence with logit normalization.",
      "On the importance of gradients for detecting distributional shifts in the wild.",
      "React: Out-of-distribution detection with rectified activations.",
      "Energy-based out-of-distribution detection.",
      "How to exploit hyperspherical embeddings for out-of-distribution detection?",
      "VOS: learning what you don’t know by virtual outlier synthesis.",
      "Out-of-distribution detection with boundary aware learning.",
      "Training confidence-calibrated classifiers for detecting out-of distribution samples.",
      "Out-of-distribution detection and generation using soft brownian offset sampling and autoencoders.",
      "Mg-gan: A multi-generator model preventing out-of-distribution samples in pedestrian trajectory prediction.",
      "Nas-ood: Neural architecture search for out-of-distribution generalization.",
      "Codes: Chamfer out-of-distribution examples against overconfidence issue.",
      "VITA: A multi-source vicinal transfer augmentation method for out-of-distribution generalization."
    ],
    "citations": [
      3251,
      3296,
      3270,
      3277,
      7707,
      517721,
      6981,
      536706,
      3234,
      251250,
      453980,
      172225,
      94179,
      368804
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 84,
      "Final_Score": 83.8,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 84\nFinal_Score: 84"
    },
    "overall_score": 84.63999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "ym0ubZrsmm",
        "forum": "ym0ubZrsmm",
        "number": 7098,
        "title": "Image Background Serves as Good Proxy for Out-of-distribution Data",
        "normalized_title": "image background serves as good proxy for out of distribution data",
        "abstract": "Out-of-distribution (OOD) detection empowers the model trained on the closed image set to identify unknown data in the open world. Though many prior techniques have yielded considerable improvements in this research direction, two crucial obstacles still remain. Firstly, a unified perspective has yet to be presented to view the developed arts with individual designs, which is vital for providing insights into future work. Secondly, we expect sufficient natural OOD supervision to promote the generation of compact boundaries between the in-distribution (ID) and OOD data without collecting explicit OOD samples. To tackle these issues, we propose a general probabilistic framework to interpret many existing methods and an OOD-data-free model, namely $\\textbf{S}$elf-supervised $\\textbf{S}$ampling for $\\textbf{O}$OD $\\textbf{D}$etection (SSOD). SSOD efficiently exploits natural OOD signals from the ID data based on the local property of convolution. With these supervisions, it jointly optimizes the OOD detection and conventional ID classification in an end-to-end manner. Extensive experiments reveal that SSOD establishes competitive state-of-the-art performance on many large-scale benchmarks, outperforming the best previous method by a large margin, e.g., reporting $\\textbf{-6.28}$% FPR95 and $\\textbf{+0.77}$% AUROC on ImageNet, $\\textbf{-19.01}$% FPR95 and $\\textbf{+3.04}$% AUROC on CIFAR-10, and top-ranked performance on hard OOD datasets, i.e., ImageNet-O and OpenImage-O.",
        "authors": [
          "Sen Pei"
        ],
        "decision": null,
        "pdate": 1705410997979,
        "cdate": 1695455317621
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "ym0ubZrsmm",
          "forum": "ym0ubZrsmm",
          "number": 7098,
          "title": "Image Background Serves as Good Proxy for Out-of-distribution Data",
          "normalized_title": "image background serves as good proxy for out of distribution data",
          "abstract": "Out-of-distribution (OOD) detection empowers the model trained on the closed image set to identify unknown data in the open world. Though many prior techniques have yielded considerable improvements in this research direction, two crucial obstacles still remain. Firstly, a unified perspective has yet to be presented to view the developed arts with individual designs, which is vital for providing insights into future work. Secondly, we expect sufficient natural OOD supervision to promote the generation of compact boundaries between the in-distribution (ID) and OOD data without collecting explicit OOD samples. To tackle these issues, we propose a general probabilistic framework to interpret many existing methods and an OOD-data-free model, namely $\\textbf{S}$elf-supervised $\\textbf{S}$ampling for $\\textbf{O}$OD $\\textbf{D}$etection (SSOD). SSOD efficiently exploits natural OOD signals from the ID data based on the local property of convolution. With these supervisions, it jointly optimizes the OOD detection and conventional ID classification in an end-to-end manner. Extensive experiments reveal that SSOD establishes competitive state-of-the-art performance on many large-scale benchmarks, outperforming the best previous method by a large margin, e.g., reporting $\\textbf{-6.28}$% FPR95 and $\\textbf{+0.77}$% AUROC on ImageNet, $\\textbf{-19.01}$% FPR95 and $\\textbf{+3.04}$% AUROC on CIFAR-10, and top-ranked performance on hard OOD datasets, i.e., ImageNet-O and OpenImage-O.",
          "authors": [
            "Sen Pei"
          ],
          "decision": null,
          "pdate": 1705410997979,
          "cdate": 1695455317621
        }
      ]
    }
  },
  {
    "paper_id": 27201,
    "title": "Visual-Text Cross Alignment:Refining the Similarity Score in Vision-Language Models",
    "abstract": "It has recently been discovered that using a pre-trainedvision-language model(VLM), e.g., CLIP, to align a whole query image with several finer text descriptions generated by a large language model can significantly enhance zero-shot performance.\nHowever, in this paper, we empirically find that the finer descriptions tend to align more effectively withlocal areas of the query imagerather than the whole image, and then we theoretically validate this finding.\nThus, we present a method calledweighted visual-text cross alignment(WCA).\nThis method begins with alocalized visual promptingtechnique, designed to identify local visual areas within the query image.\nThe local visual areas are thencross-alignedwith the finer descriptions by creating a similarity matrix using the pre-trained VLM.\nTo determine how well a query image aligns with each category, we develop a score function based on the weighted similarities in this matrix.\nExtensive experiments demonstrate that our method significantly improves zero-shot performance across various datasets, achieving results that are even comparable to few-shot learning methods.\nThe code is available atgithub.com/tmlr-group/WCA.",
    "introduction": "Following the significant advancements of large-scale pre-training in natural language processing(Devlin et al.,2018; Radford et al.,2018,2019; Brown et al.,2020), the CLIP model(Radford et al.,2021)scales up its pre-training data throughaligningimages and the corresponding natural language captions in the shared latent space, which achieves remarkable performance in zero-shot classification. Despite its achievements, CLIP’s performance exhibits notable sensitivity to the prompts used during the inference stage(Radford et al.,2021; Zhou et al.,2022b). For example,Zhou et al. (2022b)has highlighted that changing the prompt from “a photo of [CLASS]” to “a photo of a [CLASS]” can lead to a performance boost of 6%. Crafting effective prompts is crucial but it requires significant time, effort, and domain-specific knowledge(Zhou et al.,2022b), making it challenging to deploy such models in practical applications.\n\n\nFigure 1:Aligning an entire image with a detailed text description results in lower scaled cosine similarity, as shown on the left. Aligning the description with a specific image part, such as the detailed red patch (on the right), increases the score.\n\n\nTo address the above issue, a promising solution is to uselarge language models(LLMs) to generate several finer text descriptions of each category(Menon & Vondrick,2022; Pratt et al.,2023). This strategy helps reduce the manual effort in creating prompts and, more importantly, does not necessitate additional turning, thereby more easily preserving models’ generalization abilities. The ability to generalize is a crucial issue in prompt-learning methods(Li et al.,2022b; Wang et al.,2022; Wu et al.,2023; Tanwisuth et al.,2023)as these methods tend to overfit training data. LLM-based visual-text alignment emphasizesglobal matching, namely, text descriptions are aligned with the whole image.\n\n\nFigure 2:We show different zero-shot visual-text alignment methods: (a) CLIP, (b) Visual-Text Alignment via LLMs(Menon & Vondrick,2022; Pratt et al.,2023), and (c) Weighted Visual-Text Cross Alignment (ours). Unlike (a) and (b), (c) utilizes a localized visual prompting technique to enhance alignment by ensuring that detailed descriptions match precisely with specific areas of the visual content.\n\n\nHowever, in this paper, we find:\n\n\nFiner-grained text descriptions may align more accurately with the specific area of an image but not necessarily with the image as a whole.\n\n\nThis is because finer descriptions often contain detailed and finer-grained visual concepts, such as “a woodpecker has a straight and pointed bill”, may not align precisely with an entire image, as demonstrated in Figure1.\nIn expanding this, it becomes clear that the complexity of images often contains a myriad of details that a description might capture only in part. These finer elements, while crucial, might lead to a misalignment when the objective is to correlate the entire image with its description. Additionally, we provide a theoretical analysis in Section4to gain a deeper understanding of the issue. This discrepancy suggests a need for more nuanced approaches in visual-text alignment, where the focus is not just on the whole query image (i.e., the global alignment) but also on recognizing and aligning different areas within the query image.\n\n\nTo this end, we propose a method calledweighted visual-textcrossalignment(WCA). This method concentrates on local and regional visual areas within the query image that can align better with the fine-grained text descriptions of each category. This can be achieved throughlocalized visual prompting, where the image is prompted to focus on localized visual elements, such as via cropping. These visual elements are thencross-alignedwith their corresponding detailed text descriptions, leading to a similarity matrix. Furthermore, we introduce a score function based on theweighted similaritiesin the matrix and use the score to see how well a query image is aligned with each category.\n\n\nA key feature of WCA is its consideration of the varying importance of localized visual elements and text descriptions within the similarity matrix, ensuring that each similarity value in this matrix contributes differently to the overall similarity aggregation. The importance of a specific localized visual element is quantified by its cosine similarity with the query image, where a high score suggests that the element captures the main semantic content of the query image. Similarly, the relevance of a text description can be estimated by its similarity to the category label it corresponds to.\nTherefore, our method achieves accurate visual-text alignment scores efficiently without additional models or training, making it a highly efficient approach.\nWe demonstrate the distinction between WCA and the current methods of visual-text alignment in Figure2.\nOur empirical results show that WCA significantly enhances zero-shot performance across various datasets, even comparable with few-shot methods.\n\n\nTo summarize, our main contributions are outlined as follows:(i)We have identified and conducted a theoretical analysis of the issue where aligning an entire image with finer text descriptions results in suboptimal performance.(ii)We introduce a method, WCA, that performsweighted cross alignmentbetween the finer descriptions and local visual areas using localized visual prompting.(iii)Our extensive experiments validate our theoretical hypothesis and demonstrate the efficacy of our method, significantly surpassing the state-of-the-art methods without the need for extra models or data.(iv)We offer insights into the key factors that contribute to the effectiveness of our method.\n\n",
    "related_work": "Vision-language models.Vision-language models (VLMs) pre-trained on large-scale data have shown efficacy in enhancing representation learning capabilities(Cho et al.,2021; Kim et al.,2021; Xue et al.,2021; Li et al.,2021; Wang et al.,2021; Li et al.,2023). CLIP(Radford et al.,2021)underwent training on a corpus of 400 million paired images and texts, exhibiting robust transferable ability and exceptional zero-shot performance. In a similar vein, the introduction of ALIGN(Jia et al.,2021)demonstrates that despite being pre-trained on datasets containing image-text pairs with considerable noise, the scale of the training corpus can compensate for this noise and is capable of learning superior representations. Subsequent works, including FLAVA(Singh et al.,2022), Florence(Yuan et al.,2021), BLIP(Li et al.,2022a), and so on, have continued to advance this paradigm, contributing further to the field.\n\n\nTextual prompting in vision-language models.Despite CLIP exhibiting superior zero-shot capabilities, the effectiveness of its application in downstream tasks is significantly influenced by the choice of prompts, as noted byRadford et al. (2021)andZhou et al. (2022b).Zhou et al. (2022b)highlight that selecting the optimal prompt is complex and time-intensive, often requiring prompt tuning. To address this,Menon & Vondrick (2022)andPratt et al. (2023)leverage the knowledge embedded in LLMs, such as GPT-3(Brown et al.,2020)for the automatic generation of class-specific descriptions. These descriptions, particularly focusing on the discriminating features of image categories, are then aligned with the query image. This method has been shown to be effective as it enriches the textual representation by incorporating LLMs.Roth et al. (2023)examine this phenomenon and introduce the WaffleCLIP framework, which replaces LLM-generated descriptions with random character and word descriptions, eliminating the need to query LLMs and offering a cost-effective alternative. However, our work diverges from these approaches as they do not engage in visual prompting techniques.\n\n\nVisual prompting in vision-language models.In contrast to text-based prompting in VLMs, visual prompting aims to process the visual input accordingly.Yao et al. (2024)color image regions and utilize a captioning model to identify objects based on color predictions.Bahng et al. (2022)experiment with learning the image perturbation, keeping the model parameters unchanged. These approaches, along with the studies(Jia et al.,2022; Tu et al.,2023)require at least a few samples from downstream tasks. The RedCircle introduced byShtedritski et al. (2023)suggests that highlighting an object with a red circle can direct the model’s attention to that region, but it requires manual annotation.Yang et al. (2024)propose FGVP, which uses an extra model SAM(Kirillov et al.,2023), to identify objects first and then employBlur Reverse Masksto enhance the semantic localization capability of areas around the objects, reducing the need for manual annotation but adding complexity. Our method differs from these as we do not need downstream data, manual annotation, or additional models.\n\n\nTest time prompt tuning in vision-language models.While fine-tuning prompts can adapt pre-trained VLMs to specific downstream tasks, this approach requires labeled training data, which can be costly and unavailable for zero-shot tasks. Test-time prompt tuning (TPT), as introduced byShu et al. (2022), addresses this issue by learning adaptive prompts for individual test samples through the generation of multiple randomly augmented views. The goal is to optimize text prompts in an unsupervised manner. However, naive augmentation methods may lead to overly simplistic variations in test data. To address this,Feng et al. (2023)proposed DiffTPT, which uses diffusion models to augment test samples with richer visual appearance variations.\nIn contrast, our method, while also applied during testing, does not involve the same tuning processes as TPT and DiffTPT. Instead, it leverages the strengths of pre-trained VLMs in a different manner, potentially offering a more efficient approach. Our method avoids the need for extensive data augmentation and fine-tuning procedures, which are typically required by TPT and DiffTPT to enhance their performance. By directly utilizing the inherent capabilities of pre-trained VLMs, our approach simplifies the alignment process.\n\n\nFigure 3:Overview of weighted visual-text cross alignment (WCA). The process begins withlocalized visual prompting, where the input image𝒙𝒙\\bm{x}bold_italic_xis divided into localized patches, such as{𝒙1,𝒙2,𝒙3}subscript𝒙1subscript𝒙2subscript𝒙3\\{\\bm{x}_{1},\\bm{x}_{2},\\bm{x}_{3}\\}{ bold_italic_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , bold_italic_x start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , bold_italic_x start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT }. These patches are encoded by an image encoder to produce visual features. Thetext promptingstage utilizes a large language model to generate detailed textual descriptions{𝒚1,𝒚2,𝒚3}subscript𝒚1subscript𝒚2subscript𝒚3\\{\\bm{y}_{1},\\bm{y}_{2},\\bm{y}_{3}\\}{ bold_italic_y start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , bold_italic_y start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , bold_italic_y start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT }for a given class label𝒚𝒚\\bm{y}bold_italic_y(e.g., “woodpecker”). The WCA calculates alignment scores between visual features and textual features, using patch weights{w1,w2,w3}subscript𝑤1subscript𝑤2subscript𝑤3\\{w_{1},w_{2},w_{3}\\}{ italic_w start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_w start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , italic_w start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT }and text weights{v1,v2,v3}subscript𝑣1subscript𝑣2subscript𝑣3\\{v_{1},v_{2},v_{3}\\}{ italic_v start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , italic_v start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT , italic_v start_POSTSUBSCRIPT 3 end_POSTSUBSCRIPT }. The final score is computed by summing the visual-text similarity matrix.\n\n",
    "origin_citations": [
      "Unifying vision-and-language tasks via text generation.",
      "Vilt: Vision-and-language transformer without convolution or region supervision.",
      "Probing inter-modality: Visual parsing with self-attention for vision-and-language pre-training.",
      "Align before fuse: Vision and language representation learning with momentum distillation.",
      "Simvlm: Simple visual language model pretraining with weak supervision.",
      "Progressive video summarization via multimodal self-supervised learning.",
      "Learning transferable visual models from natural language supervision.",
      "Scaling up visual and vision-language representation learning with noisy text supervision.",
      "Flava: A foundational language and vision alignment model.",
      "Florence: A new foundation model for computer vision.",
      "Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation.",
      "Learning to prompt for vision-language models.",
      "Visual classification via description from large language models.",
      "What does a platypus look like? generating customized prompts for zero-shot image classification.",
      "Language models are few-shot learners.",
      "Waffling around for performance: Visual classification with random words and broad concepts.",
      "Cpt: Colorful prompt tuning for pre-trained vision-language models.",
      "Exploring visual prompts for adapting large-scale models.",
      "Visual prompt tuning.",
      "Visual query tuning: Towards effective usage of intermediate representations for parameter and memory efficient transfer learning.",
      "What does clip know about a red circle? visual prompt engineering for vlms.",
      "Fine-grained visual prompting.",
      "Segment anything.",
      "Test-time prompt tuning for zero-shot generalization in vision-language models.",
      "Diverse data augmentation with diffusions for effective test-time prompt tuning."
    ],
    "citations": [
      119879,
      5284,
      5245,
      2422,
      255505,
      1639,
      5247,
      2468,
      338632,
      55893,
      3306,
      458652,
      229682,
      679,
      304825,
      365453,
      1038295,
      3304,
      222681,
      18052,
      306296,
      2512,
      963144,
      180051
    ],
    "extract_rate": 0.96,
    "article_scores": {
      "Content_Quality": 84,
      "Publication_Potential": 89,
      "Final_Score": 85.5,
      "raw_text": "Content_Quality: 84\nPublication_Potential: 89\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 84,
      "Synthesis_Analysis": 85,
      "Final_Score": 84.0,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 84\nSynthesis_Analysis: 85\nFinal_Score: 84"
    },
    "overall_score": 84.6,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "JymXv7mkrQ",
        "forum": "JymXv7mkrQ",
        "number": 3077,
        "title": "Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models",
        "normalized_title": "visual text cross alignment refining the similarity score in vision language models",
        "abstract": "It has recently been discovered that using a pre-trained *vision-language model* (VLM), e.g., CLIP, to align a whole query image with several finer text descriptions generated by a large language model can significantly enhance zero-shot performance. However, in this paper, we empirically find that the finer descriptions tend to align more effectively with *local areas of the query image* rather than the whole image, and then we theoretically validate this finding. Thus, we present a method called *weighted visual-text cross alignment* (WCA). This method begins with a *localized visual prompting* technique, designed to identify local visual areas within the query image. The local visual areas are then *cross-aligned* with the finer descriptions by creating a similarity matrix using the pre-trained VLM. To determine how well a query image aligns with each category, we develop a score function based on the weighted similarities in this matrix. Extensive experiments demonstrate that our method significantly improves zero-shot performance across various datasets, achieving results that are even comparable to few-shot learning methods.",
        "authors": [
          "Jinhao Li",
          "Haopeng Li",
          "Sarah Monazam Erfani",
          "Lei Feng",
          "James Bailey",
          "Feng Liu"
        ],
        "decision": null,
        "pdate": 1714610327721,
        "cdate": 1706603714303
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "JymXv7mkrQ",
          "forum": "JymXv7mkrQ",
          "number": 3077,
          "title": "Visual-Text Cross Alignment: Refining the Similarity Score in Vision-Language Models",
          "normalized_title": "visual text cross alignment refining the similarity score in vision language models",
          "abstract": "It has recently been discovered that using a pre-trained *vision-language model* (VLM), e.g., CLIP, to align a whole query image with several finer text descriptions generated by a large language model can significantly enhance zero-shot performance. However, in this paper, we empirically find that the finer descriptions tend to align more effectively with *local areas of the query image* rather than the whole image, and then we theoretically validate this finding. Thus, we present a method called *weighted visual-text cross alignment* (WCA). This method begins with a *localized visual prompting* technique, designed to identify local visual areas within the query image. The local visual areas are then *cross-aligned* with the finer descriptions by creating a similarity matrix using the pre-trained VLM. To determine how well a query image aligns with each category, we develop a score function based on the weighted similarities in this matrix. Extensive experiments demonstrate that our method significantly improves zero-shot performance across various datasets, achieving results that are even comparable to few-shot learning methods.",
          "authors": [
            "Jinhao Li",
            "Haopeng Li",
            "Sarah Monazam Erfani",
            "Lei Feng",
            "James Bailey",
            "Feng Liu"
          ],
          "decision": null,
          "pdate": 1714610327721,
          "cdate": 1706603714303
        }
      ]
    }
  },
  {
    "paper_id": 47433,
    "title": "DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing",
    "abstract": "We introduceDRESS, a novel approach for generating stylized large language model (LLM) responses through representation editing. Existing methods like prompting and fine-tuning are either insufficient for complex style adaptation or computationally expensive, particularly in tasks like NPC creation or character role-playing. Our approach leverages the over-parameterized nature of LLMs to disentangle a style-relevant subspace within the model’s representation space to conduct representation editing, ensuring a minimal impact on the original semantics.\nBy applying adaptive editing strengths, we dynamically adjust the steering vectors in the style subspace to maintain both stylistic fidelity and semantic integrity. We develop two stylized QA benchmark datasets to validate the effectiveness ofDRESS, and the results demonstrate significant improvements compared to baseline methods such as prompting and ITI. In short,DRESSis a lightweight, train-free solution for enhancing LLMs with flexible and effective style control, making it particularly useful for developing stylized conversational agents.111Codes and benchmark datasets are available athttps://github.com/ArthurLeoM/DRESS-LLM.",
    "introduction": "Figure 1:An illustrative example of representation editing forShakespeare-style responses.\n\n\nLarge language models (LLMs) like GPT-4(Achiam et al.,2023)and LLaMA-3(Dubey et al.,2024)have demonstrated exceptional performance across a range of natural language processing (NLP) tasks including question-answering. This evokes the wide use of LLMs as conversational agents(Weizenbaum,1966)for various applications, including psychological counseling(Li et al.,2023a), creating gaming NPCs (non-player characters)(Cox & Ooi,2023)and character simulacra(Shao et al.,2023).\nWhile LLMs are adept at providing accurate and coherent answers, they lack the intrinsic ability to tailor responses in a specificlanguage style. Language style(Jin et al.,2022)is linguistically defined as the manner of expressing the semantics, depicted by multiple attributes like personality, emotion, authorship, era background, etc.\nStylized responses are crucial for LLM agents as the style can shape the interaction tone, making the agents more immersive and engaging, and ensuring that responses are empathetic and appropriately tailored to the user’s emotional states. Hence, crafting the language style is essential for shaping the specific image and personality of conversational agents.\nTherefore, we aim to solve the following question:How to make LLMs respond to user questions in a specific style?\n\n\nCurrently, there are two main approaches to achieving stylized responses - prompting with few-shot demonstrations and fine-tuning.\nPrompting methods(Park et al.,2023)leverage the in-context learning ability(Brown et al.,2020)of LLMs by using a description of the target style along with few-shot examples to generate stylized responses.\nHowever, simply prompting LLMs is no longer proper as instructions are plain and insufficient to describe a certain style comprehensively, and demonstrations could severely increase the sequence length, increasing the risk of lost-in-the-middle(Liu et al.,2024).\nA better way is to conduct supervised fine-tuning (SFT)(Ma et al.,2024)with target style response data(Shao et al.,2023), where LLM’s outputs are adapted to the target style distribution by adjusting the model parameters. Yet this approach is overly burdensome, particularly for scenarios like game NPC construction. Each character requires a separate fine-tuning process, making the creation of multiple characters extremely costly in terms of time and computational resources.\nTherefore, it is necessary to develop aneffectiveandefficientstrategy to reach our goal.\n\n\nRepresentation editing(Burns et al.,2023; Turner et al.,2023)has recently been widely used to control specific behaviors of LLMs (e.g., truthfulness enhancement(Zou et al.,2023), knowledge editing(Hernandez et al.,2023), etc.).\nSince it operates solely on the representation space without optimizing the parameter space, it islightweight, train-free, and highly efficient. Additionally, it leverages large amounts of data to compute generalizable steering vectors for depicting specific model functions, making it highlyeffective.\nBuilding on this insight, our approach attempts to utilize representation editing methods to craft the style of LLM output.\nSpecifically, as shown in Figure1, we aim to solve a steering vector that is added to LLM’s activations during inference, shifting the representations to the direction of another language style (e.g., poetic and rhythmic Shakespearean early modern English).\nThis approach fulfills our need to combine the efficiency of a train-free method with the effectiveness of data-driven steering for stylizing LLM responses.\n\n\nHowever, when building stylized conversational agents, it is also crucial to ensure the response quality alongside stylization. In other words, generating stylized responses must not compromise the original semantics. This presents a significant technical challenge for our representation editing approach:How to solve a steering vector minimizing the influence on the underlying semantics?Recent research observes that in the extremely wide and high-dimensional space of over-parameterized LLMs, activations can be assumed to be approximately orthogonal with high probability(Wang & Zhu,2023). This implies that the different language functions are likely to reside in orthogonal and disentangled linear subspaces(Ortiz-Jimenez et al.,2023).\nHence, our insight is to identify a style-relevant and semantic-isolated subspace from the representation space to edit within.\nBuilding on this insight, we proposeDRESS(DisentanglingRepresentationEditing inStyleSubspace), comprising the following strategies to progressively locate the style subspaces and perform semantic-isolated style steering.1) Attention head filtering: It has been demonstrated that different attention heads tend to perform varying functions(Ge et al.,2024). Hence we use probing techniques to identify the attention heads that are more closely related to styles and edit within those heads.2) Style subspace filtering: To further eliminate the style-irrelevant components in the selected attention heads, we conduct subspace filtering by seeking a subspace supported by style-related bases, so that the impact on semantics could be minimized.3) Adaptive editing strength: We employ adaptive editing strength on each subspace basis and each generated token to provide higher flexibility and avoid excessively intense editions that could harm the semantics.\nCompared to previous methods(Zou et al.,2023; Li et al.,2023b)relying on a single steering vector for editing, our approach offers greater flexibility and expressiveness by introducing a higher-rank subspace to represent style. Meanwhile, it filters out style-irrelevant noises within the steering vector, allowing for better semantic preservation.\n\n\nTo validate the effectiveness of our approach, we construct an evaluation benchmark comprising two specific stylized question-answering datasets of different languages (i.e.,Shakespeare-style in English andDream of the Red Chamber-style in Chinese222Here we select English, the most widely used language globally, and Chinese, the language with the largest number of native speakers, as our two examples. For dataset details, please refer to Section5.1.). The objective evaluation metrics include style intensity, semantic preservation, and fluency, following traditional criteria(Jin et al.,2022, Section 3). Additionally, we utilize the GPT-4 rating as a surrogate for human evaluation(Zheng et al.,2023), serving as an overall assessment metric to comprehensively evaluate the model’s capabilities.\n\n\nTo summarize, we highlight our contributions as follows. We proposed a lightweight and train-free representation editing method dubbedDRESSbased on the decoupling of language style subspaces to enable stylized LLM QA systems, which lays a fundamental groundwork for constructing humanoid conversational agents.\nTechnically, we propose three mechanisms to progressively isolate the style-relevant subspace from the entire representation space, improving the expressiveness of the style and ensuring that the semantics of LLMs remain unaffected.\nFinally, we introduced a benchmark to evaluate the response quality of stylized QA.DRESSshows significant improvements over SOTA baselines, including SFT, prompting, and other representation editing methods, demonstrating the effectiveness of our method.\n\n",
    "related_work": "Recently, there has been a line of research embarking on controlling the behavior of LLMs through representation editing, most of which focuses on truthfulness enhancement(Zou et al.,2023; Li et al.,2023b), knowledge editing(Todd et al.,2023; Hernandez et al.,2023), etc. This technique is based on the linear representation hypothesis(Elhage et al.,2022)supposing that most high-level concepts are represented linearly as directions in LLMs, which is theoretically supported by the approximate orthogonality assumption under overparameterized networks(Wang & Zhu,2023), and practically demonstrated by the success of linear probing techniques(Alain & Yoshua,2016; Belinkov,2022).\n\n\nThe primary objective of representation editing is to identify some steering vectors and add them to some layers of the forward pass of LLMs to introduce certain attributes (i.e., language style in this work or truthfulness, etc.) into the LLM outputs. Mean-Centring(Jorgensen et al.,2023)computes the steering directions using the mean difference between paired activations. RepE(Zou et al.,2023, Representation Engineering)applies PCA to the set of difference vectors and selects the principal component as the steering vector. CCS(Burns et al.,2023, Contrast Consistence Search)obtains the steering vector through the probing vector that well classifies the activation pairs. ITI(Li et al.,2023b, Inference-Time Intervention)further enhances CCS by locating attribute-relevant attention heads.\nHowever, due to the intricacy of language attributes, it is insufficient to depict them with a single direction as in the aforementioned works.\nTrFr(Chen et al.,2024b, Truth Forest)proposes a specific combination of several vectors under orthogonality regularization to enhance the expressiveness of the target attribute.\nNevertheless, none of the methods above attempt to explicitly disentangle the attribute subspace from the entire representation space to avoid affecting the original semantics. Moreover, previous works overlook the varying importance of different attribute components across various contexts, which can adversely affect the quality of the outputs.\nIn this work, we proposeDRESSto solve the problems.DRESScomprises three progressive mechanisms to isolate the attribute-relevant subspace and conduct adaptive editing in order to enhance the expressiveness and flexibility of steering, meanwhile ensuring the semantics are preserved.\n\n",
    "origin_citations": [
      "Representation engineering: A top-down approach to ai transparency.",
      "Inference-time intervention: Eliciting truthful answers from a language model.",
      "Function vectors in large language models.",
      "Inspecting and editing knowledge representations in language models.",
      "Toy models of superposition.",
      "Overparameterized random feature regression with nearly orthogonal data.",
      "Understanding intermediate layers using linear classifier probes.",
      "Probing classifiers: Promises, shortcomings, and advances.",
      "Improving activation steering in language models with mean-centring.",
      "Discovering latent knowledge in language models without supervision.",
      "Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning."
    ],
    "citations": [
      323900,
      2448,
      313510,
      20868,
      177986,
      32744,
      275358,
      472998,
      273056,
      2390,
      346164
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 80,
      "Final_Score": 83.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 80\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 85,
      "Final_Score": 85.3,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 85\nFinal_Score: 85"
    },
    "overall_score": 84.58,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "mNVR9jJYqK",
        "forum": "mNVR9jJYqK",
        "number": 9459,
        "title": "DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing",
        "normalized_title": "dressing up llm efficient stylized question answering via style subspace editing",
        "abstract": "We introduce DRESS, a novel approach for generating stylized large language model (LLM) responses through representation editing. Existing methods like prompting and fine-tuning are either insufficient for complex style adaptation or computationally expensive, particularly in tasks like NPC creation or character role-playing. Our approach leverages the over-parameterized nature of LLMs to disentangle a style-relevant subspace within the model's representation space to conduct representation editing, ensuring a minimal impact on the original semantics. By applying adaptive editing strengths, we dynamically adjust the steering vectors in the style subspace to maintain both stylistic fidelity and semantic integrity. We develop two stylized QA benchmark datasets to validate the effectiveness of DRESS, and the results demonstrate significant improvements compared to baseline methods such as prompting and ITI. In short, DRESS is a lightweight, train-free solution for enhancing LLMs with flexible and effective style control, making it particularly useful for developing stylized conversational agents. Codes and benchmark datasets are available at https://github.com/ArthurLeoM/DRESS-LLM.",
        "authors": [
          "Xinyu Ma",
          "Yifeng Xu",
          "Yang Lin",
          "Tianlong Wang",
          "Xu Chu",
          "Xin Gao",
          "Junfeng Zhao",
          "Yasha Wang"
        ],
        "decision": null,
        "pdate": 1737562655300,
        "cdate": 1727423324915
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "mNVR9jJYqK",
          "forum": "mNVR9jJYqK",
          "number": 9459,
          "title": "DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing",
          "normalized_title": "dressing up llm efficient stylized question answering via style subspace editing",
          "abstract": "We introduce DRESS, a novel approach for generating stylized large language model (LLM) responses through representation editing. Existing methods like prompting and fine-tuning are either insufficient for complex style adaptation or computationally expensive, particularly in tasks like NPC creation or character role-playing. Our approach leverages the over-parameterized nature of LLMs to disentangle a style-relevant subspace within the model's representation space to conduct representation editing, ensuring a minimal impact on the original semantics. By applying adaptive editing strengths, we dynamically adjust the steering vectors in the style subspace to maintain both stylistic fidelity and semantic integrity. We develop two stylized QA benchmark datasets to validate the effectiveness of DRESS, and the results demonstrate significant improvements compared to baseline methods such as prompting and ITI. In short, DRESS is a lightweight, train-free solution for enhancing LLMs with flexible and effective style control, making it particularly useful for developing stylized conversational agents. Codes and benchmark datasets are available at https://github.com/ArthurLeoM/DRESS-LLM.",
          "authors": [
            "Xinyu Ma",
            "Yifeng Xu",
            "Yang Lin",
            "Tianlong Wang",
            "Xu Chu",
            "Xin Gao",
            "Junfeng Zhao",
            "Yasha Wang"
          ],
          "decision": null,
          "pdate": 1737562655300,
          "cdate": 1727423324915
        }
      ]
    }
  },
  {
    "paper_id": 32700,
    "title": "3D Geometric Shape Assembly via Efficient Point Cloud Matching",
    "abstract": "Learning to assemble geometric shapes into a larger target structure is a pivotal task in various practical applications.\nIn this work, we tackle this problem by establishing local correspondences between point clouds of part shapes in both coarse- and fine-levels.\nTo this end, we introduce Proxy Match Transform (PMT), an approximate high-order feature transform layer that enables reliable matching between mating surfaces of parts while incurring low costs in memory and computation.\nBuilding upon PMT, we introduce a new framework, dubbed Proxy Match TransformeR (PMTR), for the geometric assembly task.\nWe evaluate the proposed PMTR on the large-scale 3D geometric shape assembly benchmark dataset of Breaking Bad and demonstrate its superior performance and efficiency compared to state-of-the-art methods. Project page:https://nahyuklee.github.io/pmtr.",
    "introduction": "Shape assembly aims to determine the precise placement of each constituent part and construct a larger target shape as a whole.\nThis task holds paramount significance, especially in the context of various applications encompassing robotics(Wang & Hauser,2019; Zakka et al.,2020; Zeng et al.,2021), manufacturing(Tian et al.,2022), computer graphics(Li et al.,2012), and computer-aided design(Chen et al.,2015; Jacobson,2017).\nDespite its pivotal role in industrial productivity and the plethora of applications, the field of shape assembly remains relatively underexplored in the literature due to the intricate challenge it presents:\ndemands a comprehensive understanding of geometric structures and analyses of pairwise relationships between local surfaces of input parts to establish accurate assembly.\n\n\nThere have been several recent attempts(Schor et al.,2019; Li et al.,2020a; Wu et al.,2020; Li et al.,2020b; Huang et al.,2020; Narayan et al.,2022; Chen et al.,2022; Wu et al.,2023b)to address the task of shape assembly, but these methods fall short of achieving accurate assembly.\nThey typically represent each part as a global embedding and perform regression to predict a placement for each part.\nThe global encoding strategy for each part, while simplifying the process, greatly limits local information by collapsing spatial resolutions, which is necessary to localize the mating surface.\nIndeed, accurate shape assembly requires a detailed analysis of both fine- and coarse-level spatial information of the parts in recognizing mating surfaces and establishing correspondences between the surfaces.\nTherefore, a promising approach would be to retain the spatially rich part representations during the encoding phase and analyze pairwise local correspondence relationships between them for reliable localization and matching of mating surfaces.\n\n\nIn the realm of correspondence analysis within image matching, prior methods(Rocco et al.,2018; Min & Cho,2021; Kim et al.,2022; Min et al.,2021; Rocco et al.,2020)typically utilize a high-order feature transform,i.e., high-dimensional convolution or attention, to achieve the objectives of localizing relevant instances and establishing correspondences between them.\nThe high-order feature transforms, which assess structural patterns of correlations in high-dimensional spaces, have been empirically validated for their efficacy in identifying accurate visual matches.\nHowever, the quadratic complexity with respect to input spatial resolution still remains as a significant drawback,limiting their application to only low-resolution (coarse-grained) inputs.\nSuch a limitation becomes particularly problematic in the context of geometric assembly since meticulous alignment between partsrequires analyzing high-resolution (fine-grained)to precisely identify ‘geometric compatibility’ between mating surfaces to match.\n\n\nIn this paper, we address this issue by introducing a new form of low-complexity high-order feature transform layer, dubbedProxy Match Transform (PMT), to tackle the challenges of geometric shape assembly.\nThe layer is designed to align analogous local embeddings in feature space,e.g., points on mating surfaces, with sub-quadratic complexity, thus offering a low-complexity yet high-order approach as illustrated in Fig.1.\nWe theoretically prove that the proposed PMT layer can effectively approximate the conventional high-order feature transforms(Rocco et al.,2018; Choy et al.,2020; Min & Cho,2021)under particular conditions.\nTo demonstrate its efficacy, we incorporate the PMT layer into a coarse-to-fine matching framework, Proxy Match TransformeR (PMTR), that uses PMTs for both coarse- and fine-level matching steps to establish correspondences on mating surfaces.\nWe compare our results with the recent state of the arts and provide a thorough performance analysis on the standard geometric shape assembly benchmark of Breaking Bad(Sellán et al.,2022).\nThe experiments demonstrate that our method significantly outperforms existing approaches while being computationally efficient compared to the baselines.\n\n\nOur main contributions can be summarized as follows:•We introduce Proxy Match Transform (PMT), a low-complexity high-order feature transform layer that effectively refines the matching of the feature pair.•Our theoretical analysis shows that PMT effectively approximates high-order feature transform while incurring sub-quadratic memory and time complexity.•Performance improvements in geometric shape assembly over the state-of-the-art baselines demonstrate the effectiveness and efficiency of our approach.\n\n",
    "related_work": "3D shape assembly & registration.Previous research in generative models for 3D objects has primarily focused on building objects through the combination of basic 3D primitives.\nA prevalent approach trains specialized models tailored to individual object classes, enabling the assembly of objects from volumetric primitives such as cuboids(Tulsiani et al.,2017).\nIn contrast,Khan et al. (2019)proposes a unified model that can generate cuboid primitives in various classes.\nAdditionally, variational autoencoders (VAEs) have been employed to model objects as compositions of cuboids, offering robust abstractions that distill local geometric details and elucidate object correspondences(Kingma & Welling,2014; Jones et al.,2020).\n\n\nParallel to these developments, research in the part assembly has aimed to construct complete objects from predefined semantic parts.\nThe method ofLi et al. (2020b)predicts translations and rotations for part point clouds to assemble a target object from an image reference.\nExtending this,Narayan et al. (2022); Huang et al. (2020)have conceptualized part assembly as a graph learning challenge, utilizing iterative message passing techniques to integrate parts into cohesive objects.\nThese approaches heavily rely on the PartNet dataset(Mo et al.,2019)to ensure semantic correspondence between the assembled parts and the target models, demonstrating that while geometric shapes are foundational, semantic cues can significantly guide and streamline the assembly process.\nOur research diverges from these methods by focusing on the assembly of parts without predefined semantics.\nA closely related methodology is that ofChen et al. (2022), which also tackles the problem of 3D shape assembly by integrating implicit shape reconstruction, providing a relevant benchmark.\n\n\nAdditionally, the concept of 3D shape assembly overlaps with the domain of 3D registration, especially in scenarios characterized by low overlap between a pair of point clouds.\nTechniques such as those proposed byHuang et al. (2021)andYu et al. (2021)leverage self-attention and cross-attention mechanisms within and across point cloud features to transform 3D features, facilitating enhanced matching accuracy.Qin et al. (2022)further advances this by mapping transformation-invariant data to the positional embeddings of transformer layers, optimizing the matching process in low-overlap conditions.\nDespite their efficacy, the practical application of these methods in fine-grained matching scenarios is often constrained by thequadratic complexityassociated with their matching layers, highlighting a critical area for improvement in computational efficiency and scalability.\nOur work addresses these challenges by proposing a novel approach that optimizes the computational demands of feature matching while maintaining high robustness.\n\n\nFigure 1:Given a correlation score at position(𝐱,𝐲)𝐱𝐲(\\mathbf{x},\\mathbf{y})( bold_x , bold_y )(the edge between highlighted nodes) and its neighboring scores (all other edges), vanilla high-order feature transform (shown on the left) leads to quadratic complexity due to its demand for memory-intensive pairwise correlation scores.\nThe product of two PMTs (shown on the right) effectively approximates this high-order transform only with sub-quadratic complexity by avoiding direct construction of correlation scores, instead exchanging information through a low-dimensional proxy tensor.\nThe red/blue nodes and black edges represent the source/target features and the correlation scores between them, respectively.\n\n\nHigh-order feature transform for matching.High-order feature transforms are essential in (both image and point cloud) matching tasks, helping to establish consensus among correspondences within a high-dimensional space.\nInitially introduced byRocco et al. (2018), the concept of a learning-based neighborhood consensus supports the identification of accurate matches by leveraging neighboring ambiguous matches between 2D images.\nThis approach has also been adapted for 3D registration tasks, notably byChoy et al. (2020), who utilized a 6D sparse convolutional layer to filter out outlier correspondences.\nGiven the high computational complexity associated with high-order feature transforms, several studies have proposed methods to reduce this burden.\nTechniques such as decomposing high-dimensional convolutional kernels(Min et al.,2021)and sparsifying the correlation map with top-k𝑘kitalic_kscores(Rocco et al.,2020)have been effective.\nFurther,Shi et al. (2023)enhanced matching efficiency by creating a sparse correlation matrix through the grouping of input tokens, significantly reducing the number of tokens involved.\nMore recent advances have integrated the self-attention mechanism to utilize global feature consensus effectively, although these methodologies, proposed byCho et al. (2021)andKim et al. (2022), come at a higher computational cost.\n\n\nOur work introduces the Proxy Match Transform (PMT), which simplifies existing high-order feature transforms to significantly reduce computational demands.\nWe apply PMT in a coarse-to-fine approach, identifying reliable correspondences between the mating surfaces of input parts and subsequently refining them for precise assembly.\nThere have been several approaches relevant to ours such as leveraging local geometric cues for assembly byLu et al. (2023), the linear approximations in convolutional networks byDenton et al. (2014), sparse attention mechanisms byZaheer et al. (2021), low-rank approximations of self-attention byChen et al. (2020), and Gaussian kernel approximations byChen et al. (2021).\nHowever, unlike these methods, which primarily improve processing within asingle feature, PMT uniquely addresses the challenge of efficient matching betweentwo distinct features, improving both computational efficiency and feature correspondence analysis, which are essential for diverse applications including geometric shape assembly.\n\n",
    "origin_citations": [
      "Learning shape abstractions by assembling volumetric primitives.",
      "Unsupervised primitive discovery for improved 3d generative modeling.",
      "Auto-encoding variational bayes.",
      "Shapeassembly: Learning to generate programs for 3d shape structure synthesis.",
      "Learning 3d part assembly from a single image.",
      "Rgl-net: A recurrent graph learning framework for progressive part assembly.",
      "Generative 3d part assembly via dynamic graph learning.",
      "Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding.",
      "Neural shape mating: Self-supervised object assembly with adversarial shape priors.",
      "Predator: Registration of 3d point clouds with low overlap.",
      "Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration.",
      "Geometric transformer for fast and robust point cloud registration.",
      "Neighbourhood consensus networks.",
      "Deep global registration.",
      "Hypercorrelation squeeze for few-shot segmentation.",
      "Efficient neighbourhood consensus networks via submanifold sparse convolutions.",
      "Clustergnn: Cluster-based coarse-to-fine graph neural network for efficient feature matching.",
      "Cats: Cost aggregation transformers for visual correspondence.",
      "Transformatcher:match-to-match attention for semantic correspondence.",
      "Jigsaw: Learning to assemble multiple fractured objects.",
      "Exploiting linear structure within convolutional networks for efficient evaluation.",
      "Big bird: Transformers for longer sequences.",
      "Compressed self-attention for deep metric learning with low-rank approximation.",
      "Skyformer: Remodel self-attention with gaussian kernel and nyström method."
    ],
    "citations": [
      155199,
      268902,
      5680,
      310842,
      253268,
      408998,
      349705,
      359670,
      145516,
      431419,
      512206,
      88324,
      386516,
      191629,
      29716,
      191956,
      368738,
      195778,
      147154,
      257147,
      6314,
      1499,
      511422
    ],
    "extract_rate": 0.9583333333333334,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 86,
      "Final_Score": 83.89999999999999,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 86\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 84.56,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "FYQIgQWH3d",
        "forum": "FYQIgQWH3d",
        "number": 9656,
        "title": "3D Geometric Shape Assembly via Efficient Point Cloud Matching",
        "normalized_title": "3d geometric shape assembly via efficient point cloud matching",
        "abstract": "Learning to assemble geometric shapes into a larger target structure is a pivotal task in various practical applications. In this work, we tackle this problem by establishing local correspondences between point clouds of part shapes in both coarse- and fine-levels. To this end, we introduce Proxy Match Transform (PMT), an approximate high-order feature transform layer that enables reliable matching between mating surfaces of parts while incurring low costs in memory and compute. Building upon PMT, we introduce a new framework, dubbed Proxy Match TransformeR (PMTR), for the geometric assembly task. We evaluate the proposed PMTR on the large-scale 3D geometric shape assembly benchmark dataset of Breaking Bad and demonstrate its superior performance and efficiency compared to state-of-the-art methods. Project page: https://nahyuklee.github.io/pmtr",
        "authors": [
          "Nahyuk Lee",
          "Juhong Min",
          "Junha Lee",
          "Seungwook Kim",
          "Kanghee Lee",
          "Jaesik Park",
          "Minsu Cho"
        ],
        "decision": null,
        "pdate": 1714610499552,
        "cdate": 1706873296185
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "FYQIgQWH3d",
          "forum": "FYQIgQWH3d",
          "number": 9656,
          "title": "3D Geometric Shape Assembly via Efficient Point Cloud Matching",
          "normalized_title": "3d geometric shape assembly via efficient point cloud matching",
          "abstract": "Learning to assemble geometric shapes into a larger target structure is a pivotal task in various practical applications. In this work, we tackle this problem by establishing local correspondences between point clouds of part shapes in both coarse- and fine-levels. To this end, we introduce Proxy Match Transform (PMT), an approximate high-order feature transform layer that enables reliable matching between mating surfaces of parts while incurring low costs in memory and compute. Building upon PMT, we introduce a new framework, dubbed Proxy Match TransformeR (PMTR), for the geometric assembly task. We evaluate the proposed PMTR on the large-scale 3D geometric shape assembly benchmark dataset of Breaking Bad and demonstrate its superior performance and efficiency compared to state-of-the-art methods. Project page: https://nahyuklee.github.io/pmtr",
          "authors": [
            "Nahyuk Lee",
            "Juhong Min",
            "Junha Lee",
            "Seungwook Kim",
            "Kanghee Lee",
            "Jaesik Park",
            "Minsu Cho"
          ],
          "decision": null,
          "pdate": 1714610499552,
          "cdate": 1706873296185
        }
      ]
    }
  },
  {
    "paper_id": 9346,
    "title": "InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists",
    "abstract": "Recent advances in generative diffusion models have enabled text-controlled synthesis of realistic and diverse images with impressive quality. Despite these remarkable advances, the application of text-to-image generative models in computer vision for standard visual recognition tasks remains limited. The current de facto approach for these tasks is to design model architectures and loss functions that are tailored to the task at hand. In this paper, we develop aunified language interfacefor computer vision tasks that abstracts away task-specific design choices and enables task execution by following natural language instructions. Our approach involves casting multiple computer vision tasks as text-to-image generation problems. Here, thetextrepresents aninstructiondescribing the task, and the resultingimageis a visually-encodedtask output. To train our model, we pool commonly-used computer vision datasets covering a range of tasks, including segmentation, object detection, depth estimation, and classification. We then use a large language model to paraphrase prompt templates that convey the specific tasks to be conducted on each image, and through this process, we create a multi-modal and multi-task training dataset comprising input and output images along with annotated instructions. Following the InstructPix2Pix architecture, we apply instruction-tuning to a text-to-image diffusion model using our constructed dataset, steering its functionality from a generative model to an instruction-guided multi-task vision learner. Experiments demonstrate that our model, dubbedInstructCV, performs competitively compared to other generalist and task-specific vision models. Moreover, it exhibits compelling generalization capabilities to unseen data, categories, and user instructions.",
    "introduction": "Recent work on text-to-image models has achieved impressive performance in image synthesis[1,2,3]. Particularly, diffusion models[4,5,6,7]have demonstrated remarkable capabilities of transforming diverse text prompts into realistic images, even for novel concepts. Models like DALL·E[2]and Stable Diffusion[8]highlight this progress, now finding use in real-world applications. However, despite these impressive results, generative text-to-image models have so far not been exploited as a unified basis for standard visual recognition tasks. Instead, the predominant approach for these tasks is to design dedicated task-specific architectures and loss functions[9,10,11], foregoing the opportunity to learn generalizable representations across heterogeneous problem domains and data landscapes.\n\n\nPrevious attempts to create unified models for computer vision tasks have predominantly relied onprompt tuningapproaches in conjunction with sequence-to-sequence architectures[9,12,13,10,11,14,15,16,17]. This general framework enables conditioning on input images as well as task-specific prompts by representing image pixels and (trainable) prompts as sequences of discrete tokens. When trained on multi-task datasets, the resulting output tokens align with the desired outcomes for the respective tasks prompted. One illustrative example of this approach is Pix2Seq[16], which follows an autoregressive language modeling approach for processing tokenized image pixels and task identifier codes. Another example includes a class of methods based onvisual prompting, which defines task-specific prompts in pixel space, unifying multiple vision tasks within a common “inpainting” framework[18,19]. In both examples, the task-specific prompts steer a single architecture to execute multiple vision tasks. However, these prompts consist of (uninterpretable) numerical values derived from specific training datasets, which may limit their ability to generalize to new datasets, tasks, or categories.\n\n\nIn this paper, we propose a unified model for computer vision tasks that conducts a given task by following natural language instructions (Fig.1). Our framework, dubbedInstructCV, repurposes generative text-to-image models to create a universal language interface for vision tasks. It does so by casting multiple computer vision tasks as text-to-image generation problems, where textual prompts (instructions) serve as explicit task descriptors, guiding the generation process to produce the visual task output corresponding to the input image. By conditioning on natural language descriptions of vision tasks, InstructCV enhances the representation of semantic coherence between images and language prompts, improving the model’s generalization capabilities to new human-written instructions and new categories compared to prior “generalist” vision models[16,17,12,20,15,21,22,23,24,25,13,26,18,19].\n\n\nTo train InstructCV, we follow aninstruction tuningapproach applied to a pretrained conditional diffusion model (Stable Diffusion). We generate the instruction tuning data by constructing a multi-modal, multi-task training dataset that comprises tuples of textual instructions, input images and visually-encoded task outputs. We do so by first combining several standard computer vision datasets across multiple tasks including segmentation, object detection, depth estimation, and classification. Next, in order to create heterogeneous and semantically rich textual instructions, we use a large language model (LLM) to paraphrase prompt templates for each vision task. Finally, we encode the output of the vision task associated with each instruction in the form of an output image (e.g., a masking pattern for semantic segmantation). Using this dataset, we utilize the InstructPix2Pix architecture[27]to instruction-tune a text-to-image diffusion model, transforming its functionality from a generative image synthesis model into an instruction-guided multi-task vision learner.\n\n\nOur experiments demonstrate that InstructCV achieves competitive results compared to other vision generalist and task-specific vision models. Particularly, InstructCV displays compelling generalization properties, surpassing the performance of state-of-the-art vision generalist models on external datasets as well as on unseen prompts in open-vocabulary segmentation tasks.\n\n\nFigure 2:Pictorial depiction of the InstructCV training pipeline.(a) We pool multiple computer vision datasets to construct a multi-modal and multi-task set of image pairs, where the target of each task is visually encoded in the form of an output image. Starting with a set of task-specific prompt templates, we sample a new instruction for each training point by using an LLM to rephrase the template for the corresponding task. (b) Using the dataset in (a), we finetune a diffusion model to produce the output𝐯⁢(𝐲)𝐯𝐲{\\bf v}({\\bf y})bold_v ( bold_y )given an image𝐱𝐱{\\bf x}bold_x& an instructionℐℐ\\mathcal{I}caligraphic_I.Figure 2:Pictorial depiction of the InstructCV training pipeline.(a) We pool multiple computer vision datasets to construct a multi-modal and multi-task set of image pairs, where the target of each task is visually encoded in the form of an output image. Starting with a set of task-specific prompt templates, we sample a new instruction for each training point by using an LLM to rephrase the template for the corresponding task. (b) Using the dataset in (a), we finetune a diffusion model to produce the output𝐯⁢(𝐲)𝐯𝐲{\\bf v}({\\bf y})bold_v ( bold_y )given an image𝐱𝐱{\\bf x}bold_x& an instructionℐℐ\\mathcal{I}caligraphic_I.\n\n",
    "related_work": "Repurposing diffusion models for vision tasks.Diffusion models have achieved impressive performance in image generation[36,4,38,39,40], text-to-image synthesis[6,8,2,1], as well as generation of other modalities such as video[41]and audio[42]. The idea of repurposing diffusion models to tackle standard computer vision tasks has been considered before to develop models for object detection[43]and open-vocabulary panoptic segmentation[44]. These approaches were limited to single-task settings with specialized loss functions and (unimodal) architectures. Contrarily, InstructCV provides a unified architecture for multi-task learning, with a natural language interface that enhances generalization to new datasets and categories. The idea of “instruction-tuning” text-to-image diffusion models was introduced in[27]with the objective of finetuning the model to follow editing instructions. InstructCV builds on this framework to adapt text-to-image models for performing conventional visual recognition tasks. To our knowledge, this is one of the earliest efforts in this direction.\n\n\nVision Generalists.Several prior attempts have aimed to develop unified models capable of executing multiple vision tasks within a single, shared architecture. Motivated by successes of LLMs, recent work has attempted to design such generalist models based on sequence-to-sequence architectures. Among these, models such as Florence[9], OFA[10], CoCa[11]and BEiT-3[14], learn general representation encoder, which require individual finetuning to each specific downstream task. Methods such as Unified-IO[15]and Pix2Seq-v2[17]build single architectures that are capable of performing multiple vision tasks via prompt tuning. However, the sequence-based operation of these models results in slow inference speeds, and the tuned prompts may not generalize to unseen datasets/categories. Another line of work proposes vision transformer-based architectures that frame different vision tasks as inpainting problems[18,19]. This work focuses on in-context learning based on visual prompts and does not consider language-based instructions, which we believe is a more natural interface for general-purpose models. To the best of our knowledge, the only generalist model that supports a language-based interface for vision tasks similar to that of InstructCV is the VisionLLM model developed in[13]. This is an LLM-based framework that treats images as a foreign language and aligns vision-centric tasks with language tasks that can be flexibly defined using language-based instructions. VisionLLM and InstructCV share a common objective but use different approaches. VisionLLM finetunes a pre-trained LLM using vision-centric tasks, whereas InstructCV finetunes a text-to-image model by substituting image captions with instructional text. We were unable to empirically compare the two models as the code for VisionLLM was not available at the time of writing this paper.\n\n\n\n\n",
    "origin_citations": [
      "Generative modeling by estimating gradients of the data distribution.",
      "Denoising diffusion probabilistic models.",
      "Cascaded diffusion models for high fidelity image generation.",
      "Palette: Image-to-image diffusion models.",
      "Image super-resolution via iterative refinement.",
      "GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models.",
      "High-resolution image synthesis with latent diffusion models.",
      "Hierarchical text-conditional image generation with clip latents.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Clipstyler: Image style transfer with a single text condition.",
      "Diffusiondet: Diffusion model for object detection, 2022.",
      "ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models.",
      "Instructpix2pix: Learning to follow image editing instructions.",
      "Florence: A new foundation model for computer vision, 2021.",
      "Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework.",
      "Coca: Contrastive captioners are image-text foundation models.",
      "Image as a foreign language: Beit pretraining for all vision and vision-language tasks, 2022.",
      "Unified-io: A unified model for vision, language, and multi-modal tasks.",
      "A unified sequence interface for vision tasks.",
      "Images speak in images: A generalist painter for in-context visual learning.",
      "Visual prompting via image inpainting.",
      "Visionllm: Large language model is also an open-ended decoder for vision-centric tasks, 2023."
    ],
    "citations": [
      5921,
      4827,
      5923,
      276339,
      894646,
      2559,
      2240,
      2558,
      5926,
      176197,
      153347,
      245286,
      2235,
      338632,
      89931,
      5246,
      1582,
      2471,
      609541,
      222952,
      3001,
      261177
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 88,
      "Final_Score": 83.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 84.52000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "Nu9mOSq7eH",
        "forum": "Nu9mOSq7eH",
        "number": 7458,
        "title": "InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists",
        "normalized_title": "instructcv instruction tuned text to image diffusion models as vision generalists",
        "abstract": "Recent advances in generative diffusion models have enabled text-controlled synthesis of realistic and diverse images with impressive quality. Despite these remarkable advances, the application of text-to-image generative models in computer vision for standard visual recognition tasks remains limited. The current de facto approach for these tasks is to design model architectures and loss functions that are tailored to the task at hand. In this paper, we develop a unified language interface for computer vision tasks that abstracts away task specific design choices and enables task execution by following natural language instructions. Our approach involves casting multiple computer vision tasks as text-to-image generation problems. Here, the text represents an instruction describing the task, and the resulting image is a visually-encoded task output. To train our model, we pool commonly-used computer vision datasets covering a range of tasks, including segmentation, object detection, depth estimation, and classification. We then use a large language model to paraphrase prompt templates that convey the specific tasks to be conducted on each image, and through this process, we create a multi-modal and multi-task training dataset comprising input and output images along with annotated instructions. Following the InstructPix2Pix architecture, we apply instruction-tuning to a text-to-image diffusion model using our constructed dataset, steering its functionality from a generative model to an instruction-guided multi-task vision learner. Experiments demonstrate that our model, dubbed InstructCV, performs competitively compared to other generalist and task-specific vision models. Moreover, it exhibits compelling generalization capabilities to unseen data, categories, and user instructions.",
        "authors": [
          "Yulu Gan",
          "Sungwoo Park",
          "Alexander Marcel Schubert",
          "Anthony Philippakis",
          "Ahmed Alaa"
        ],
        "decision": null,
        "pdate": 1705411010151,
        "cdate": 1695474312459
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "Nu9mOSq7eH",
          "forum": "Nu9mOSq7eH",
          "number": 7458,
          "title": "InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists",
          "normalized_title": "instructcv instruction tuned text to image diffusion models as vision generalists",
          "abstract": "Recent advances in generative diffusion models have enabled text-controlled synthesis of realistic and diverse images with impressive quality. Despite these remarkable advances, the application of text-to-image generative models in computer vision for standard visual recognition tasks remains limited. The current de facto approach for these tasks is to design model architectures and loss functions that are tailored to the task at hand. In this paper, we develop a unified language interface for computer vision tasks that abstracts away task specific design choices and enables task execution by following natural language instructions. Our approach involves casting multiple computer vision tasks as text-to-image generation problems. Here, the text represents an instruction describing the task, and the resulting image is a visually-encoded task output. To train our model, we pool commonly-used computer vision datasets covering a range of tasks, including segmentation, object detection, depth estimation, and classification. We then use a large language model to paraphrase prompt templates that convey the specific tasks to be conducted on each image, and through this process, we create a multi-modal and multi-task training dataset comprising input and output images along with annotated instructions. Following the InstructPix2Pix architecture, we apply instruction-tuning to a text-to-image diffusion model using our constructed dataset, steering its functionality from a generative model to an instruction-guided multi-task vision learner. Experiments demonstrate that our model, dubbed InstructCV, performs competitively compared to other generalist and task-specific vision models. Moreover, it exhibits compelling generalization capabilities to unseen data, categories, and user instructions.",
          "authors": [
            "Yulu Gan",
            "Sungwoo Park",
            "Alexander Marcel Schubert",
            "Anthony Philippakis",
            "Ahmed Alaa"
          ],
          "decision": null,
          "pdate": 1705411010151,
          "cdate": 1695474312459
        }
      ]
    }
  },
  {
    "paper_id": 46813,
    "title": "Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models",
    "abstract": "Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process.\nWhile learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations.\nTo overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing.\nIt employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder.\nThis method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary.\nWe demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations.\nAdditionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge.\nHierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.",
    "introduction": "Figure 1:Schematic of the proposed hierarchical architecture.\nThe input text is first split into words, with each word prepended by a special token [W]. These words are passed through the encoder and the activation at the position of the [W] token is selected as the word embedding𝐞isuperscript𝐞𝑖\\mathbf{e}^{i}bold_e start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPT. The sequence of word embeddings𝐞isuperscript𝐞𝑖\\mathbf{e}^{i}bold_e start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPTis then processed by the backbone to produce abstractpredictive word embeddings𝐩isuperscript𝐩𝑖\\mathbf{p}^{i}bold_p start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPT. The decoder then maps𝐩isuperscript𝐩𝑖\\mathbf{p}^{i}bold_p start_POSTSUPERSCRIPT italic_i end_POSTSUPERSCRIPTto probabilites for the characters of thenextword. During inference, given a text with a partial word, complete words are processed by the encoder and backbone. The decoder thenrecursivelycompletes the remaining characters of the incomplete word and the completed word enters the encoder in aword recursion.\n\n\nTokenization plays a fundamental role in natural language processing (NLP) as it breaks down text into units that computational models can process.\nTwo fundamental approaches are character-level and word-level tokenization.\nWhile character-level tokenization uses the “atomic” units of text and enjoys a small vocabulary size, it leads to long sequences with a high computational and memory cost.\nConversely, word-level tokenization leads to short sequences but suffers from extremely large vocabulary sizes and the inability to process out-of-vocabulary words.\n\n\nSubword tokenization has emerged as a compromise between these two extremes and has become the standard.\nCommon subword tokenizers are trained—separately from the model—on a reference corpus of text.\nFor example, Byte Pair Encoding(BPE; Gage,1994; Sennrich et al.,2016)builds a vocabulary starting from individual bytes and iteratively merging adjacent pairs of tokens that occur most frequently in the corpus until the desired vocabulary size is reached.\nThe resulting subword vocabulary leads to good sequence length compression on the reference corpus, while maintaining the ability to handle out-of-vocabulary words using a byte fallback.\n\n\nHowever, subword tokenizers come with several downsides.\nFirst, contemporary models routinely use vocabulary sizes in the hundreds of thousands, making the corresponding embedding matrices and output heads extremely large.\nFor instance, for the 8B model of the Llama-3 family(Dubey et al.,2024), with a vocabulary size of128128128128k, embedding and head account for roughly 13% of the model’s parameter footprint.\nSecondly, the tokenizer is fitted in a separate step and not included in the end-to-end learning process of the model.\nThis becomes problematic when a pretrained model is applied to, or finetuned on text from different domains or languages to which the tokenizer is not attuned(see, e.g., Ali et al.,2024; Petrov et al.,2023; Deiseroth et al.,2024).\nFinally, spelling mistakes or variations can lead to drastically different token sequences for semantically close inputs and thereby degrade model performance.\n\n\nTo address these shortcomings, we investigate a hierarchical architecture as shown in Figure1, which combines character-level and word-level processing. We first split the text into words.\nThe characters of each word are processed by a small character-level encoder module, which maps them to a word embedding.\nThe resulting sequence of word embeddings is then processed by a larger backbone model.\nThe outputs of the backbone are treated as abstract “predictive” word embeddings and are decoded back to characters by another small character-level decoder module.\nAs we will demonstrate, the character-level modules can be kept very small, having fewer parameters than the token-level embedding and head they replace.\nEncoder, backbone and decoder are transformer models and the entire system can be trained end-to-end, without the need for a fixed, trained tokenizer.\n\n\nOur core contributions are:•We revisit, refine and thoroughly investigate ahierarchical architecturefor autoregressive language modelling that combines character- and word-level processing.\nIts design eliminates the need for a fixed word or subword-level vocabulary, and does not require separate tokenizer training.\nThrough a careful computational cost analysis and comprehensive architecture sweeps, we identify optimal hierarchical model configurations across various compute budgets.•We conductextensive experiments, comparing our proposed model against state-of-the-art subword tokenizer-based models and a competing hierarchical approach in compute-matched experiments on identical data.\nWe demonstrate that our approach scales effectively to the 7B scale, consistently matching the performance of the baseline models.•We demonstrate that our proposed model is significantlymore robustto perturbations of its input.•Finally, we show that our model enjoyssuperior finetunabilityon out-of-distribution data, such as new languages or domains, outperforming tokenizer-based architectures which struggle with adaption to new vocabularies.\nIn particular, during continued pretraining on an unseen language, our model achieves superior performance on the target language, retains more of its previously acquired knowledge, while training almost twice as fast.\n\n\nThe remainder of this paper is organised as follows.\nSection2introduces our hierarchical architecture, including a detailed computational cost analysis.\nRelated work is discussed in Section3.\nIn Section4, we present our experimental setup and results.\nSection5concludes the paper with final remarks and potential directions for future research.\n\n",
    "related_work": "Thegranularityof computational models for natural language has been a prominent question in NLP research since its inception(see, e.g., a review by Mielke et al.,2021).\nFor many years, the question of character- vs word-level modelling has dominated the discussion, until subword tokenization methods, such as Byte-Pair Encoding(BPE, Sennrich et al.,2016), have emerged as a middle ground, balancing the flexibility of characters with the semantic coherence of words.\nOver time, subword tokenizers have become the dominant approach, gaining widespread adoption.\nThe current landscape of NLP largely treats subword tokens as fundamental, indivisible units that are determined through a preprocessing step prior to model training.\n\n\nSome prior works have augmented word or subword token embeddings with character-level information(e.g., Ma et al.,2020; El Boukkouri et al.,2020; Aguilar et al.,2021).\nMore recently, the T-FREE method(Deiseroth et al.,2024)operates at the word level while incorporating character information through specialized embedding and output layers.\nThese approaches have either not tackled generative modelling or still require a fixed vocabulary for generation.\n\n\nAnother line of prior work aims to enable character- or byte-level language modelling at contemporary scales.\nWhile some attempts have been made to train purely character-level transformers(Al-Rfou et al.,2019; Choe et al.,2019; Xue et al.,2022)these have ultimately not kept pace with subword-level models.\nMoving closer to the present work, some authors have presented “hybrid” approaches, e.g., using downsampling mechanisms(Charformer; Tay et al.,2022), or cross-attention with a latent “bottleneck” sequence(Perceiver AR; Hawthorne et al.,2022)to internally condense the large sequence lengths generated by character-level models.\nIn the following, we describe in detail the most closely-related works.\n\n\nSun et al. (2023)use a hierarchical character-word architecture for BERT-style masked language modelling.\nAs input to the decoder, the authors concatenate the backbone output for thei𝑖iitalic_i-th word with the per-character encoder outputs for thesameword.\nWe devise a generative variant of this architecture.\nWe introduce a shift by one, concatenating with embeddings of thenextword and use the raw character embeddings rather than the encoder outputs, since the bidirectional encoder would otherwise “leak” information about future characters.\nOur work also substantially scales up this modelling paradigm, experimenting with models up to the 7B parameter scale, compared toSun et al. (2023)who use models around the 100M scale.\n\n\nThe MegaByte architecture(Yu et al.,2023)uses a backbone-decoder architecture for generative modelling.\nInstead of splitting byte sequences into words, MegaByte chunks it into fixed-size patches of subsequent bytes.\nTheir architecture does not use an encoder module; the input to the backbone is simply the concatenation of the embeddings of the bytes within a patch.\nNote that this restricts the architecture to the use of fixed-size patches.\nAt the decoder, the backbone output isaddedto the byte sequences, which is prepended with a padding token.Yu et al. (2023)experiment with language models up to 320M parameters for the baseline model.\nWe compare our approach experimentally to MegaByte below.\n\n\nThawani et al. (2023)propose the hierarchical architecture most closely related to the present work.\nA notable difference is that they prepend not one but four [W] tokens to each word in order to increase model capacity when going from encoder to backbone.\nThis incurs drastically higher cost in en- and decoder compared to our approach of incrasing the hidden dimension.Thawani et al. (2023)experiment with models up to 77M parameters on datasets with fewer than 10M characters and a context window of only 192 characters (or the token equivalent thereof).\nUnfortunately, their experiments are not compute-matched and, by our calculation, assign 4x more compute to the hierarchical architecture compared to the baseline.\n\n\nFinally, in work that appeared concurrently with the preparation of the present paper,Slagle (2024)propose a byte-level model that applies additional Transformer layers to a subset of the input bytes.\nThey investigate both a fixed-size spacing as well as a split rule that marks only certain bytes as “split bytes”, including whitespaces and punctuation.\nThe byte-level layers are not restricted to individual words/chunks and instead use sliding window attention, precluding inference-time performance improvement via caching (AppendixA.4).\nExperiments are compute-matched and scaled up to the 1B models trained on 80B bytes and do not include downstream evaluations.\nNone of the above papers investigate robustness or finetunability.\n\n",
    "origin_citations": [
      "Between words and characters: A brief history of open-vocabulary modeling and tokenization in nlp.",
      "Neural machine translation of rare words with subword units.",
      "CharBERT: Character-aware pre-trained language model.",
      "CharacterBERT: Reconciling ELMo and BERT for word-level open-vocabulary representations from characters.",
      "Char2Subword: Extending the subword embedding space using robust character compositionality.",
      "T-FREE: tokenizer-free generative llms via sparse representations for memory-efficient embeddings.",
      "Character-level language modeling with deeper self-attention.",
      "Bridging the gap for tokenizer-free language models.",
      "ByT5: Towards a token-free future with pre-trained byte-to-byte models.",
      "Charformer: Fast character transformers via gradient-based subword tokenization.",
      "General-purpose, long-context autoregressive modeling with perceiver AR.",
      "From characters to words: Hierarchical pre-trained language model for open-vocabulary language understanding.",
      "MEGABYTE: predicting million-byte sequences with multiscale transformers.",
      "Learn your tokens: Word-pooled tokenization for language modeling.",
      "Spacebyte: Towards deleting tokenization from large language modeling."
    ],
    "citations": [
      537083,
      8559,
      199369,
      437032,
      435966,
      965584,
      524785,
      41456,
      395538,
      4737,
      876188,
      259275,
      114715,
      315322,
      134655
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 91,
      "Final_Score": 88.2,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 91\nFinal_Score: 88"
    },
    "related_work_scores": {
      "Content_Coherence": 75,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 75\nCitation_Quality: 85\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 84.47999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "tU074jg2vS",
        "forum": "tU074jg2vS",
        "number": 11586,
        "title": "Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models",
        "normalized_title": "hierarchical autoregressive transformers combining byte and word level processing for robust adaptable language models",
        "abstract": "Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process. While learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations. To overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing. It employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder. This method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary. We demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations. Additionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge. Hierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.",
        "authors": [
          "Pit Neitemeier",
          "Björn Deiseroth",
          "Constantin Eichenberg",
          "Lukas Balles"
        ],
        "decision": null,
        "pdate": 1737562788539,
        "cdate": 1727462053737
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "tU074jg2vS",
          "forum": "tU074jg2vS",
          "number": 11586,
          "title": "Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models",
          "normalized_title": "hierarchical autoregressive transformers combining byte and word level processing for robust adaptable language models",
          "abstract": "Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process. While learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations. To overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing. It employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder. This method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary. We demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations. Additionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge. Hierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.",
          "authors": [
            "Pit Neitemeier",
            "Björn Deiseroth",
            "Constantin Eichenberg",
            "Lukas Balles"
          ],
          "decision": null,
          "pdate": 1737562788539,
          "cdate": 1727462053737
        }
      ]
    }
  },
  {
    "paper_id": 27182,
    "title": "DiffCut: Catalyzing Zero-Shot Semantic Segmentation with Diffusion Features and Recursive Normalized Cut",
    "abstract": "Foundation models have emerged as powerful tools across various domains including language, vision, and multimodal tasks. While prior works have addressed unsupervised image segmentation, they significantly lag behind supervised models. In this paper, we use a diffusion UNet encoder as a foundation vision encoder and introduce DiffCut, an unsupervised zero-shot segmentation method that solely harnesses the output features from the final self-attention block. Through extensive experimentation, we demonstrate that the utilization of these diffusion features in a graph based segmentation algorithm, significantly outperforms previous state-of-the-art methods on zero-shot segmentation.\n Specifically, we leverage arecursive Normalized Cutalgorithm that softly regulates the granularity of detected objects and produces well-defined segmentation maps that precisely capture intricate image details. Our work highlights the remarkably accurate semantic knowledge embedded within diffusion UNet encoders that could then serve as foundation vision encoders for downstream tasks.Project page:https://diffcut-segmentation.github.ioFigure 1:Unsupervised zero-shot image segmentation.OurDiffCutmethod\n exploits features from a diffusion UNet encoder in a graph-basedrecursivepartitioning algorithm. Compared to DiffSeg[1],\n DiffCut provides finely detailed segmentation maps that more closely align with semantic concepts.",
    "introduction": "Foundation models have emerged as powerful tools across various domains, including language[2,3,4], vision[5,6,7], and multimodal tasks[8,9,10,11,12,13]. Pretrained on extensive datasets, these models exhibit unparalleled generalization capabilities, marking a significant departure from training models from scratch to efficiently adapting pretrained foundation models[14,15,16,17]. Utilizing pretrained models is particularly vital for dense visual tasks, alleviating the need for large annotated datasets specific to each domain.\n While prior works[18,19,20,21,22]have addressed unsupervised image segmentation, they significantly lag behind supervised models[23,24,25,26].\n Recently, SAM[27], proposed a model that can produce fine-grained class-agnostic masks which achieves outstanding zero-shot transfer to any images. Still, it requires a huge annotated segmentation dataset as well as significant training resources. Therefore, in this work, we investigate an alternative direction: unsupervised and zero-shot segmentation under the most constraining\n conditions, where no segmentation annotations or prior knowledge on the target dataset are available.\n\n\nRecently, several methods have emerged to address unsupervised object detection by framing it as a graph partitioning problem, utilizing self-supervised ViT features[28,5]. LOST[29]proposes to localize a unique object in a image by exploiting the inverse degree information to find a seed patch. TokenCut[30]splits the graph in two subsets given a bipartition. FOUND[31]and MaskCut[32]extend these approaches by addressing the single object discovery limitation. While being able to localize multiple objects, the latter methods remain constrained to identify a pre-determined number of objects, making them ill-suited for a task of unsupervised image segmentation which inherently requires to adapt the number of segment to uncover to the visual content.\n\n\nConversely, text-to-image diffusion models[33,34,35]can produce high-quality visual content from textual descriptions[36,37,38], indicating implicit learning of a wide range of visual concepts.\n Recent works have tried to leverage diverse internal representations of such models for localization or segmentation tasks. Several methods[39,40,41,42,43]opt to exploit image-text interactions within cross-attention modules but are ultimately constrained by the need for meticulous input prompt design. Concurrently,[44]identifies semantic correspondences between image pixels and spatial locations of low-dimensional feature maps by modulating cross-attention modules. This method proves to be computationally intensive as it requires numerous forward inferences. On the other hand, DiffSeg[1]segment images by iteratively merging self-attention maps which only depict local correlation between patches.\n\n\nIn this work, we introduce DiffCut, a new method for zero-shot image segmentation which solely harnesses the encoder features of a pre-trained diffusion model in arecursivegraph partitioning algorithm\n to produce fine-grained segmentation maps. Importantly, our method does not require any label from downstream segmentation datasets and its backbone has not been pre-trained on dense pixel annotations such as SAM[27].\n We observe inFig.1that DiffCut produces sharp segments that nicely outline object boundaries.\n In comparison with the recent state-of-the-art unsupervised zero-shot segmentation method DiffSeg[1], the segments yielded by DiffCut, are better aligned with the semantic visual concepts,e.g.DiffCut is able to uncover the urban area as well as the boats in the middle row image.\nOur main contributions are as follows:\n\n\n•We leverage the features from the final self-attention block of a diffusion UNet encoder, for the task of unsupervised image segmentation. In this context, we demonstrate that exploiting the inner patch-level alignment yields superior performance compared to merging self-attention maps as done in DiffSeg[1].•Compared to existing graph based object localization methodse.g.TokenCut or MaskCut[30,32], we push further and take advantage of arecursive Normalized Cutalgorithm to generate dense segmentation maps. Via a partitioning threshold, the method is able to regulate the granularity of detected objects and consequently adapt the number of segments to the visual content.•We perform extensive experiments to validate the effectiveness of DiffCut and show that it significantly outperforms state-of-the-art methods for unsupervised segmentation on standard benchmarks, reducing the gap with fully supervised models.\n\n\nIn addition, we exhibit the remarkable semantic coherence emerging in our chosen diffusion features by measuring their patch-level alignment, which surpasses other backbones such as CLIP[8]or DINOv2[5]. Our ablation studies further reveal the relevance of these diffusion features as well as therecursivepartitioning approach which proves to provide robust segmentation performance. Finally, we show that DiffCut can be extended to an open-vocabulary setting with a straightforward process leveraging aconvolutionalCLIP, which even tops most dedicated methods on this task.\n\n",
    "related_work": "Semantic segmentation.Semantic segmentation consists in partitioning an image into a set of segments, each corresponding to a specific semantic concept. While supervised semantic segmentation has been widely explored[45,46,47,27], unsupervised and zero-shot transfer segmentation for any images with previously unseen categories remains significantly more challenging and much less investigated.\n For example, most works in unsupervised segmentation require access to the target data for unsupervised adaption[21,20,19,18]. Therefore, these methods cannot segment images that are not seen during the adaptation. Recently, DiffSeg[1]moved a step forward by proposing an unsupervised and zero-shot approach that can produce quality segmentation maps without any prior knowledge on the underlying visual content.\n\n\nSegmentation with Text Supervision.Recent works have shown that learning accurate segmentation maps is possible with text supervision, overcoming the cost of dense annotations. These works are mostly based on image-text contrastive learning[48,49,50,51], and usually exploit the features of CLIP[52,53,54]. MaskCLIP[52]leverages CLIP to get pseudo labels used to train a typical image segmentation model. ReCO[53]uses CLIP for dataset curation and get a reference image embedding for each class that is used to obtain the final segmentation. CLIPpy[48]proposes minimal modifications to CLIP to get dense labels. SegCLIP[54]continues to train CLIP with additional reconstruction and superpixel-based KL loss to enhance localization. TCL[50]learns a region-text alignment to get precise segmentation masks. GroupViT[49]also learns masks from text supervision and is based on a hierarchical grouping mechanism. Similarly, ViewCo[51]proposes a contrastive learning between multiple views/crops of the image and the text.\n\n\nGraph-based Object Detection.Built on top of self-supervised ViT features, various methods frame the problem of object detection as a graph partitioning problem. LOST[29]aims at detecting salient object in an image using the degree of the nodes in the graph and a seed expansion mechanism. Based onNormalized Cut (NCut)[55], FOUND[31]proposes to identify all background patches, hence discovering all object patches as a by-product with no need for a prior knowledge of the number of objects or their relative size with respect to the background. TokenCut[30]detects one single salient object in each image with a uniqueNCutbipartition. In an attempt to adapt TokenCut to multi-objects localization, MaskCut[32]first localizes an object and disconnects its corresponding patches to the rest of the graph before repeating the process a pre-determined number of times.\n As these graph partitioning methods are only able to uncover a fixed number of segments, they are inadequate for a task of image segmentation.\n\n\nSegmentation with Diffusion Models.Diffusion models can produce high-quality visual content given a text prompt, indicating implicit learning of a wide range of visual concepts and the ability of grounding these concepts in images. Therefore their internal representations appear as good candidates for visual localization tasks[56,57,58].\n ODISE[59]is one of the first training-based approaches to build a fully supervised panoptic image segmentor on top of diffusion features. Several other methods[40,41,42]leverage attention modules for localization or segmentation tasks. DiffuMask[42]uses the cross-modal grounding between a text input and an image in cross-attention modules to segment the referred object in a synthetic image. However, DiffuMask can only be applied to a generated image.\n In a zero-shot setting,[41]harnesses the image-text interaction via cross-attention score maps to complete self-attention maps and segment grounded objects. EmerDiff[44]opts not to exploit image-text interactions in cross-attention modules. Instead, it identifies semantic correspondences between image pixels and spatial locations by modulating the values of a sub-region of feature maps in low-resolution cross-attention layers. These cross-attention based methods eventually prove to be highly computationally intensive as multiple forward inferences are often required. On the other hand, DiffSeg[1]proposes an iterative merging process based on measuring KL divergence among self-attention maps to merge them into valid segmentation masks. However, it appears that self-attention score maps only depict very local correlation between patches.\n\n",
    "origin_citations": [
      "Masked-attention mask transformer for universal image segmentation.",
      "Max-deeplab: End-to-end panoptic segmentation with mask transformers.",
      "Per-pixel classification is not all you need for semantic segmentation.",
      "Segment anything.",
      "Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering.",
      "Unsupervised semantic segmentation by distilling feature correspondences.",
      "Invariant information clustering for unsupervised image classification and segmentation.",
      "Acseg: Adaptive conceptualization for unsupervised semantic segmentation.",
      "Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion.",
      "Perceptual grouping in contrastive vision-language models.",
      "Groupvit: Semantic segmentation emerges from text supervision.",
      "Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs.",
      "Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency.",
      "Extract free dense labels from clip.",
      "Reco: Retrieve and co-segment for zero-shot transfer.",
      "Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation.",
      "Localizing objects with self-supervised transformers and no labels.",
      "Normalized cuts and image segmentation.",
      "Unsupervised object localization: Observing the background to discover objects.",
      "Tokencut: Segmenting objects in images and videos with self-supervised transformer and normalized cut.",
      "Cut and learn for unsupervised object detection and instance segmentation.",
      "A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence.",
      "Diffusiondet: Diffusion model for object detection.",
      "Diffedit: Diffusion-based semantic image editing with mask guidance.",
      "Open-vocabulary panoptic segmentation with text-to-image diffusion models.",
      "Diffusionseg: Adapting diffusion towards unsupervised object discovery.",
      "Diffusion model is secretly a training-free open vocabulary semantic segmenter.",
      "Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models.",
      "Emerdiff: Emerging pixel-level semantic knowledge in diffusion models."
    ],
    "citations": [
      152903,
      4768,
      1756,
      2512,
      506173,
      503116,
      113280,
      459243,
      405099,
      336467,
      174081,
      223717,
      819482,
      153057,
      411138,
      242928,
      364391,
      911948,
      230880,
      212967,
      258742,
      245286,
      335814,
      77161,
      495861,
      129579,
      494911,
      125456
    ],
    "extract_rate": 0.9655172413793104,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 80,
      "Final_Score": 83.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 80\nFinal_Score: 83.5"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 78,
      "Final_Score": 85.1,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 92\nSynthesis_Analysis: 78\nFinal_Score: 85"
    },
    "overall_score": 84.46,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "N0xNf9Qqmc",
        "forum": "N0xNf9Qqmc",
        "number": 3966,
        "title": "DiffCut: Catalyzing Zero-Shot Semantic Segmentation with Diffusion Features and Recursive Normalized Cut",
        "normalized_title": "diffcut catalyzing zero shot semantic segmentation with diffusion features and recursive normalized cut",
        "abstract": "Foundation models have emerged as powerful tools across various domains including language, vision, and multimodal tasks. While prior works have addressed unsupervised semantic segmentation, they significantly lag behind supervised models. In this paper, we use a diffusion UNet encoder as a foundation vision encoder and introduce DiffCut, an unsupervised zero-shot segmentation method that solely harnesses the output features from the final self-attention block. Through extensive experimentation, we demonstrate that using these diffusion features in a graph based segmentation algorithm, significantly outperforms previous state-of-the-art methods on zero-shot segmentation. Specifically, we leverage a recursive Normalized Cut algorithm that regulates the granularity of detected objects and produces well-defined segmentation maps that precisely capture intricate image details. Our work highlights the remarkably accurate semantic knowledge embedded within diffusion UNet encoders that could then serve as foundation vision encoders for downstream tasks.",
        "authors": [
          "Paul Couairon",
          "Mustafa Shukor",
          "Jean-Emmanuel HAUGEARD",
          "Matthieu Cord",
          "Nicolas THOME"
        ],
        "decision": null,
        "pdate": 1727287735928,
        "cdate": 1715345257842
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "N0xNf9Qqmc",
          "forum": "N0xNf9Qqmc",
          "number": 3966,
          "title": "DiffCut: Catalyzing Zero-Shot Semantic Segmentation with Diffusion Features and Recursive Normalized Cut",
          "normalized_title": "diffcut catalyzing zero shot semantic segmentation with diffusion features and recursive normalized cut",
          "abstract": "Foundation models have emerged as powerful tools across various domains including language, vision, and multimodal tasks. While prior works have addressed unsupervised semantic segmentation, they significantly lag behind supervised models. In this paper, we use a diffusion UNet encoder as a foundation vision encoder and introduce DiffCut, an unsupervised zero-shot segmentation method that solely harnesses the output features from the final self-attention block. Through extensive experimentation, we demonstrate that using these diffusion features in a graph based segmentation algorithm, significantly outperforms previous state-of-the-art methods on zero-shot segmentation. Specifically, we leverage a recursive Normalized Cut algorithm that regulates the granularity of detected objects and produces well-defined segmentation maps that precisely capture intricate image details. Our work highlights the remarkably accurate semantic knowledge embedded within diffusion UNet encoders that could then serve as foundation vision encoders for downstream tasks.",
          "authors": [
            "Paul Couairon",
            "Mustafa Shukor",
            "Jean-Emmanuel HAUGEARD",
            "Matthieu Cord",
            "Nicolas THOME"
          ],
          "decision": null,
          "pdate": 1727287735928,
          "cdate": 1715345257842
        }
      ]
    }
  },
  {
    "paper_id": 29619,
    "title": "Approximately Equivariant Neural Processes",
    "abstract": "Equivariant deep learning architectures exploit symmetries in learning problems to improve the sample efficiency of neural-network-based models and their ability to generalise. However, when modelling real-world data, learning problems are often notexactlyequivariant, but only approximately. For example, when estimating the global temperature field from weather station observations, local topographical features like mountains break translation equivariance.\nIn these scenarios, it is desirable to construct architectures that can flexibly depart from exact equivariance in a data-driven way. Current approaches to achieving this cannot usually be applied out-of-the-box to any architecture and symmetry group. In this paper, we develop a general approach to achieving this using existing equivariant architectures.\nOur approach is agnostic to both the choice of symmetry group and model architecture, making it widely applicable. We consider the use of approximately equivariant architectures in neural processes (NPs), a popular family of meta-learning models. We demonstrate the effectiveness of our approach on a number of synthetic and real-world regression experiments, showing that approximately equivariant NP models can outperform both their non-equivariant and strictly equivariant counterparts.",
    "introduction": "The development of equivariant deep learning architectures has spearheaded many advancements in machine learning, including CNNs(LeCun et al.,1989), group equivariant CNNs(Cohen and Welling,2016a; Finzi et al.,2020), transformers(Vaswani et al.,2017), DeepSets(Zaheer et al.,2017), and GNNs(Scarselli et al.,2008). When appropriate, equivariances provide useful inductive biases that can drastically improve sample complexity(Mei et al.,2021)and generalisation capabilities(Elesedy and Zaidi,2021; Bulusu et al.,2021; Petrache and Trivedi,2023)by exploiting symmetries present in the data. Yet, real-world data are seldom strictly equivariant. As an example, consider modelling daily precipitation across space and time. Such data may be close to translation equivariant—and therefore translation equivariance serves as a useful inductive bias—however, it is clear that it is not strictly translation equivariant due to geographical and seasonal variations. More generally, whilst there are aspects of modelling problems which are universal and exhibit symmetries such as equivariance (e.g. atmospheric physics) there are often unknown local factors (e.g. precise local topography) which break these symmetries and which the models do not have access to. It is thus desirable for the model to be able to depart from strict equivariance when necessary; that is, to developapproximately equivariantmodels.\n\n\nA family of models that have benefited greatly from the use of equivariant deep learning architectures are neural processes(NPs; Garnelo et al.,2018a,b). However, many of the problem domains that NPs are applied to exhibit only approximate equivariance. It is therefore desirable to build approximately equivariant NPs through the use of approximately equivariant deep learning architectures. Whilst there exist several approaches to constructing approximately equivariant architectures, they are limited to CNN- and MLP-based models(Wang et al.,2022a; van der Ouderaa et al.,2022; Finzi et al.,2021; Romero and Lohit,2022)—which restricts their applicability to certain symmetry groups—and often require modifications to the loss function used to train the models(Finzi et al.,2021; Kim et al.,2023a). As there exist NPs that utilise a variety of architectures, such as transformers and GNNs(Carr and Wingate,2019; Nguyen and Grover,2022; Kim et al.,2019; Feng et al.,2022), these approaches are not sufficient. We address this shortcoming through the development of an approach to constructing approximately equivariant models that is both architecture and symmetry-group agnostic. Importantly, our approach can be realised without modifying the core architecture of existing equivariant models, enabling use in a variety of equivariant NPs such as the convolutional conditional NP(ConvCNP; Gordon et al.,2019), the equivariant CNP(EquivCNP; Kawano et al.,2021), the steerable CNP(SteerCNP; Holderrieth et al.,2021), the translation equivariant transformer NP(TE-TNP; Ashman et al.,2024), and the relational CNP(RCNP; Huang et al.,2023).\n\n\nWe outline our core contributions as follows:1.We demonstrate that under certain technical conditions, such as Hölder regularity and compactness, any non-equivariant mapping between function spaces can be approximated by an equivariant mapping with fixed additional inputs (Theorems3and2). This provides insight into how to construct approximately equivariant models, and generalises several existing approaches to relaxing equivariant constraints.2.We apply this result and the insights it provides to construct approximately equivariant versions of several popular equivariant NPs. The modifications required arevery simple, yet effective: we demonstrate improved performance relative to both strictly equivariant and non-equivariant models on a number of spatio-temporal regression tasks.\n\n",
    "related_work": "We consider the supervised learning setting, where𝒳𝒳\\mathcal{X}caligraphic_X,𝒴𝒴\\mathcal{Y}caligraphic_Ydenote the input and output spaces, and(𝐱,𝐲)∈𝒳×𝒴𝐱𝐲𝒳𝒴(\\mathbf{x},\\mathbf{y})\\in\\mathcal{X}\\times\\mathcal{Y}( bold_x , bold_y ) ∈ caligraphic_X × caligraphic_Ydenotes an input-output pair. Let𝒮=⋃N=0∞(𝒳×𝒴)N𝒮superscriptsubscript𝑁0superscript𝒳𝒴𝑁\\mathcal{S}=\\bigcup_{N=0}^{\\infty}(\\mathcal{X}\\times\\mathcal{Y})^{N}caligraphic_S = ⋃ start_POSTSUBSCRIPT italic_N = 0 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ∞ end_POSTSUPERSCRIPT ( caligraphic_X × caligraphic_Y ) start_POSTSUPERSCRIPT italic_N end_POSTSUPERSCRIPTbe a collection of all finite data sets, which includes the empty set∅\\varnothing∅. Let𝒮M=(𝒳×𝒴)Msubscript𝒮𝑀superscript𝒳𝒴𝑀\\mathcal{S}_{M}=\\left(\\mathcal{X}\\times\\mathcal{Y}\\right)^{M}caligraphic_S start_POSTSUBSCRIPT italic_M end_POSTSUBSCRIPT = ( caligraphic_X × caligraphic_Y ) start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPTbe the collection ofM𝑀Mitalic_Minput-output pairs and𝒮≤M=⋃m=1M𝒮msubscript𝒮absent𝑀superscriptsubscript𝑚1𝑀subscript𝒮𝑚\\mathcal{S}_{\\leq M}=\\bigcup_{m=1}^{M}\\mathcal{S}_{m}caligraphic_S start_POSTSUBSCRIPT ≤ italic_M end_POSTSUBSCRIPT = ⋃ start_POSTSUBSCRIPT italic_m = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_M end_POSTSUPERSCRIPT caligraphic_S start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPTbe the collection of at mostM𝑀Mitalic_Mpairs. We denote a context and target set with𝒟c,𝒟t∈𝒮subscript𝒟𝑐subscript𝒟𝑡𝒮\\mathcal{D}_{c},\\ \\mathcal{D}_{t}\\in\\mathcal{S}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ caligraphic_S, where|𝒟c|=Ncsubscript𝒟𝑐subscript𝑁𝑐|\\mathcal{D}_{c}|=N_{c}| caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT | = italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT,|𝒟t|=Ntsubscript𝒟𝑡subscript𝑁𝑡|\\mathcal{D}_{t}|=N_{t}| caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT | = italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT. Let𝐗c∈(𝒳)Ncsubscript𝐗𝑐superscript𝒳subscript𝑁𝑐\\mathbf{X}_{c}\\in(\\mathcal{X})^{N_{c}}bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ∈ ( caligraphic_X ) start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝐘c∈(𝒴)Ncsubscript𝐘𝑐superscript𝒴subscript𝑁𝑐\\mathbf{Y}_{c}\\in(\\mathcal{Y})^{N_{c}}bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ∈ ( caligraphic_Y ) start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT end_POSTSUPERSCRIPTbe the inputs and corresponding outputs of𝒟csubscript𝒟𝑐\\mathcal{D}_{c}caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT, and let𝐗t∈(𝒳)Ntsubscript𝐗𝑡superscript𝒳subscript𝑁𝑡\\mathbf{X}_{t}\\in(\\mathcal{X})^{N_{t}}bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ ( caligraphic_X ) start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPT,𝐘t∈(𝒴)Ntsubscript𝐘𝑡superscript𝒴subscript𝑁𝑡\\mathbf{Y}_{t}\\in(\\mathcal{Y})^{N_{t}}bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ∈ ( caligraphic_Y ) start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPTbe defined analogously. We denote a single task asξ=(𝒟c,𝒟t)=((𝐗c,𝐘c),(𝐗t,𝐘t))𝜉subscript𝒟𝑐subscript𝒟𝑡subscript𝐗𝑐subscript𝐘𝑐subscript𝐗𝑡subscript𝐘𝑡\\xi=(\\mathcal{D}_{c},\\mathcal{D}_{t})=((\\mathbf{X}_{c},\\mathbf{Y}_{c}),(%\n\\mathbf{X}_{t},\\mathbf{Y}_{t}))italic_ξ = ( caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) = ( ( bold_X start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT , bold_Y start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) , ( bold_X start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT , bold_Y start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) ). Let𝒫⁢(𝒳)𝒫𝒳\\mathcal{P}(\\mathcal{X})caligraphic_P ( caligraphic_X )denote the collection of𝒴𝒴\\mathcal{Y}caligraphic_Y-valued stochastic processes on𝒳𝒳\\mathcal{X}caligraphic_X. LetΘΘ\\Thetaroman_Θdenote the parameter space for some family of probability densities, e.g. means and variances for the space of all Gaussian densities.\n\n\n2.1Neural ProcessesNPs(Garnelo et al.,2018a,b)can be viewed as neural-network-based parametrisations ofprediction mapsπ:𝒮→𝒫⁢(𝒳):𝜋→𝒮𝒫𝒳\\pi\\colon\\mathcal{S}\\to\\mathcal{P}(\\mathcal{X})italic_π : caligraphic_S → caligraphic_P ( caligraphic_X )from data sets𝒮𝒮\\mathcal{S}caligraphic_Sto predictions𝒫⁢(𝒳)𝒫𝒳\\mathcal{P}(\\mathcal{X})caligraphic_P ( caligraphic_X ), where predictions are represented by𝒴𝒴\\mathcal{Y}caligraphic_Y-valued stochastic processes on𝒳𝒳\\mathcal{X}caligraphic_X(Foong et al.,2020; Bruinsma,2022). Throughout, we denote the density of the finite-dimensional distribution ofπ⁢(𝒟)𝜋𝒟\\pi(\\mathcal{D})italic_π ( caligraphic_D )at inputs𝐗𝐗\\mathbf{X}bold_Xbyp(⋅∣𝐗,𝒟)p(\\,\\cdot\\mid\\mathbf{X},\\mathcal{D})italic_p ( ⋅ ∣ bold_X , caligraphic_D ).\nIn this work, we restrict our attention to conditional NPs(CNPs; Garnelo et al.,2018a), which only target marginal predictive distributions by assuming that the predictive densities factorise:p⁢(𝐘|𝐗,𝒟)=∏np⁢(𝐲n|𝐱n,𝒟)𝑝conditional𝐘𝐗𝒟subscriptproduct𝑛𝑝conditionalsubscript𝐲𝑛subscript𝐱𝑛𝒟p(\\mathbf{Y}|\\mathbf{X},\\mathcal{D})=\\prod_{n}p(\\mathbf{y}_{n}|\\mathbf{x}_{n},%\n\\mathcal{D})italic_p ( bold_Y | bold_X , caligraphic_D ) = ∏ start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT italic_p ( bold_y start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT | bold_x start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , caligraphic_D ). We denote all parameters of a CNP byω𝜔\\omegaitalic_ω.\nCNPs are trained in a meta-learning fashion, in which the expected predictive log-probability is maximised:ωML=arg⁢maxω⁡ℒML⁢(ω)whereℒML⁢(ω)=𝔼p⁢(ξ)⁢[∑n=1Ntlog⁡pω⁢(𝐲t,n|𝐱t,n,𝒟c)].formulae-sequencesubscript𝜔MLsubscriptargmax𝜔subscriptℒML𝜔wheresubscriptℒML𝜔subscript𝔼𝑝𝜉delimited-[]superscriptsubscript𝑛1subscript𝑁𝑡subscript𝑝𝜔conditionalsubscript𝐲𝑡𝑛subscript𝐱𝑡𝑛subscript𝒟𝑐\\textstyle\\omega_{\\text{ML}}=\\operatorname*{arg\\,max}_{\\omega}\\mathcal{L}_{%\n\\text{ML}}(\\omega)\\quad\\text{where}\\quad\\mathcal{L}_{\\text{ML}}(\\omega)=%\n\\mathbb{E}_{p(\\xi)}\\big{[}\\sum_{n=1}^{N_{t}}\\log p_{\\omega}(\\mathbf{y}_{t,n}|%\n\\mathbf{x}_{t,n},\\mathcal{D}_{c})\\big{]}.italic_ω start_POSTSUBSCRIPT ML end_POSTSUBSCRIPT = start_OPERATOR roman_arg roman_max end_OPERATOR start_POSTSUBSCRIPT italic_ω end_POSTSUBSCRIPT caligraphic_L start_POSTSUBSCRIPT ML end_POSTSUBSCRIPT ( italic_ω ) where caligraphic_L start_POSTSUBSCRIPT ML end_POSTSUBSCRIPT ( italic_ω ) = blackboard_E start_POSTSUBSCRIPT italic_p ( italic_ξ ) end_POSTSUBSCRIPT [ ∑ start_POSTSUBSCRIPT italic_n = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_N start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT end_POSTSUPERSCRIPT roman_log italic_p start_POSTSUBSCRIPT italic_ω end_POSTSUBSCRIPT ( bold_y start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT | bold_x start_POSTSUBSCRIPT italic_t , italic_n end_POSTSUBSCRIPT , caligraphic_D start_POSTSUBSCRIPT italic_c end_POSTSUBSCRIPT ) ] .(1)Here, the expectation is taken with respect to the distribution over tasksp⁢(ξ)𝑝𝜉p(\\xi)italic_p ( italic_ξ ). In practice, we only have access to a finite number of tasks for training, so the expectation is approximated with an average over tasks.\nThe global maximum is achieved if and only if the model recovers the ground-truth marginals(Proposition 3.26 by Bruinsma,2022).\n\n\n2.2Group EquivarianceWe consider equivariance with respect to transformations in some groupG𝐺Gitalic_G. For example,G𝐺Gitalic_Gcan be the group of translations or the group of rotations. Mathematically, a groupG𝐺Gitalic_Gis a set endowed with a binary operationG×G→G→𝐺𝐺𝐺G\\times G\\to Gitalic_G × italic_G → italic_G(denoted as multiplication) such that(i)(f⁢g)⁢h=f⁢(g⁢h)𝑓𝑔ℎ𝑓𝑔ℎ(fg)h=f(gh)( italic_f italic_g ) italic_h = italic_f ( italic_g italic_h )for allf,g,h∈G𝑓𝑔ℎ𝐺f,g,h\\in Gitalic_f , italic_g , italic_h ∈ italic_G;(ii)there exists an identity elemente∈G𝑒𝐺e\\in Gitalic_e ∈ italic_Gsuch thate⁢g=g⁢e=g𝑒𝑔𝑔𝑒𝑔eg=ge=gitalic_e italic_g = italic_g italic_e = italic_gfor allg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G; and(iii)every elementg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_Ghas an inverseg−1∈Gsuperscript𝑔1𝐺g^{-1}\\in Gitalic_g start_POSTSUPERSCRIPT - 1 end_POSTSUPERSCRIPT ∈ italic_G.AG𝐺Gitalic_G-space is a spaceX𝑋Xitalic_Xfor which there exists a functionG×X→X→𝐺𝑋𝑋G\\times X\\to Xitalic_G × italic_X → italic_Xcalled a group action (again denoted by multiplication) such that(i)e⁢x=x𝑒𝑥𝑥ex=xitalic_e italic_x = italic_xfor allx∈X𝑥𝑋x\\in Xitalic_x ∈ italic_X; and(ii)f⁢(g⁢x)=(f⁢g)⁢x𝑓𝑔𝑥𝑓𝑔𝑥f(gx)=(fg)xitalic_f ( italic_g italic_x ) = ( italic_f italic_g ) italic_xfor allf,g∈G𝑓𝑔𝐺f,g\\in Gitalic_f , italic_g ∈ italic_Gandx∈X𝑥𝑋x\\in Xitalic_x ∈ italic_X.The notion of group equivariance is used to describe mappings for which, when the input to the mapping is transformed by someg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G, the output of the mapping is transformed equivalently. This is formalised in the following definition.Definition 1(G𝐺Gitalic_G-equivariance).LetX𝑋Xitalic_XandY𝑌Yitalic_YbeG𝐺Gitalic_G-spaces. Call a mappingρ:X→Y:𝜌→𝑋𝑌\\rho\\colon X\\rightarrow Yitalic_ρ : italic_X → italic_YG𝐺Gitalic_G-equivariant ifρ⁢(g⁢x)=g⁢ρ⁢(x)𝜌𝑔𝑥𝑔𝜌𝑥\\rho(gx)=g\\rho(x)italic_ρ ( italic_g italic_x ) = italic_g italic_ρ ( italic_x )for allg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_Gandx∈X𝑥𝑋x\\in Xitalic_x ∈ italic_X.\n\n\n2.3Group-Equivariant Conditional Neural ProcessesThe property ofG𝐺Gitalic_G-equivariance is particularly useful in NPs when modelling ground-truth stochastic processes that exhibitG𝐺Gitalic_G-stationarity:Definition 2(G𝐺Gitalic_G-stationary stochastic process).Let𝒳𝒳\\mathcal{X}caligraphic_Xbe aG𝐺Gitalic_G-space.\nWe say that a stochastic processP∈𝒫⁢(𝒳)𝑃𝒫𝒳P\\in\\mathcal{P}(\\mathcal{X})italic_P ∈ caligraphic_P ( caligraphic_X )isG𝐺Gitalic_G-stationary if, forf∼Psimilar-to𝑓𝑃f\\sim Pitalic_f ∼ italic_Pand allg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G,f⁢(⋅)𝑓⋅f(\\,\\cdot\\,)italic_f ( ⋅ )is equal in distribution tof(g⋅)f(g\\,\\cdot\\,)italic_f ( italic_g ⋅ ).LetP∈𝒫⁢(𝒳)𝑃𝒫𝒳P\\in\\mathcal{P}(\\mathcal{X})italic_P ∈ caligraphic_P ( caligraphic_X )be a ground-truth stochastic process. For a dataset𝒟∈𝒮𝒟𝒮\\mathcal{D}\\in\\mathcal{S}caligraphic_D ∈ caligraphic_S, defineπP⁢(𝒟)subscript𝜋𝑃𝒟\\pi_{P}(\\mathcal{D})italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D )by integratingP𝑃Pitalic_Pagainst some densityπP′⁢(𝒟)subscriptsuperscript𝜋′𝑃𝒟\\pi^{\\prime}_{P}(\\mathcal{D})italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ), such thatd⁢πP⁢(𝒟)=πP′⁢(𝒟)⁢d⁢Pdsubscript𝜋𝑃𝒟subscriptsuperscript𝜋′𝑃𝒟d𝑃\\mathrm{d}\\pi_{P}(\\mathcal{D})=\\pi^{\\prime}_{P}(\\mathcal{D})\\,\\mathrm{d}Proman_d italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ) = italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ) roman_d italic_P.πP′⁢(𝒟)subscriptsuperscript𝜋′𝑃𝒟\\pi^{\\prime}_{P}(\\mathcal{D})italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D )is the Radon-Nikodym derivative of the posterior stochastic process with respect to the prior; intuitively,πP′⁢(𝒟)⁢(f)subscriptsuperscript𝜋′𝑃𝒟𝑓\\pi^{\\prime}_{P}(\\mathcal{D})(f)italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ) ( italic_f )is proportional to the likelihood,πP′⁢(𝒟)⁢(f)=p⁢(𝒟∣f)/p⁢(𝒟)subscriptsuperscript𝜋′𝑃𝒟𝑓𝑝conditional𝒟𝑓𝑝𝒟\\pi^{\\prime}_{P}(\\mathcal{D})(f)=p(\\mathcal{D}\\mid f)/p(\\mathcal{D})italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ) ( italic_f ) = italic_p ( caligraphic_D ∣ italic_f ) / italic_p ( caligraphic_D ), so it determines how the data is observed (e.g. under which noise?). Assume thatπP′⁢(∅)∝1proportional-tosubscriptsuperscript𝜋′𝑃1\\pi^{\\prime}_{P}(\\varnothing)\\propto 1italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( ∅ ) ∝ 1so thatπP⁢(∅)=Psubscript𝜋𝑃𝑃\\pi_{P}(\\varnothing)=Pitalic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( ∅ ) = italic_P. We say thatπP′subscriptsuperscript𝜋′𝑃\\pi^{\\prime}_{P}italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTisG𝐺Gitalic_G-invariant ifπP′⁢(g⁢𝒟)∘g=πP′⁢(𝒟)subscriptsuperscript𝜋′𝑃𝑔𝒟𝑔subscriptsuperscript𝜋′𝑃𝒟\\pi^{\\prime}_{P}(g\\mathcal{D})\\circ g=\\pi^{\\prime}_{P}(\\mathcal{D})italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( italic_g caligraphic_D ) ∘ italic_g = italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPT ( caligraphic_D ).Proposition 1(G𝐺Gitalic_G-stationarity andG𝐺Gitalic_G-equivariance).The ground-truth stochastic processP𝑃Pitalic_PisG𝐺Gitalic_G-stationary andπP′subscriptsuperscript𝜋′𝑃\\pi^{\\prime}_{P}italic_π start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTisG𝐺Gitalic_G-invariant if, and only if,πPsubscript𝜋𝑃\\pi_{P}italic_π start_POSTSUBSCRIPT italic_P end_POSTSUBSCRIPTisG𝐺Gitalic_G-equivariant.Proof.This is Thm 2.1 byAshman et al. (2024)with translations replaced by applications ofg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G.\n∎Thus, for data generated from a stochastic process that is approximatelyG𝐺Gitalic_G-stationary, incorporatingG𝐺Gitalic_G-equivariance into NP approximations of the corresponding prediction map can serve as a useful inductive bias that can help generalisation and improve parameter efficiency. Intuitively, requiring a NP to be equivariant effectively ties parameters together, which significantly reduces the search space during optimisation, enabling better solutions to be found with fewer data. The improved generalisation capabilities ofG𝐺Gitalic_G-equivariant models is formalised byElesedy and Zaidi (2021); Bulusu et al. (2021); Petrache and Trivedi (2023).For a CNP, assume that every marginalp(⋅|𝐱,𝒟)p(\\,\\cdot\\,|\\mathbf{x},\\mathcal{D})italic_p ( ⋅ | bold_x , caligraphic_D )is in some fixed parametric family with parametersθ⁢(𝐱;𝒟)∈Θ𝜃𝐱𝒟Θ\\theta(\\mathbf{x};\\mathcal{D})\\in\\Thetaitalic_θ ( bold_x ; caligraphic_D ) ∈ roman_Θ. For example,θ𝜃\\thetaitalic_θcould consist of the mean and variance for a Gaussian distribution. LetC⁢(𝒳,Θ)𝐶𝒳ΘC(\\mathcal{X},\\Theta)italic_C ( caligraphic_X , roman_Θ )denote the set of continuous functions𝒳→Θ→𝒳Θ\\mathcal{X}\\rightarrow\\Thetacaligraphic_X → roman_Θ. For a CNP, define the associatedparameter mapΦ:𝒮→C⁢(𝒳,Θ):Φ→𝒮𝐶𝒳Θ\\Phi\\colon\\mathcal{S}\\to C(\\mathcal{X},\\Theta)roman_Φ : caligraphic_S → italic_C ( caligraphic_X , roman_Θ )byΦ⁢(𝒟)⁢(𝐱)=θ⁢(𝐱;𝒟)Φ𝒟𝐱𝜃𝐱𝒟\\Phi(\\mathcal{D})(\\mathbf{x})=\\theta(\\mathbf{x};\\mathcal{D})roman_Φ ( caligraphic_D ) ( bold_x ) = italic_θ ( bold_x ; caligraphic_D ). Intuitively, the parameter mapΦΦ\\Phiroman_Φmaps a dataset𝒟𝒟\\mathcal{D}caligraphic_Dto a functionΦ⁢(𝒟)Φ𝒟\\Phi(\\mathcal{D})roman_Φ ( caligraphic_D )giving the parametersΦ⁢(𝒟)⁢(𝐱)Φ𝒟𝐱\\Phi(\\mathcal{D})(\\mathbf{x})roman_Φ ( caligraphic_D ) ( bold_x )for every input𝐱𝐱\\mathbf{x}bold_x. Assume that𝒳𝒳\\mathcal{X}caligraphic_Xis aG𝐺Gitalic_G-space. We turn𝒮𝒮\\mathcal{S}caligraphic_SandC⁢(𝒳,Θ)𝐶𝒳ΘC(\\mathcal{X},\\Theta)italic_C ( caligraphic_X , roman_Θ )intoG𝐺Gitalic_G-spaces by applyingg𝑔gitalic_gto the inputs:g⁢𝒟∈𝒮𝑔𝒟𝒮g\\mathcal{D}\\in\\mathcal{S}italic_g caligraphic_D ∈ caligraphic_Sconsists of the input–output pairs(g⁢𝐱n,𝐲n)𝑔subscript𝐱𝑛subscript𝐲𝑛(g\\mathbf{x}_{n},\\mathbf{y}_{n})( italic_g bold_x start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , bold_y start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT ), andg⁢θ⁢(𝐱;𝒟)∈C⁢(𝒳,Θ)𝑔𝜃𝐱𝒟𝐶𝒳Θg\\theta(\\mathbf{x};\\mathcal{D})\\in C(\\mathcal{X},\\Theta)italic_g italic_θ ( bold_x ; caligraphic_D ) ∈ italic_C ( caligraphic_X , roman_Θ )is defined byg⁢θ⁢(𝐱;𝒟)=θ⁢(g−1⁢𝐱;𝒟)𝑔𝜃𝐱𝒟𝜃superscript𝑔1𝐱𝒟g\\theta(\\mathbf{x};\\mathcal{D})=\\theta(g^{-1}\\mathbf{x};\\mathcal{D})italic_g italic_θ ( bold_x ; caligraphic_D ) = italic_θ ( italic_g start_POSTSUPERSCRIPT - 1 end_POSTSUPERSCRIPT bold_x ; caligraphic_D ).Definition 3(G𝐺Gitalic_G-equivariant CNP).A CNP isG𝐺Gitalic_G-equivariant if the associated parameter mapΦΦ\\Phiroman_ΦisG𝐺Gitalic_G-equivariant:Φ⁢(g⁢𝒟)=g⁢Φ⁢(𝒟)Φ𝑔𝒟𝑔Φ𝒟\\Phi(g\\mathcal{D})=g\\Phi(\\mathcal{D})roman_Φ ( italic_g caligraphic_D ) = italic_g roman_Φ ( caligraphic_D )for all datasets𝒟𝒟\\mathcal{D}caligraphic_D.To parametrise aG𝐺Gitalic_G-equivariant CNP, we must parametrise the associated parameter mapΦΦ\\Phiroman_Φ.\nHere we present a general construction byKawano et al. (2021).Theorem 1(Representation ofG𝐺Gitalic_G-equivariant CNPs, Theorem 2 byKawano et al. (2021)).Let𝒴⊆ℝD𝒴superscriptℝ𝐷\\mathcal{Y}\\subseteq\\mathbb{R}^{D}caligraphic_Y ⊆ blackboard_R start_POSTSUPERSCRIPT italic_D end_POSTSUPERSCRIPTbe compact. Consider an appropriate collection𝒮≤M′⊆𝒮≤Msubscriptsuperscript𝒮′absent𝑀subscript𝒮absent𝑀\\mathcal{S}^{\\prime}_{\\leq M}\\subseteq\\mathcal{S}_{\\leq M}caligraphic_S start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT ≤ italic_M end_POSTSUBSCRIPT ⊆ caligraphic_S start_POSTSUBSCRIPT ≤ italic_M end_POSTSUBSCRIPT. Then a functionΦ:𝒮≤M′→C⁢(𝒳,Θ):Φ→subscriptsuperscript𝒮′absent𝑀𝐶𝒳Θ\\Phi\\colon\\mathcal{S}^{\\prime}_{\\leq M}\\rightarrow C(\\mathcal{X},\\Theta)roman_Φ : caligraphic_S start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT start_POSTSUBSCRIPT ≤ italic_M end_POSTSUBSCRIPT → italic_C ( caligraphic_X , roman_Θ )is continuous, permutation-invariant, andG𝐺Gitalic_G-equivariant if and only if it is of the formΦ⁢(𝒟)=ρ⁢(fenc⁢(𝒟))wherefenc⁢((𝐱1,𝐲1),…,(𝐱m,𝐲m))=∑i=1mϕ⁢(𝐲i)⁢ψ⁢(⋅,𝐱i)formulae-sequenceΦ𝒟𝜌subscript𝑓enc𝒟wheresubscript𝑓encsubscript𝐱1subscript𝐲1…subscript𝐱𝑚subscript𝐲𝑚superscriptsubscript𝑖1𝑚italic-ϕsubscript𝐲𝑖𝜓⋅subscript𝐱𝑖\\textstyle\\Phi(\\mathcal{D})=\\rho(f_{\\text{enc}}(\\mathcal{D}))\\quad\\text{where}%\n\\quad f_{\\text{enc}}((\\mathbf{x}_{1},\\mathbf{y}_{1}),\\ldots,(\\mathbf{x}_{m},%\n\\mathbf{y}_{m}))=\\sum_{i=1}^{m}\\phi(\\mathbf{y}_{i})\\psi(\\,\\cdot\\,,\\mathbf{x}_{%\ni})roman_Φ ( caligraphic_D ) = italic_ρ ( italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( caligraphic_D ) ) where italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( ( bold_x start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT , bold_y start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT ) , … , ( bold_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT , bold_y start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) ) = ∑ start_POSTSUBSCRIPT italic_i = 1 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_m end_POSTSUPERSCRIPT italic_ϕ ( bold_y start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT ) italic_ψ ( ⋅ , bold_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT )(2)for some continuousϕ:𝒴→ℝ2⁢D:italic-ϕ→𝒴superscriptℝ2𝐷\\phi\\colon\\mathcal{Y}\\rightarrow\\mathbb{R}^{2D}italic_ϕ : caligraphic_Y → blackboard_R start_POSTSUPERSCRIPT 2 italic_D end_POSTSUPERSCRIPT,\nan appropriateG𝐺Gitalic_G-invariant positive-definite kernelψ:𝒳2→ℝ:𝜓→superscript𝒳2ℝ\\psi\\colon\\mathcal{X}^{2}\\rightarrow\\mathbb{R}italic_ψ : caligraphic_X start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT → blackboard_R(i.e.ψ⁢(𝐱n,𝐱m)=ψ⁢(g⁢𝐱n,g⁢𝐱m)𝜓subscript𝐱𝑛subscript𝐱𝑚𝜓𝑔subscript𝐱𝑛𝑔subscript𝐱𝑚\\psi(\\mathbf{x}_{n},\\mathbf{x}_{m})=\\psi(g\\mathbf{x}_{n},g\\mathbf{x}_{m})italic_ψ ( bold_x start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , bold_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT ) = italic_ψ ( italic_g bold_x start_POSTSUBSCRIPT italic_n end_POSTSUBSCRIPT , italic_g bold_x start_POSTSUBSCRIPT italic_m end_POSTSUBSCRIPT )for allg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G),\nand some continuous andG𝐺Gitalic_G-equivariantρ:ℍ→C⁢(𝒳,Θ):𝜌→ℍ𝐶𝒳Θ\\rho\\colon\\mathbb{H}\\rightarrow C(\\mathcal{X},\\Theta)italic_ρ : blackboard_H → italic_C ( caligraphic_X , roman_Θ )withℍℍ\\mathbb{H}blackboard_Han appropriateG𝐺Gitalic_G-invariant space of functions (i.e. closed underg∈G𝑔𝐺g\\in Gitalic_g ∈ italic_G).Theorem1naturally gives rise to architectures which can be deconstructed into two components: anencoderand adecoder. The encoder,fenc:𝒮→𝒵:subscript𝑓enc→𝒮𝒵f_{\\text{enc}}\\colon\\mathcal{S}\\rightarrow\\mathcal{Z}italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT : caligraphic_S → caligraphic_Z, maps datasets to some embedding space𝒵𝒵\\mathcal{Z}caligraphic_Z. The decoder,ρ:𝒵→C⁢(𝒳,Θ):𝜌→𝒵𝐶𝒳Θ\\rho\\colon\\mathcal{Z}\\rightarrow C(\\mathcal{X},\\Theta)italic_ρ : caligraphic_Z → italic_C ( caligraphic_X , roman_Θ ), takes this representation and maps to a function that gives the parameters of the CNP’s predictive distributions:p⁢(𝐲|𝐱,𝒟)=p⁢(𝐲|ρ⁢(fenc⁢(𝒟))⁢(𝐱))𝑝conditional𝐲𝐱𝒟𝑝conditional𝐲𝜌subscript𝑓enc𝒟𝐱p(\\mathbf{y}|\\mathbf{x},\\mathcal{D})=p(\\mathbf{y}|\\rho(f_{\\text{enc}}(\\mathcal%\n{D}))(\\mathbf{x}))italic_p ( bold_y | bold_x , caligraphic_D ) = italic_p ( bold_y | italic_ρ ( italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( caligraphic_D ) ) ( bold_x ) )whereρ⁢(fenc⁢(𝒟))⁢(𝐱)∈Θ𝜌subscript𝑓enc𝒟𝐱Θ\\rho(f_{\\text{enc}}(\\mathcal{D}))(\\mathbf{x})\\in\\Thetaitalic_ρ ( italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( caligraphic_D ) ) ( bold_x ) ∈ roman_Θ.\nFor simplicity, we often view the decoder as a function𝒵×𝒳→Θ→𝒵𝒳Θ\\mathcal{Z}\\times\\mathcal{X}\\rightarrow\\Thetacaligraphic_Z × caligraphic_X → roman_Θand more simply writeρ⁢(fenc⁢(𝒟),𝐱)𝜌subscript𝑓enc𝒟𝐱\\rho(f_{\\text{enc}}(\\mathcal{D}),\\mathbf{x})italic_ρ ( italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( caligraphic_D ) , bold_x ).\nInTheorem1, both the encoderfencsubscript𝑓encf_{\\text{enc}}italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPTand decoderρ𝜌\\rhoitalic_ρareG𝐺Gitalic_G-equivariant.\nMany neural process architectures are of this form,\nincluding the ConvCNP(Gordon et al.,2019), the EquivCNP(Kawano et al.,2021), the RCNP(Huang et al.,2023), and the TE-TNP(Ashman et al.,2024).111We note that the form of the embedded datasetfenc⁢(𝒟)subscript𝑓enc𝒟f_{\\text{enc}}(\\mathcal{D})italic_f start_POSTSUBSCRIPT enc end_POSTSUBSCRIPT ( caligraphic_D )differs slightly in the RCP and TE-TNP. However, in both cases the form inEquation2can be recovered as special cases. We discuss this more inAppendixB.We discuss the specifics of each of these architectures inAppendixB.\n\n",
    "origin_citations": [
      "Conditional neural processes.",
      "Neural processes.",
      "Meta-learning stationary stochastic process prediction with convolutional neural processes.",
      "Convolutional Conditional Neural Processes.",
      "Translation-equivariant transformer neural processes.",
      "Provably strict generalisation benefit for equivariant models.",
      "Generalization capabilities of translationally equivariant neural networks.",
      "Approximation-generalization trade-offs under (approximate) group equivariance.",
      "Group equivariant conditional neural processes.",
      "Convolutional conditional neural processes.",
      "Practical equivariances via relational conditional neural processes."
    ],
    "citations": [
      5945,
      7224,
      202311,
      69526,
      776016,
      473890,
      506763,
      257463,
      474553,
      69526,
      49911
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 78,
      "Final_Score": 82.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 78\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 84,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 78,
      "Final_Score": 85.4,
      "raw_text": "Content_Coherence: 84\nCitation_Quality: 92\nSynthesis_Analysis: 78\nFinal_Score: 85"
    },
    "overall_score": 84.4,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "dqT9MC5NQl",
        "forum": "dqT9MC5NQl",
        "number": 6859,
        "title": "Approximately Equivariant Neural Processes",
        "normalized_title": "approximately equivariant neural processes",
        "abstract": "Equivariant deep learning architectures exploit symmetries in learning problems to improve the sample efficiency of neural-network-based models and their ability to generalise. However, when modelling real-world data, learning problems are often not *exactly* equivariant, but only approximately. For example, when estimating the global temperature field from weather station observations, local topographical features like mountains break translation equivariance. In these scenarios, it is desirable to construct architectures that can flexibly depart from exact equivariance in a data-driven way. Current approaches to achieving this cannot usually be applied out-of-the-box to any architecture and symmetry group. In this paper, we develop a general approach to achieving this using existing equivariant architectures. Our approach is agnostic to both the choice of symmetry group and model architecture, making it widely applicable. We consider the use of approximately equivariant architectures in neural processes (NPs), a popular family of meta-learning models. We demonstrate the effectiveness of our approach on a number of synthetic and real-world regression experiments, showing that approximately equivariant NP models can outperform both their non-equivariant and strictly equivariant counterparts.",
        "authors": [
          "Matthew Ashman",
          "Cristiana Diaconu",
          "Adrian Weller",
          "Wessel P Bruinsma",
          "Richard E. Turner"
        ],
        "decision": null,
        "pdate": 1727287827429,
        "cdate": 1715606371445
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "dqT9MC5NQl",
          "forum": "dqT9MC5NQl",
          "number": 6859,
          "title": "Approximately Equivariant Neural Processes",
          "normalized_title": "approximately equivariant neural processes",
          "abstract": "Equivariant deep learning architectures exploit symmetries in learning problems to improve the sample efficiency of neural-network-based models and their ability to generalise. However, when modelling real-world data, learning problems are often not *exactly* equivariant, but only approximately. For example, when estimating the global temperature field from weather station observations, local topographical features like mountains break translation equivariance. In these scenarios, it is desirable to construct architectures that can flexibly depart from exact equivariance in a data-driven way. Current approaches to achieving this cannot usually be applied out-of-the-box to any architecture and symmetry group. In this paper, we develop a general approach to achieving this using existing equivariant architectures. Our approach is agnostic to both the choice of symmetry group and model architecture, making it widely applicable. We consider the use of approximately equivariant architectures in neural processes (NPs), a popular family of meta-learning models. We demonstrate the effectiveness of our approach on a number of synthetic and real-world regression experiments, showing that approximately equivariant NP models can outperform both their non-equivariant and strictly equivariant counterparts.",
          "authors": [
            "Matthew Ashman",
            "Cristiana Diaconu",
            "Adrian Weller",
            "Wessel P Bruinsma",
            "Richard E. Turner"
          ],
          "decision": null,
          "pdate": 1727287827429,
          "cdate": 1715606371445
        }
      ]
    }
  },
  {
    "paper_id": 31838,
    "title": "LoRA-GA: Low-Rank Adaptation with Gradient Approximation",
    "abstract": "Fine-tuning large-scale pretrained models is prohibitively expensive in terms of computational and memory costs. LoRA, as one of the most popular Parameter-Efficient Fine-Tuning (PEFT) methods, offers a cost-effective alternative by fine-tuning an auxiliary low-rank model that has significantly fewer parameters. Although LoRA reduces the computational and memory requirements significantly at each iteration, extensive empirical evidence indicates that it converges at a considerably slower rate compared to full fine-tuning, ultimately leading to increased overall compute and often worse test performance. In our paper, we perform an in-depth investigation of the initialization method of LoRA and show that careful initialization (without any change of the architecture and the training algorithm) can significantly enhance both efficiency and performance. In particular, we introduce a novel initialization method, LoRA-GA (LowRankAdaptation withGradientApproximation), which aligns the gradients of low-rank matrix product with those of full fine-tuning at the first step. Our extensive experiments demonstrate that LoRA-GA achieves a convergence rate comparable to that of full fine-tuning (hence being significantly faster than vanilla LoRA as well as various recent improvements) while simultaneously attaining comparable or even better performance. For example, on the subset of the GLUE dataset with T5-Base, LoRA-GA outperforms LoRA by 5.69% on average. On larger models such as Llama 2-7B, LoRA-GA shows performance improvements of 0.34, 11.52%, and 5.05% on MT-bench, GSM8K, and Human-eval, respectively. Additionally, we observe up to 2-4 times convergence speed improvement compared to vanilla LoRA, validating its effectiveness in accelerating convergence and enhancing model performance. Code is available atgithub.",
    "introduction": "Fine-tuning large language models (LLMs) is essential for enabling advanced techniques such as instruction fine-tuning[1], reinforcement learning from human feedback (RLHF)[2], and adapting models to specific downstream applications. However, the computational and storage costs associated with full fine-tuning are prohibitively high, particularly as model sizes continue to grow.\nTo address these challenges, methods of Parameter-Efficient Fine-Tuning (PEFT) (see e.g.,[3]), such as Low-Rank Adaptation (LoRA)[4], have emerged and gained significant attention.\n\n\nInstead of updating the parameters of the model directly, LoRA incorporates auxilary low-rank matricesB𝐵Bitalic_BandA𝐴Aitalic_Ainto the linear layers of models (such as theQ,K,V𝑄𝐾𝑉Q,K,Vitalic_Q , italic_K , italic_V, andO𝑂Oitalic_Omatrices in a self-attention block[5]), while keeping the original layer weightsW𝑊Witalic_Wfixed. The modified layer is represented asy=(W+η⁢B⁢A)⁢x𝑦𝑊𝜂𝐵𝐴𝑥y=(W+\\eta BA)xitalic_y = ( italic_W + italic_η italic_B italic_A ) italic_x, wherex𝑥xitalic_xis the input of that layer,y𝑦yitalic_yis the output, andη𝜂\\etaitalic_ηis the scaling factor. This approach significantly reduces the number of parameters that need to be fine-tuned, thereby lowering the computational and memory costs at each step.\n\n\nDespite these benefits, extensive empirical evidence\n(see e.g.,[6,7,8,9]) shows that LoRA converges significantly slower compared to full finetune. This slower convergence often increases overall computational costs (measured in Floating Point Operations) and can sometimes lead to worse test performance. In our experiments, we typically observe that LoRA requires 5-6x more iterations and FLOPs to reach the same performance as full fine-tuning under the same learning rate, as shown in Figure1.\n\n\nFigure 1:(Left) Training loss curves of Llama 2-7B on MetaMathQA to training steps. LoRA-GA converges as quickly as full fine-tuning and outperforms LoRA. (Right) Initialization procedures used in LoRA and LoRA-GA. The key difference is that LoRA-GA initializes adapters using the eigenvectors of the gradient matrix, as opposed to random initialization with a scaling factor.\n\n\nTo study the cause of slow convergence, we perform an in-depth investigation of the initialization strategy of LoRA’s adapter weights. It is known that fine-tuning pretrained models using the same objective (e.g., language modeling) often converges faster than re-initializing new parameters (e.g., a classification head)[10]. This observation leads us to question whether the slow convergence of vanilla LoRA might be attributed to the default random initialization of adapter weights (LoRA initializesA𝐴Aitalic_Ausing Kaiming initialization[11]and setsB𝐵Bitalic_Bto zero[4]). In our experiments, we find that different initialization strategies for LoRA can significantly impact the results, and its default initialization is suboptimal.\n\n\nIn pursuit of a convergence rate comparable to full fine-tuning, we aim for initialization so that the update ofB⁢A𝐵𝐴BAitalic_B italic_Amatches the update ofW𝑊Witalic_Wclosely. Previous work suggests that gradient descent operates in a low-dimensional subspace[12,13]. If we can closely approximate the gradients of the full model at the initial step, subsequent steps can also be approximated, potentially accelerating the convergence of LoRA.\n\n\nTo this end, we introduce a novel initialization method, LoRA-GA (LowRankGradientApproximation). By initializingAinitsubscript𝐴initA_{\\text{init}}italic_A start_POSTSUBSCRIPT init end_POSTSUBSCRIPTandBinitsubscript𝐵initB_{\\text{init}}italic_B start_POSTSUBSCRIPT init end_POSTSUBSCRIPTwith the eigenvectors of the full gradient matrix, the gradient of the low-rank productB⁢A𝐵𝐴BAitalic_B italic_Aaligns with the direction of the gradient of the full weight matrixW𝑊Witalic_W. Mathematically, we aim to ensure that:Δ⁢(B⁢A)≈ζ⁢Δ⁢W,for some non-zero positive constant⁢ζ.Δ𝐵𝐴𝜁Δ𝑊for some non-zero positive constant𝜁\\Delta(BA)\\approx\\zeta\\Delta W,\\quad\\text{for some non-zero positive constant}%\n\\ \\zeta.roman_Δ ( italic_B italic_A ) ≈ italic_ζ roman_Δ italic_W , for some non-zero positive constant italic_ζ .\n\n\nOur contributions can be summarized as follows:\n\n\n1.We propose LoRA-GA , a novel initialization method for LoRA that accelerates convergence by approximating the gradients of the low-rank matrices with ones of the full weight matrix.\n\n\n2.We identify the scaling factor under non-zero initialization, which ensures the variance of adapter outputs is invariant to the rank of the adapter and the dimension of the input.\n\n\n3.We validate LoRA-GA through extensive experiments, demonstrating significant performance improvements and faster convergence compared to vanilla LoRA. Specifically, LoRA-GA outperforms LoRA by 5.69% on the GLUE[14]subset with T5-Base[15], and by 0.34, 11.52%, and 5.05% on MT-bench[16], GSM8K[17], and HumanEval[18]with Llama 2-7B[19], respectively, while achieving up to 2-4 times faster convergence.\n\n",
    "related_work": "2.1InitializationThe significance of maintaining variance stability during initialization has been widely acknowledged to prevent the occurrence of diminishing or exploding phenomena.\nXavier initialization[20]ensures stability in both the forward and backward passes of a network under a linear activation function. He initialization[11]extends this solution to networks using ReLU activation.\nDistinct from these, LSUV initialization[21]selects a mini-batch of data, performing a forward pass to determine the output variance, and subsequently normalizing it to ensure stability.\nTensor program (see e.g.,[22]) has emerged as a powerful\ntechnique for tuning various hyperparameters, including the initialization, for large models.\n\n\n2.2Parameter-Efficient Fine-Tuning (PEFT)To fine-tune increasingly large language models within the constraints of limited hardware resources, researchers have developed various Parameter-Efficient Fine-Tuning (PEFT) methods. One approach is Adapter-based methods[23,24,25,26], which incorporate new layers into existing layers of a model. By fine-tuning only these inserted layers (typically with much few parameters), resource consumption is significantly reduced. However, this approach introduces additional latency during both the forward and backward passes, as the computation must traverse the newly added layers. Another approach is Soft Prompt-based methods[10,27,28,29,30], which prepend learnable soft tokens (prompts) to the model’s input to adapt the model to specific tasks. This approach leverages the pre-trained model’s inherent capabilities, needing only appropriate prompts to adapt to downstream tasks. Despite its effectiveness, this method also incurs additional computational overhead\nand hence latency during inference.\n\n\n2.3LoRA’s VariantsLoRA is one of the most popular PEFT methods that introduces the product of low-rank matrices alongside existing layers to approximate weight changes during fine-tuning. Several methods have been proposed to improve the structure of LoRA. AdaLoRA[31]dynamically prunes insignificant weights during fine-tuning using SVD, allowing more rank allocation to important areas within a fixed parameter budget. DoRA[8]enhances the model’s expressiveness by adding learnable magnitudes to the direction adjustments made by low-rank matrix products. Additionally, LoHA[32]and LoKr[33]employ Hamiltonian and Kronecker products, respectively.Despite these advancements, vanilla LoRA remains the most popular method due to its robust library and hardware support. Therefore, improving LoRA without altering its structure and at a low cost is crucial. Several recent methods focus on this aspect. ReLoRA[34]suggests periodically merging learned adapters into the weight matrices to enhance LoRA’s expressibility. LoRA+[35]proposes using different learning rates for the two matrices in LoRA to improve convergence. rsLoRA[36]introduces a new scaling factor to make the scale of the output invariant to rank. Although our stable scale approach appears similar to rsLoRA, rsLoRA assumesB⁢A=0𝐵𝐴0BA=0italic_B italic_A = 0initialization, makingr𝑟ritalic_rinvariant to the updateΔ⁢B⁢AΔ𝐵𝐴\\Delta BAroman_Δ italic_B italic_A. In contrast, our stable scale ensures that non-zero initializedB⁢A𝐵𝐴BAitalic_B italic_Aremains invariant to both rank and input dimension from the start.Recently, PiSSA[37]proposes to\ninitializingA𝐴Aitalic_AandB𝐵Bitalic_Bto approximate the original matrixW𝑊Witalic_W, by\nperforming SVD onW𝑊Witalic_W. Our method, however, is based on a very different idea, that is to approximate the gradient ofW𝑊Witalic_W, which involves performing SVD on sampled gradients and properly scaling the initialized matrices, as detailed in SectionE.\n\n",
    "origin_citations": [
      "Understanding the difficulty of training deep feedforward neural networks.",
      "Delving deep into rectifiers: Surpassing human-level performance on imagenet classification, 2015.",
      "All you need is a good init.",
      "Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer.",
      "Parameter-efficient transfer learning for nlp.",
      "Towards a unified view of parameter-efficient transfer learning, 2022.",
      "Adamix: Mixture-of-adaptations for parameter-efficient model tuning.",
      "Adapterfusion: Non-destructive task composition for transfer learning.",
      "Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.",
      "The power of scale for parameter-efficient prompt tuning, 2021.",
      "Residual prompt tuning: Improving prompt tuning with residual reparameterization, 2023.",
      "Gpt understands, too.",
      "Prefix-tuning: Optimizing continuous prompts for generation.",
      "Adalora: Adaptive budget allocation for parameter-efficient fine-tuning, 2023.",
      "Dora: Weight-decomposed low-rank adaptation, 2024.",
      "Fedpara: Low-rank hadamard product for communication-efficient federated learning.",
      "Krona: Parameter efficient tuning with kronecker adapter.",
      "Relora: High-rank training through low-rank updates, 2023.",
      "Lora+: Efficient low rank adaptation of large models, 2024.",
      "A rank stabilization scaling factor for fine-tuning with lora, 2023.",
      "Pissa: Principal singular values and singular vectors adaptation of large language models, 2024."
    ],
    "citations": [
      3205,
      407869,
      518119,
      7567,
      342105,
      146767,
      419165,
      408717,
      8536,
      116307,
      7789,
      1591,
      495508,
      97191,
      94151,
      139643,
      185540,
      466147,
      22829,
      448949
    ],
    "extract_rate": 0.9523809523809523,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 76,
      "Citation_Quality": 89,
      "Synthesis_Analysis": 83,
      "Final_Score": 83.30000000000001,
      "raw_text": "Content_Coherence: 76\nCitation_Quality: 89\nSynthesis_Analysis: 83\nFinal_Score: 83"
    },
    "overall_score": 84.34,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "VaLAWrLHJv",
        "forum": "VaLAWrLHJv",
        "number": 16871,
        "title": "LoRA-GA: Low-Rank Adaptation with Gradient Approximation",
        "normalized_title": "lora ga low rank adaptation with gradient approximation",
        "abstract": "Fine-tuning large-scale pretrained models is prohibitively expensive in terms of computational and memory costs. LoRA, as one of the most popular Parameter-Efficient Fine-Tuning (PEFT) methods, offers a cost-effective alternative by fine-tuning an auxiliary low-rank model that has significantly fewer parameters. Although LoRA reduces the computational and memory requirements significantly at each iteration, extensive empirical evidence indicates that it converges at a considerably slower rate compared to full fine-tuning, ultimately leading to increased overall compute and often worse test performance. In our paper, we perform an in-depth investigation of the initialization method of LoRA and show that careful initialization (without any change of the architecture and the training algorithm) can significantly enhance both efficiency and performance. In particular, we introduce a novel initialization method, LoRA-GA (Low Rank Adaptation with Gradient Approximation), which aligns the gradients of low-rank matrix product with those of full fine-tuning at the first step. Our extensive experiments demonstrate that LoRA-GA achieves a convergence rate comparable to that of full fine-tuning (hence being significantly faster than vanilla LoRA as well as various recent improvements) while simultaneously attaining comparable or even better performance. For example, on the subset of the GLUE dataset with T5-Base, LoRA-GA outperforms LoRA by 5.69% on average. On larger models such as Llama 2-7B, LoRA-GA shows performance improvements of 0.34, 11.52%, and 5.05% on MTbench, GSM8k, and Human-eval, respectively. Additionally, we observe up to 2-4 times convergence speed improvement compared to vanilla LoRA, validating its effectiveness in accelerating convergence and enhancing model performance.",
        "authors": [
          "Shaowen Wang",
          "Linxi Yu",
          "Jian Li"
        ],
        "decision": null,
        "pdate": 1727288139386,
        "cdate": 1715775111558
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "VaLAWrLHJv",
          "forum": "VaLAWrLHJv",
          "number": 16871,
          "title": "LoRA-GA: Low-Rank Adaptation with Gradient Approximation",
          "normalized_title": "lora ga low rank adaptation with gradient approximation",
          "abstract": "Fine-tuning large-scale pretrained models is prohibitively expensive in terms of computational and memory costs. LoRA, as one of the most popular Parameter-Efficient Fine-Tuning (PEFT) methods, offers a cost-effective alternative by fine-tuning an auxiliary low-rank model that has significantly fewer parameters. Although LoRA reduces the computational and memory requirements significantly at each iteration, extensive empirical evidence indicates that it converges at a considerably slower rate compared to full fine-tuning, ultimately leading to increased overall compute and often worse test performance. In our paper, we perform an in-depth investigation of the initialization method of LoRA and show that careful initialization (without any change of the architecture and the training algorithm) can significantly enhance both efficiency and performance. In particular, we introduce a novel initialization method, LoRA-GA (Low Rank Adaptation with Gradient Approximation), which aligns the gradients of low-rank matrix product with those of full fine-tuning at the first step. Our extensive experiments demonstrate that LoRA-GA achieves a convergence rate comparable to that of full fine-tuning (hence being significantly faster than vanilla LoRA as well as various recent improvements) while simultaneously attaining comparable or even better performance. For example, on the subset of the GLUE dataset with T5-Base, LoRA-GA outperforms LoRA by 5.69% on average. On larger models such as Llama 2-7B, LoRA-GA shows performance improvements of 0.34, 11.52%, and 5.05% on MTbench, GSM8k, and Human-eval, respectively. Additionally, we observe up to 2-4 times convergence speed improvement compared to vanilla LoRA, validating its effectiveness in accelerating convergence and enhancing model performance.",
          "authors": [
            "Shaowen Wang",
            "Linxi Yu",
            "Jian Li"
          ],
          "decision": null,
          "pdate": 1727288139386,
          "cdate": 1715775111558
        }
      ]
    }
  },
  {
    "paper_id": 34299,
    "title": "Diffusion Feedback Helps CLIP See Better",
    "abstract": "Contrastive Language-Image Pre-training (CLIP), which excels at abstracting open-world representations across domains and modalities, has become a foundation for a variety of vision and multimodal tasks.\nHowever, recent studies reveal that CLIP has severe visual shortcomings, such as which can hardly distinguish orientation, quantity, color, structure,etc.\nThese visual shortcomings also limit the perception capabilities of multimodal large language models (MLLMs) built on CLIP.\nThe main reason could be that the image-text pairs used to train CLIP are inherently biased, due to the lack of the distinctiveness of the text and the diversity of images.\nIn this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process.\nWe introduceDIVA, which uses theDIffusion model as aVisualAssistant for CLIP.\nSpecifically,DIVAleverages generative feedback from text-to-image diffusion models to optimize CLIP representations, with only images (without corresponding text).\nWe demonstrate thatDIVAimproves CLIP’s performance on the challenging MMVP-VLM benchmark which assesses fine-grained visual abilities to a large extent (e.g., 3-7% ↑), and enhances the performance of MLLMs and vision models on multimodal understanding and segmentation tasks.\nExtensive evaluation on 29 image classification and retrieval benchmarks confirms that our framework preserves CLIP’s strong zero-shot capabilities.\nThe code will be available athttps://github.com/baaivision/DIVA.",
    "introduction": "Contrastive language-image pre-training (CLIP)(Radford et al.,2021b)has been widely applied to various multimodal understanding and generation tasks, including open-domain image classification(Sun et al.,2024d; Zhang et al.,2022; Zhu et al.,2023), text-to-image retrieval(Luo et al.,2023; Baldrati et al.,2022; Sain et al.,2023),\nvisual grounding(Wang et al.,2022; Yu et al.,2023; Wang et al.,2024b;a),\nand text-to-image generation(Frans et al.,2022; Bar-Tal et al.,2022; Rombach et al.,2022a; Crowson et al.,2022; Ramesh et al.,2022; Vinker et al.,2022).\nThis widespread application is due to CLIP’s excellent visual representation ability, learned from large-scale data. Thus, enhancing CLIP’s representation and capabilities is crucial for advancing downstream tasks.\n\n\nSince the introduction of CLIP(Radford et al.,2021b), numerous subsequent studies on CLIP models have emerged in recent years.\nThese studies have utilized training techniques pre-train(Sun et al.,2023;2024b; Fang et al.,2023; Xu et al.,2023a; Zhai et al.,2023; Shi et al.,2024)and fine-tune(Wei et al.,2023; Zhang et al.,2024)CLIP models, achieving improved performance and unlocking new abilities. However, these approaches still suffer from unavoidable limitations, as they heavily rely on image-text data pairs and cannot work on image-only data.\n\n\nFigure 1:Left:The existing CLIP models mostly suffer from the inability to distinguish visual details. After enhancing the visual capabilities with ourDIVA, the sensitivity of CLIP to visual details has greatly improved.Right:Our proposedDIVAconsistently boosts the performance of various CLIP models(Radford et al.,2021b; Fang et al.,2023; Xu et al.,2023a; Zhai et al.,2023)on MMVP-VLM benchmark that evaluates the visual capabilities of vision-language models.\n\n\nAs noted by recent works(Kim et al.,2023; Zeng et al.,2021; Zhang et al.,2024; Tong et al.,2024b;a), despite its excellent zero-shot performance, CLIP suffers from certain perceptual understanding limitations due to the contrastive learning paradigm and the noisy image-text pairs used in training. These limitations include an inability to accurately comprehend long texts and to perceive fine-grained differences in similar images. While some studies have attempted to address the text comprehension issue(Zhang et al.,2024), research on improving CLIP’s fine-grained visual perception remains underexplored. The ability to perceive visual detail is crucial for foundation models, and the lack of this capability in CLIP directly affects the performance of vision and multimodal models that use CLIP as a vision encoder(Tong et al.,2024b;a).\n\n\nThus, in this work, we focus on addressing CLIP’s inability to distinguish fine-grained visual details via self-supervised learning (SSL) paradigm.\nBased on text-to-image diffusion models generating realistic images with much details, we explore leveraging generative feedback from diffusion models to optimize CLIP representations.\nBy conditioning diffusion models with CLIP’s densely recapped visual features and applying reconstruction loss to CLIP optimization,\nwe leverage theDIffusion model as aVisualAssistant for CLIP, thus the name of our approach,DIVA.\nOur results highlight thatDIVAgreatly enhances CLIP’s performance on MMVP-VLM benchmark measuring visual abilities of V-L models, and improves multimodal large language models (MLLMs) and vision models on multimodal and vision understanding tasks.\nBesides,DIVAmaintains CLIP’s excellent zero-shot performance on 29 image classification and retrieval benchmarks.\n\n\nOur main contributions can be summarized as follows:•Concentrating on overcoming CLIP’s visual shortcomings in perceiving fine-grained details, we present the first work to exploit the potential of leveraging generative feedback from text-to-image diffusion models to optimize CLIP model’s discriminative representations.•We propose a simple self-supervised frameworkDIVAfor CLIP’s representation optimization. Coupled with our visual dense recap scheme,DIVAconditions diffusion models with dense visual features from CLIP and incorporates image reconstruction loss for optimization.•OurDIVAgreatly boosts CLIP’s visual perception capability and improves its performance on MMVP-VLM benchmark, further enhancing MLLMs and vision models on multimodal and visual understanding tasks. Meanwhile, our results on 29 image classification and retrieval benchmarks show thatDIVAmaintains CLIP’s original excellent zero-shot performance.\n\n",
    "related_work": "CLIP Models & MLLMs.The introduction of CLIP(Radford et al.,2021b)has significantly advanced multimodal learning. Since its debut, a series of CLIP models have emerged(Sun et al.,2023; Fang et al.,2023; Xu et al.,2023a; Zhai et al.,2023), enhancing performance and unlocking new capabilities through improved pre-training techniques and model architectures.\nOn this basis, CLIP has been widely adopted as a foundation model, serving as a backbone for various applications such as image segmentation(Li et al.,2022a; Xu et al.,2022; Shan et al.,2024; Xu et al.,2023c; Liang et al.,2023; Zhou et al.,2023), object detection(Gu et al.,2021; Li et al.,2022b; Subramanian et al.,2022)and video understanding(Bose et al.,2023; Lin et al.,2022; Castro & Heilbron,2022; Xu et al.,2021; Rasheed et al.,2023; Tang et al.,2024). Its ability to align language and vision has led to superior results on these tasks compared to traditional methods.\nMoreover, CLIP has driven the development of MLLMs(Liu et al.,2024b;a; Sun et al.,2024c;a). Combining strong visual understanding with advanced large language models facilitates more sophisticated interactions between vision and language. Recent works have highlighted inherent visual flaws in the CLIP models and MLLMs using CLIP as the visual encoder(Tong et al.,2024b;a). To address this, some research has incorporated multiple vision encoders to achieve more precise and comprehensive visual perception(Kar et al.,2024; Jiang et al.,2023; Tong et al.,2024b). However, this approach increases computational costs and memory usage. There has been no research directly enhancing CLIP’s visual perception capabilities to better serve MLLMs. Thus, the main focus of our work is to fundamentally overcome CLIP’s visual perception shortcomings, directly benefiting both vision models and multimodal MLLMs that use CLIP as a backbone.\n\n\nDiffusion Models for Representation Learning.Diffusion models(Ho et al.,2020; Song et al.,2020)have made remarkable progress in various generative tasks, such as image generation(Rombach et al.,2022b; Saharia et al.,2022; Betker et al.,2023; Zheng et al.,2024), video generation(Singer et al.,2022; Blattmann et al.,2023; Junhao Zhang et al.,2023; Ho et al.,2022), editing(Meng et al.,2021; Hu et al.,2024; Mou et al.,2023), etc.\nApart from research above, there are also many works focus on employing diffusion models for representation learning.\nSome of works leverage the intermediate activation of pre-trained diffusion models for different downstream tasks, including classification(Xiang et al.,2023), semantic segmentation(Baranchuk et al.,2021), panoptic segmentation(Xu et al.,2023b), depth estimation(Zhao et al.,2023), etc.\nOther works(Hudson et al.,2024; Pan et al.,2023)train their own diffusion models coupled with meticulously devised modules to further boosting the representation capabilities.\nBesides, Diffusion-TTA(Prabhudesai et al.,2023)aims to adapt pre-trained vision encoders to samples in testing set using feedback from a diffusion model.\nAdditionally, some methods(Guo et al.,2024; Trabucco et al.,2023; Tian et al.,2024; Azizi et al.,2023)utilize diffusion models to generate synthetic data, which is then adopted to enhance the representation capabilities of corresponding models.\nIn contrast, in our work, we mainly focus on exploring the potential of enhancing the original discriminative representations of CLIP models by directly utilizing generative feedback from the diffusion models.\nAdditionally, we aim to leverage the diffusion models to break free from the constraints of paired image-text data and construct a self-supervised framework to improve CLIP’s visual perception capabilities.\n\n",
    "origin_citations": [
      "Learning transferable visual models from natural language supervision.",
      "EVA-CLIP: Improved training techniques for clip at scale.",
      "Data filtering networks.",
      "Demystifying CLIP data.",
      "Sigmoid loss for language image pre-training.",
      "Language-driven semantic segmentation.",
      "Groupvit: Semantic segmentation emerges from text supervision.",
      "Open-vocabulary semantic segmentation with image embedding balancing.",
      "Side adapter network for open-vocabulary semantic segmentation.",
      "Open-vocabulary semantic segmentation with mask-adapted clip.",
      "Zegclip: Towards adapting clip for zero-shot semantic segmentation.",
      "Open-vocabulary object detection via vision and language knowledge distillation.",
      "Grounded language-image pre-training.",
      "Reclip: A strong zero-shot baseline for referring expression comprehension.",
      "Movieclip: Visual scene recognition in movies.",
      "Frozen clip models are efficient video learners.",
      "Fitclip: Refining large-scale pretrained image-text models for zero-shot video understanding tasks.",
      "Videoclip: Contrastive pre-training for zero-shot video-text understanding.",
      "Fine-tuned clip models are efficient video learners.",
      "Learnable feature augmentation framework for temporal action localization.",
      "Visual instruction tuning.",
      "Improved baselines with visual instruction tuning.",
      "Emu: Generative pretraining in multimodality.",
      "Generative multimodal models are in-context learners.",
      "Eyes wide shut? exploring the visual shortcomings of multimodal llms.",
      "Cambrian-1: A fully open, vision-centric exploration of multimodal llms.",
      "Brave: Broadening the visual encoding of vision-language models.",
      "From clip to dino: Visual encoders shout in multi-modal large language models.",
      "Denoising diffusion probabilistic models.",
      "Score-based generative modeling through stochastic differential equations.",
      "High-resolution image synthesis with latent diffusion models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Improving image generation with better captions.",
      "Cogview3: Finer and faster text-to-image generation via relay diffusion.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Stable video diffusion: Scaling latent video diffusion models to large datasets.",
      "Show-1: Marrying pixel and latent diffusion models for text-to-video generation.",
      "Imagen video: High definition video generation with diffusion models.",
      "Sdedit: Guided image synthesis and editing with stochastic differential equations.",
      "Instruct-imagen: Image generation with multi-modal instruction.",
      "Dragondiffusion: Enabling drag-style manipulation on diffusion models.",
      "Denoising diffusion autoencoders are unified self-supervised learners.",
      "Label-efficient semantic segmentation with diffusion models.",
      "Open-vocabulary panoptic segmentation with text-to-image diffusion models.",
      "Unleashing text-to-image diffusion models for visual perception.",
      "Soda: Bottleneck diffusion models for representation learning.",
      "Masked diffusion as self-supervised representation learner.",
      "Diffusion-tta: Test-time adaptation of discriminative models via generative feedback.",
      "Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic-domain alignment.",
      "Effective data augmentation with diffusion models.",
      "Stablerep: Synthetic images from text-to-image models make strong visual representation learners.",
      "Synthetic data from diffusion models improves imagenet classification."
    ],
    "citations": [
      1639,
      493141,
      477297,
      477663,
      493162,
      255086,
      174081,
      918656,
      226304,
      460116,
      222490,
      5325,
      7121,
      63679,
      335961,
      235219,
      500882,
      7916,
      222665,
      2238,
      322710,
      810942,
      348078,
      2455,
      704915,
      446752,
      320196,
      4827,
      431228,
      2240,
      5926,
      289951,
      176197,
      23455,
      478059,
      461035,
      96204,
      485566,
      186981,
      495891,
      152182,
      77161,
      78386,
      22100,
      180210,
      22932,
      698014,
      490432,
      308045,
      54550
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 84,
      "Final_Score": 82.6,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 85\nSynthesis_Analysis: 84\nFinal_Score: 83"
    },
    "overall_score": 84.16,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "tLFWU6izoA",
        "forum": "tLFWU6izoA",
        "number": 918,
        "title": "Diffusion Feedback Helps CLIP See Better",
        "normalized_title": "diffusion feedback helps clip see better",
        "abstract": "Contrastive Language-Image Pre-training (CLIP), which excels at abstracting open-world representations across domains and modalities, has become a foundation for a variety of vision and multimodal tasks. However, recent studies reveal that CLIP has severe visual shortcomings, such as which can hardly distinguish orientation, quantity, color, structure, etc. These visual shortcomings also limit the perception capabilities of multimodal large language models (MLLMs) built on CLIP. The main reason could be that the image-text pairs used to train CLIP are inherently biased, due to the lack of the distinctiveness of the text and the diversity of images. In this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process. We introduce DIVA, which uses the DIffusion model as a Visual Assistant for CLIP. Specifically, DIVA leverages generative feedback from text-to-image diffusion models to optimize CLIP representations, with only images (without corresponding text). We demonstrate that DIVA improves CLIP's performance on the challenging MMVP-VLM benchmark which assesses fine-grained visual abilities to a large extent (e.g., 3-7%), and enhances the performance of MLLMs and vision models on multimodal understanding and segmentation tasks. Extensive evaluation on 29 image classification and retrieval benchmarks confirms that our framework preserves CLIP's strong zero-shot capabilities. The code is publicly available at https://github.com/baaivision/DIVA.",
        "authors": [
          "Wenxuan Wang",
          "Quan Sun",
          "Fan Zhang",
          "Yepeng Tang",
          "Jing Liu",
          "Xinlong Wang"
        ],
        "decision": null,
        "pdate": 1737562212128,
        "cdate": 1726405586193
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "tLFWU6izoA",
          "forum": "tLFWU6izoA",
          "number": 918,
          "title": "Diffusion Feedback Helps CLIP See Better",
          "normalized_title": "diffusion feedback helps clip see better",
          "abstract": "Contrastive Language-Image Pre-training (CLIP), which excels at abstracting open-world representations across domains and modalities, has become a foundation for a variety of vision and multimodal tasks. However, recent studies reveal that CLIP has severe visual shortcomings, such as which can hardly distinguish orientation, quantity, color, structure, etc. These visual shortcomings also limit the perception capabilities of multimodal large language models (MLLMs) built on CLIP. The main reason could be that the image-text pairs used to train CLIP are inherently biased, due to the lack of the distinctiveness of the text and the diversity of images. In this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process. We introduce DIVA, which uses the DIffusion model as a Visual Assistant for CLIP. Specifically, DIVA leverages generative feedback from text-to-image diffusion models to optimize CLIP representations, with only images (without corresponding text). We demonstrate that DIVA improves CLIP's performance on the challenging MMVP-VLM benchmark which assesses fine-grained visual abilities to a large extent (e.g., 3-7%), and enhances the performance of MLLMs and vision models on multimodal understanding and segmentation tasks. Extensive evaluation on 29 image classification and retrieval benchmarks confirms that our framework preserves CLIP's strong zero-shot capabilities. The code is publicly available at https://github.com/baaivision/DIVA.",
          "authors": [
            "Wenxuan Wang",
            "Quan Sun",
            "Fan Zhang",
            "Yepeng Tang",
            "Jing Liu",
            "Xinlong Wang"
          ],
          "decision": null,
          "pdate": 1737562212128,
          "cdate": 1726405586193
        }
      ]
    }
  },
  {
    "paper_id": 19403,
    "title": "Graph Neural Networks for Learning Equivariant Representations of Neural Networks",
    "abstract": "Neural networks that process the parameters of other neural networks find applications in domains as diverse as classifying implicit neural representations, generating neural network weights, and predicting generalization errors.\nHowever, existing approaches either overlook the inherent permutation symmetry in the neural network or rely on intricate weight-sharing patterns to achieve equivariance, while ignoring the impact of the network architecture itself.\nIn this work, we propose to represent neural networks as computational graphs of parameters, which allows us to harness powerful graph neural networks and transformers that preserve permutation symmetry.\nConsequently, our approach enables a single model to learn from neural graphs with diverse architectures.\nWe showcase the effectiveness of our method on a wide range of tasks, including classification and editing of implicit neural representations, predicting generalization performance, and learning to optimize, while consistently outperforming state-of-the-art methods.\nThe source code is open-sourced athttps://github.com/mkofinas/neural-graphs.",
    "introduction": "How can we design neural networks that themselves takeneural network parametersas input?\nThis would allow us to make inferencesaboutneural networks, such as predicting their generalization error(Unterthiner et al.,2020), generating neural network weights(Schürholt et al.,2022a), and classifying or generating implicit neural representations(Dupont et al.,2022)without having to evaluate them on many different inputs.\nFor simplicity, let us consider a deep neural network with multiple hidden layers.\nAs a naïve approach, we can simply concatenate all flattened weights and biases into one large feature vector, from which we can then make predictions as usual.\nHowever, this overlooks an important structure in the parameters: neurons in a layer can bereorderedwhile maintaining exactly the same function(Hecht-Nielsen,1990). Reordering neurons of a neural network means permuting the preceding and following weight matrices accordingly.\nIgnoring the permutation symmetry will typically cause this model to make different predictions for different orderings of the neurons in the input neural network, even though they represent exactly the same function.\n\n\nIn general, accounting for symmetries in the input data improves the learning efficiency and underpins the field of geometric deep learning(Bronstein et al.,2021).\nRecent studies(Navon et al.,2023; Zhou et al.,2023a)confirm the effectiveness of equivariant layers forparameter spaces(the space of neural network parameters) with specially designed weight-sharing patterns.\nThese weight-sharing patterns, however, require manual adaptation to each new architectural design.\nImportantly, a single model can only process neural network parameters for a single fixed architecture.\nMotivated by these observations, we take an alternative approach to address the permutation symmetry in neural networks: we present theneural graphrepresentation that connects neural network parameters similar to the computation graph (seeFigure1).\nBy explicitly integrating the graph structure in our neural network, a single model can processheterogeneousarchitectures,i.e.architectures with varying computational graphs,\nincluding architectures with different number of layers, number of hidden dimensions, non-linearities, and different network connectivities such as residual connections.\n\n\nWe make the following contributions.\nFirst, we propose a simple and efficient representation of neural networks asneural graphs, which ensures invariance to neuron symmetries.\nThe perspective of permuting neurons rather than parameters makes our model conceptually much simpler than prior work.\nWe detail various choices on how the neural graph can encode a neural network, including the novel concept of “probe features” that represent the neuron activations of a forward pass.\nSecond, we adapt existing graph neural networks and transformers to take neural graphs as input,\nand incorporate inductive biases from neural graphs.\nIn the context of geometric deep learning, neural graphs constitute a new benchmark for graph neural networks.\nFinally, we empirically validate our proposed method on a wide range of tasks, and outperform state-of-the-art approaches by a large margin.\n\n\n\\readlist\\Nnod\\foreachitem\\N∈\\Nnod\\N\\Nnod\\prevlayerl−1𝑙1l-1italic_l - 1layerl𝑙litalic_lNeural network feedforward activation(neuroni𝑖\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}iitalic_iin layerl𝑙litalic_l)𝐱i(l)=σ⁢(𝐛i(l)+∑j𝐖i⁢j(l)⁢𝐱j(l−1))superscriptsubscript𝐱𝑖𝑙𝜎superscriptsubscript𝐛𝑖𝑙subscript𝑗superscriptsubscript𝐖𝑖𝑗𝑙superscriptsubscript𝐱𝑗𝑙1{\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}\\mathbf{%\nx}_{i}^{(l)}}=\\sigma\\mathopen{}\\mathclose{{}\\left({\\color[rgb]{1,0.5,0}%\n\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}\\mathbf{b}_{i}^{(l)}}+%\n\\displaystyle\\sum_{\\color[rgb]{0,0.6,0}\\definecolor[named]{pgfstrokecolor}{rgb%\n}{0,0.6,0}j}{\\color[rgb]{0,0.3984375,0.80078125}\\definecolor[named]{%\npgfstrokecolor}{rgb}{0,0.3984375,0.80078125}\\mathbf{W}}_{{\\color[rgb]{1,0.5,0}%\n\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}i}{\\color[rgb]{0,0.6,0}%\n\\definecolor[named]{pgfstrokecolor}{rgb}{0,0.6,0}j}}^{(l)}{\\color[rgb]{0,0.6,0%\n}\\definecolor[named]{pgfstrokecolor}{rgb}{0,0.6,0}\\mathbf{x}_{j}^{(l-1)}}}\\right)bold_x start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT = italic_σ ( bold_b start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT + ∑ start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT bold_W start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT bold_x start_POSTSUBSCRIPT italic_j end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l - 1 ) end_POSTSUPERSCRIPT )Neural network asneural graph:Nodei𝑖{\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}i}italic_ifeature:𝑽i(l)←𝐛i(l)←superscriptsubscript𝑽𝑖𝑙superscriptsubscript𝐛𝑖𝑙\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}\\bm{V}_{i%\n}^{(l)}\\leftarrow\\mathbf{b}_{i}^{(l)}bold_italic_V start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT ← bold_b start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPTEdgej→i→𝑗𝑖{\\color[rgb]{0,0.6,0}\\definecolor[named]{pgfstrokecolor}{rgb}{0,0.6,0}j}\\to{%\n\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{1,0.5,0}i}italic_j → italic_ifeature:𝑬i⁢j(l)←𝐖i⁢j(l)←superscriptsubscript𝑬𝑖𝑗𝑙superscriptsubscript𝐖𝑖𝑗𝑙{\\color[rgb]{0,0.3984375,0.80078125}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n0,0.3984375,0.80078125}\\bm{E}}_{{\\color[rgb]{1,0.5,0}\\definecolor[named]{%\npgfstrokecolor}{rgb}{1,0.5,0}i}{\\color[rgb]{0,0.6,0}\\definecolor[named]{%\npgfstrokecolor}{rgb}{0,0.6,0}j}}^{(l)}\\leftarrow{\\color[rgb]{%\n0,0.3984375,0.80078125}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n0,0.3984375,0.80078125}\\mathbf{W}}_{{\\color[rgb]{1,0.5,0}\\definecolor[named]{%\npgfstrokecolor}{rgb}{1,0.5,0}i}{\\color[rgb]{0,0.6,0}\\definecolor[named]{%\npgfstrokecolor}{rgb}{0,0.6,0}j}}^{(l)}bold_italic_E start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPT ← bold_W start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ( italic_l ) end_POSTSUPERSCRIPTFigure 1:Representing a neural network and its weights as aneural graph.\nWe assign neural network parameters to graph features by treatingbiases𝐛isubscript𝐛𝑖{\\mathbf{b}}_{i}bold_b start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPTas correspondingnode features𝑽isubscript𝑽𝑖{\\bm{V}}_{i}bold_italic_V start_POSTSUBSCRIPT italic_i end_POSTSUBSCRIPT, andweights𝐖i⁢jsubscript𝐖𝑖𝑗{\\mathbf{W}}_{{\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n1,0.5,0}i}{\\color[rgb]{0,0.6,0}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n0,0.6,0}j}}bold_W start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPTasedge features𝑬i⁢jsubscript𝑬𝑖𝑗{\\bm{E}}_{{\\color[rgb]{1,0.5,0}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n1,0.5,0}i}{\\color[rgb]{0,0.6,0}\\definecolor[named]{pgfstrokecolor}{rgb}{%\n0,0.6,0}j}}bold_italic_E start_POSTSUBSCRIPT italic_i italic_j end_POSTSUBSCRIPTconnecting the nodes in adjacent layers.\n\n",
    "related_work": "Networks for networks.A recent line of work studies how to learn representations for trained classifiers(Baker et al.,2017; Eilertsen et al.,2020; Unterthiner et al.,2020; Schürholt et al.,2021;2022b;2022a)to predict their generalization performance, or other properties that provide insight into neural networks.Dupont et al. (2022); De Luigi et al. (2023)learn low-dimensional encodings for INRs for generation or other downstream tasks.\nThese works either flatten the network parameters or compute parameter statistics and process them with a standard MLP.\nUsing statistics respects the symmetries of the neural network, but discards fine-grained details that can be important for the task.\nOn the other hand, architectures that flatten the network parameters are not equivariant with respect to the symmetries of the neural network.\nThis is undesirable because models that are functionally identical – obtained by permuting the weights and biases appropriately – can receive vastly different predictions.Schürholt et al. (2021)proposed neuron permutation augmentations to address the symmetry problem, however such an approach can require vastly more training compute due to the size of the permutation group:∏ldl!subscriptproduct𝑙subscript𝑑𝑙\\prod_{l}d_{l}!∏ start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT italic_d start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT !.\nOther related works(Peebles et al.,2022; Ashkenazi et al.,2023; Knyazev et al.,2021; Erkoç et al.,2023)encode and/or decode neural network parameters mainly for reconstruction and generation purposes.\nHowever, similarly to the aforementioned works, these do not address the symmetry problem in a principled way.\n\n\nThe symmetry problem.Three recent studies address these shortcomings and propose equivariant linear layers that achieve equivariance through intricate weight-sharing patterns(Navon et al.,2023; Zhou et al.,2023a;b).\nThey compose these layers to create both invariant and equivariant networks for networks.\nSo far, their performance on INR classification shows that there is still a gap when compared to image classification using CNNs.\nSimilar to these studies, our proposed method maintains the symmetry of the neural network.\nIn contrast to these works, by integrating the graph structure in our neural network, we are no longer limited to homogeneous architectures and can processheterogeneousarchitectures for a much wider gamut of applications.\n\n",
    "origin_citations": [
      "Accelerating Neural Architecture Search using Performance Prediction.",
      "Classifying the classifier: dissecting the weight space of neural networks.",
      "Predicting Neural Network Accuracy from Weights.",
      "Self-Supervised Representation Learning on Neural Network Weights for Model Characteristic Prediction.",
      "Model Zoos: A Dataset of Diverse Populations of Neural Network Models.",
      "Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights.",
      "From data to functa: Your data point is a function and you can treat it like one.",
      "Deep Learning on Implicit Neural Representations of Shapes.",
      "Learning to Learn with Generative Models of Neural Network Checkpoints.",
      "NeRN – Learning Neural Representations for Neural Networks.",
      "Parameter Prediction for Unseen Deep Architectures.",
      "HyperDiffusion: Generating Implicit Neural Fields with Weight-Space Diffusion.",
      "Equivariant Architectures for Learning in Deep Weight Spaces.",
      "Permutation Equivariant Neural Functionals.",
      "Neural Functional Transformers."
    ],
    "citations": [
      15721,
      131063,
      164990,
      1088384,
      176211,
      176220,
      55838,
      489651,
      176990,
      138590,
      512593,
      492427,
      212232,
      225464,
      259864
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 79,
      "Citation_Quality": 83,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.4,
      "raw_text": "Content_Coherence: 79\nCitation_Quality: 83\nSynthesis_Analysis: 85\nFinal_Score: 83"
    },
    "overall_score": 84.04,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "oO6FsMyDBt",
        "forum": "oO6FsMyDBt",
        "number": 7510,
        "title": "Graph Neural Networks for Learning Equivariant Representations of Neural Networks",
        "normalized_title": "graph neural networks for learning equivariant representations of neural networks",
        "abstract": "Neural networks that process the parameters of other neural networks find applications in domains as diverse as classifying implicit neural representations, generating neural network weights, and predicting generalization errors. However, existing approaches either overlook the inherent permutation symmetry in the neural network or rely on intricate weight-sharing patterns to achieve equivariance, while ignoring the impact of the network architecture itself. In this work, we propose to represent neural networks as computational graphs of parameters, which allows us to harness powerful graph neural networks and transformers that preserve permutation symmetry. Consequently, our approach enables a single model to encode neural computational graphs with diverse architectures. We showcase the effectiveness of our method on a wide range of tasks, including classification and editing of implicit neural representations, predicting generalization performance, and learning to optimize, while consistently outperforming state-of-the-art methods. The source code is open-sourced at https://github.com/mkofinas/neural-graphs.",
        "authors": [
          "Miltiadis Kofinas",
          "Boris Knyazev",
          "Yan Zhang",
          "Yunlu Chen",
          "Gertjan J. Burghouts",
          "Efstratios Gavves",
          "Cees G. M. Snoek",
          "David W. Zhang"
        ],
        "decision": null,
        "pdate": 1705411011331,
        "cdate": 1695476894164
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "oO6FsMyDBt",
          "forum": "oO6FsMyDBt",
          "number": 7510,
          "title": "Graph Neural Networks for Learning Equivariant Representations of Neural Networks",
          "normalized_title": "graph neural networks for learning equivariant representations of neural networks",
          "abstract": "Neural networks that process the parameters of other neural networks find applications in domains as diverse as classifying implicit neural representations, generating neural network weights, and predicting generalization errors. However, existing approaches either overlook the inherent permutation symmetry in the neural network or rely on intricate weight-sharing patterns to achieve equivariance, while ignoring the impact of the network architecture itself. In this work, we propose to represent neural networks as computational graphs of parameters, which allows us to harness powerful graph neural networks and transformers that preserve permutation symmetry. Consequently, our approach enables a single model to encode neural computational graphs with diverse architectures. We showcase the effectiveness of our method on a wide range of tasks, including classification and editing of implicit neural representations, predicting generalization performance, and learning to optimize, while consistently outperforming state-of-the-art methods. The source code is open-sourced at https://github.com/mkofinas/neural-graphs.",
          "authors": [
            "Miltiadis Kofinas",
            "Boris Knyazev",
            "Yan Zhang",
            "Yunlu Chen",
            "Gertjan J. Burghouts",
            "Efstratios Gavves",
            "Cees G. M. Snoek",
            "David W. Zhang"
          ],
          "decision": null,
          "pdate": 1705411011331,
          "cdate": 1695476894164
        }
      ]
    }
  },
  {
    "paper_id": 10880,
    "title": "Improving Generalization of Alignment with Human Preferences through Group Invariant Learning",
    "abstract": "The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences.\nAs universal AI assistants, there’s a growing expectation for them to perform consistently across various domains.\nHowever, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples.\nThis focus on quick reward gains undermines both the stability in training and the model’s ability to generalize to new, unseen data.\nIn this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains.\nGiven the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance.\nThen, we optimize the policy to perform well on challenging groups.\nLastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.",
    "introduction": "In the rapidly evolving field of language models, Reinforcement Learning from Human Feedback (RLHF) has emerged as a critical component, with the goal of aligning model outputs with human intents(Ouyang et al.,2022; Bai et al.,2022b).\nThis process integrates reward modeling, where human annotators rank different model responses based on their preference, followed by the reinforcement learning (RL) stages to refine and optimize the model’s behavior.\nGiven its universal applicability, such a model is expected to grasp a wide range of human intents and handle diverse scenarios(Askell et al.,2021).\nIn this context, the ability for generalization – consistent performance in both seen and unseen situations – becomes of paramount importance.\n\n\nRLHF faces a critical challenge in effectively generalizing human intents, raising concerns about its true capability beyond supervised settings(Casper et al.,2023).\nThis distinction is crucial as it impacts the model’s reliability in unseen situations and worst-case scenarios.\nPrevious studies indicate that RL often tends to excessively focus on simple and high-reward data while neglecting the learning of challenging samples(Ngo,2022; di Langosco et al.,2022; Zheng et al.,2023b).\nFurthermore, when there are flaws in the reward model, this behavior can lead the model into a “reward hacking” dilemma(Skalse et al.,2022; Pan et al.,2022), resulting in meaningless text outputs. All of these factors contribute to the model’s inconsistent performance on data from different groups111We interchangeably use the terms “groups” and “domains”., with poor generalization capabilities(Casper et al.,2023; Song et al.,2023).\n\n\nIn the field of deep learning, handling training data from different groups is a common challenge(Levy et al.,2020). This challenge becomes even more prominent in RL because the state-action distribution continuously changes as the policy gets optimized(Bai et al.,2022a; Xu et al.,2023).\nThis means that algorithms must learn and adapt under the constantly shifting data distribution, greatly increasing the difficulty and instability of learning.\nMost existing RL methods primarily focus on maximizing the expected future rewards, but their robustness often falls short when dealing with data from different distribution sources(Tang et al.,2019; Zhang et al.,2021).\nThe reason is that these methods often overlook the differences between data groups and fail to effectively penalize rare but catastrophic events that might occur(Tang et al.,2019; Javed et al.,2021).\nTo prevent the model’s output from deviating too much from the expected reward model, many techniques introduce a Kullback-Leibler (KL) divergence penalty(Stiennon et al.,2020; Ouyang et al.,2022).\nHowever, the weakness of this approach is that its constraint strength is typically set by the most likely outlier data, limiting the algorithm’s ability to handle challenging data(Laidlaw et al.,2023).\n\n\nIn this paper, we propose an alignment method with strong generalization capabilities aimed at achieving consistent model performance across multiple data groups.\nUnlike existing approaches, our technique not only focuses on maximizing overall expected rewards but also on reducing variance among data groups.\nBy maximizing the performance differences between different data groups, our method automatically segregates the data into distinct groups without the need for manual annotation. Through this adversarial approach, our method significantly enhances model generalization and training stability.\nFurthermore, based on the performance of each group, our approach can adaptively adjust the strength of the KL penalty term, providing a larger exploration space for finding better policies to handle challenging data.\n\n\nOur main contributions are as follows:\n\n\n•We introduce a unified framework that ensures distributionally robust alignment by dynamically adapting to various data groups, enhancing the model’s ability to handle different data distributions.•Within this framework, we develop a method for inferring group labels. This method utilizes group labels to implement adaptive KL constraints, ensuring optimal model behavior across different data subsets, thereby contributing to the robustness and stability of our model.•We empirically demonstrate that our proposed method outperforms the traditional PPO algorithm in the context of a general AI assistant and summarization settings. It exhibits outstanding generalization capabilities, significantly improving stability and performance metrics, further establishing the practical utility of our approach.\n\n",
    "related_work": "Although LLMs have promising capabilities, they are prone to exhibiting unintended behaviors, such as fabricating facts, generating biased or toxic content, or even producing harmful material for humans(Bender et al.,2021; Bommasani et al.,2021).\nTherefore, it is essential to align LLMs with human intentions and societal values. For example, they should be helpful, honest, and harmless (3H)(Ouyang et al.,2022; Bai et al.,2022b; Thoppilan et al.,2022).\nRL offers the most direct approach to achieving this goal. In RL, agents require supervision signals from reward models acting as human proxies. They are then fine-tuned through numerous iterations within the RL framework, a process known as Reinforcement Learning from Human Feedback.\nSeveral recent attempts have been made in this direction(Zhang et al.,2023; Rafailov et al.,2023; Hu et al.,2023).\n\n\nIn RL, the policy model faces significant challenges related to its generalization ability(Casper et al.,2023).\nFirstly, policies may exhibit poor generalization, especially when there is misleading correlation between the true objective and other events(McKinney et al.,2023; Tien et al.,2023).\nFurthermore, RL agents tend to seek expedient solutions, which can lead them to avoid challenging data in order to obtain high rewards, similar to what is observed in question-answering models(Turner et al.,2021; Casper et al.,2023).\nLastly, optimizing agents for imperfect rewards can result in reward hacking, leading to the generation of outputs that, while yielding high rewards, are meaningless(Skalse et al.,2022; Pan et al.,2022).\nAll of these challenges can lead to poor performance in capturing real human intent, emphasizing the necessity of learning a policy that can perform consistently across different data domains.\n\n\nMany RL algorithms focus on improving the generalization ability of policies in different environments(Javed et al.,2021; Sonar et al.,2021)and worst-case scenarios(Tang et al.,2019; Brown et al.,2020).\nHowever, most of these methods rely heavily on Bayesian neural networks(Brown et al.,2020; Javed et al.,2021), and the formulation of the problem differs from that of LLMs(Sonar et al.,2021).\nOur approach is inspired by invariant learning(Arjovsky et al.,2019; Creager et al.,2021), aiming to enhance stability in unfamiliar domains during testing by learning to find invariant features across different data group, thereby learning a more robust policy.\nIn recent research, invariant learning has been extended to scenarios that do not require prior group labels(Creager et al.,2021; Liu et al.,2021; Chen et al.,2022).\nTypically, these methods first train a reference model to acquire loss information from different data and then train an additional classifier to maximize violations of the invariant learning objective for grouping(Creager et al.,2021). In contrast, our approach employs a unified framework that iteratively performs group label inference and invariant policy learning. To the best of our knowledge, this is the first attempt to introduce the concept of group invariant learning into RL.\n\n",
    "origin_citations": [
      "On the dangers of stochastic parrots: Can language models be too big?",
      "On the opportunities and risks of foundation models.",
      "Training language models to follow instructions with human feedback.",
      "Training a helpful and harmless assistant with reinforcement learning from human feedback.",
      "Lamda: Language models for dialog applications.",
      "The wisdom of hindsight makes language models better instruction followers.",
      "Direct preference optimization: Your language model is secretly a reward model.",
      "Aligning language models with offline reinforcement learning from human feedback.",
      "Open problems and fundamental limitations of reinforcement learning from human feedback.",
      "On the fragility of learned reward functions.",
      "Causal confusion and reward misidentification in preference-based reward learning.",
      "Optimal policies tend to seek power.",
      "Defining and characterizing reward hacking.",
      "The effects of reward misspecification: Mapping and mitigating misaligned models.",
      "Policy gradient bayesian robust optimization for imitation learning.",
      "Invariant policy optimization: Towards stronger generalization in reinforcement learning.",
      "Worst cases policy gradients.",
      "Bayesian robust optimization for imitation learning.",
      "Invariant risk minimization.",
      "Environment inference for invariant learning.",
      "Just train twice: Improving group robustness without training group information.",
      "When does group invariant learning survive spurious correlations?"
    ],
    "citations": [
      1550,
      364,
      2255,
      1553,
      489739,
      5930,
      1091085,
      5929,
      302835,
      63447,
      204491,
      176929,
      255087,
      193827,
      2740,
      450954,
      520949,
      2738,
      354561,
      2766,
      79484
    ],
    "extract_rate": 0.9545454545454546,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 75,
      "Final_Score": 84.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 75\nFinal_Score: 84"
    },
    "overall_score": 84.03999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "fwCoLe3TAX",
        "forum": "fwCoLe3TAX",
        "number": 8880,
        "title": "Improving Generalization of Alignment with Human Preferences through Group Invariant Learning",
        "normalized_title": "improving generalization of alignment with human preferences through group invariant learning",
        "abstract": "The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. \nAs universal AI assistants, there's a growing expectation for them to perform consistently across various domains. \nHowever, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples.\nThis focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data.\nIn this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. \nGiven the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance.\nThen, we optimize the policy to perform well on challenging groups. \nLastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.",
        "authors": [
          "Rui Zheng",
          "Wei Shen",
          "Yuan Hua",
          "Wenbin Lai",
          "Shihan Dou",
          "Yuhao Zhou",
          "Zhiheng Xi",
          "Xiao Wang",
          "Haoran Huang",
          "Tao Gui",
          "Qi Zhang",
          "Xuanjing Huang"
        ],
        "decision": null,
        "pdate": 1705411052744,
        "cdate": 1695528001020
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "fwCoLe3TAX",
          "forum": "fwCoLe3TAX",
          "number": 8880,
          "title": "Improving Generalization of Alignment with Human Preferences through Group Invariant Learning",
          "normalized_title": "improving generalization of alignment with human preferences through group invariant learning",
          "abstract": "The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. \nAs universal AI assistants, there's a growing expectation for them to perform consistently across various domains. \nHowever, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples.\nThis focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data.\nIn this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. \nGiven the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance.\nThen, we optimize the policy to perform well on challenging groups. \nLastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.",
          "authors": [
            "Rui Zheng",
            "Wei Shen",
            "Yuan Hua",
            "Wenbin Lai",
            "Shihan Dou",
            "Yuhao Zhou",
            "Zhiheng Xi",
            "Xiao Wang",
            "Haoran Huang",
            "Tao Gui",
            "Qi Zhang",
            "Xuanjing Huang"
          ],
          "decision": null,
          "pdate": 1705411052744,
          "cdate": 1695528001020
        }
      ]
    }
  },
  {
    "paper_id": 39679,
    "title": "From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities",
    "abstract": "Multimodal Large Language Models have made significant strides in integrating visual and textual information, yet they often struggle with effectively aligning these modalities. We introduce a novel image tokenizer that bridges this gap by applying the principle of Byte-Pair Encoding (BPE) to visual data. Unlike conventional approaches that rely on separate visual encoders, our method directly incorporates structural prior information into image tokens, mirroring the successful tokenization strategies used in text-only Large Language Models. This innovative approach enables Transformer models to more effectively learn and reason across modalities. Through theoretical analysis and extensive experiments, we demonstrate that our BPE Image Tokenizer significantly enhances MLLMs’ multimodal understanding capabilities, even with limited training data. Our method not only improves performance across various benchmarks but also shows promising scalability, potentially paving the way for more efficient and capable multimodal foundation models.",
    "introduction": "The development of Multimodal Large Language Models (MLLMs) has made significant progress(Yin et al.,2023; Team et al.,2023; Liu et al.,2024b). However, these multimodal foundation models often model different modalities separately, incorporating many modality-specific designs such as specialized encoders and decoders(Liu et al.,2024b; Zhang et al.,2024; Jin et al.,2023). While this approach allows training data to align well with these modality-specific designs, it often struggles to achieve a unified understanding of multimodal information(Team,2024). The primary reason for this limitation could be that while encoders of other modalities can learn rich information, without the assistance of the corresponding decoders, LLMs cannot fully comprehend the complex patterns contained within the embeddings provided by the encoder. In other words, LLMs need to learn to interpret the token embeddings again, which is the job of the decoders of other modalities, leading to difficulties in aligning with these modalities(Baltrušaitis et al.,2018).\n\n\nRecent research has begun to explore unified token-based representations for MLLMs(Team,2024; Lu et al.,2024; Zheng et al.,2024), attempting to achieve better capabilities for multimodal information processing by moving away from modality-specific designs. This approach offers two main advantages: first, it can achieve unified information understanding, and second, it enables unified generation that can be decoded into different modalities. However, these methods simply convert various modal information into unified representations and then directly feed them into Transformer models(Vaswani,2017), hoping to learn the data just with massive computing resources. Although this approach can be effective to some extent, it consumes enormous computing resources and time while failing to achieve significant improvements in performance.\n\n\nIn comparison to the widely adopted training paradigm of text-only LLMs(Touvron et al.,2023; Achiam et al.,2023), we observe that current token-based MLLMs often overlook a crucial component: the tokenizer that explicitly merges data sources. In text-only LLMs, the most widely adopted tokenizer is the Byte-Pair Encoding (BPE) tokenizer(Sennrich,2015; Radford et al.,2019), which learns to merge characters into tokens of varying lengths based on the word frequency in the training corpus before providing them to the Transformer for learning. While many believe this approach is solely for conserving computational resources, some recent theoretical and experimental analyses suggest that using merged tokens plays a vital role in the Transformer’s learning of sequential data(Rajaraman et al.,2024; Makkuva et al.,2024; Merrill et al.,2024). In fact, if raw sequence data is provided directly, Transformer models may even fail to learn how to predict entirely.\n\n\nIn this paper, we reveal that Transformer models cannot effectively learn certain types of two-dimensional sequence data. However, when we apply operations similar to BPE tokenizers to the two-dimensional data,i.e., merging tokens based on the frequency of training data, the learning effectiveness can improve significantly. We believe the benefit comes from the addition of necessary structural prior information to the tokens during the merging process, making it easier for Transformer models to be aware of important relational information in the data during learning. Our theoretical analysis proves that tokenizing sequence data on certain types of two-dimensional data can indeed achieve smaller losses, which we further validate experimentally.\n\n\nBased on the insights provided by our theoretical analysis and experimental validation, we propose a new learning paradigm for MLLMs. By tokenizing the unified representation of multimodal data using a novel BPE Image Tokenizer, we enable Transformer models to better understand image data, allowing MLLMs built from text-only LLMs to have good multimodal capabilities. As illustrated in Figure1, after quantizing the image into token IDs, the BPE Image Tokenizer further combines these token IDs according to its learned patterns. Similar to text tokenizers, our BPE Image Tokenizer ensures that the tokens being fed to the language model decoder inherently contain more semantically meaningful information in the statistical sense. Intuitively, the tokenized IDs, which have incorporated structural prior information from the image, could provide the Transformer model with better comprehension of the specific alignment between text and image in the training data. Preliminary results further validate the effectiveness of explicit tokenization for multimodal data. We hope that the new paradigm for learning MLLMs proposed in this paper will provide better guidance for subsequent scaling-up efforts, leading to the building of more powerful MLLMs.\n\n\nFigure 1:Illustration of our BPE Image Tokenizer. The overall process begins with quantizing the image into initial token IDs. The BPE Image Tokenizer then combines these tokens based on learned patterns, similar to text tokenizers. This combination results in tokens that inherently contain more semantic information. The final tokenized sequence thus incorporates structural prior information from the image, enabling the Transformer model to deeper comprehend the alignment between visual and textual information during training. This approach facilitates more effective integration of visual data into MLLMs, enhancing their multimodal understanding capabilities.\n\n\nIn summary, we make the following key contributions:\n\n\n•We are the first to propose a new MLLM learning paradigm that explicitly tokenizes multimodal data like text-only LLMs, centered around our novel BPE Image Tokenizer.•We theoretically analyze why this learning paradigm can bring benefits and further provide corresponding experimental validation.•We design an algorithm for training the BPE Image Tokenizer and train an MLLM with this tokenizer. The performance evaluation further validates the capability enhancements this learning paradigm brings to MLLMs.\n\n",
    "related_work": "Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in various tasks, including visual question answering, image captioning, and cross-modal retrieval(Yin et al.,2023). The evolution of MLLMs can be broadly categorized into two main approaches: late-fusion models with specialized encoders and early-fusion token-based models.\n\n\nTraditional late-fusion MLLMs typically employ distinct specialized designs for different modalities(Team,2024). In this approach, modality-specific modules are trained separately using specialized data, followed by additional alignment steps to achieve late-fusion of different modalities. For instance, in the image modality, CLIP-based visual encoders(Radford et al.,2021; Fang et al.,2023)are first pre-trained on extensive image-caption datasets, then aligned with text-only LLM backbones(Touvron et al.,2023; Peng et al.,2023)using additional data. Notable examples of this approach include Flamingo(Alayrac et al.,2022), LLaVA(Liu et al.,2024b;a), IDEFICS(Laurençon et al.,2024), and Emu(Sun et al.,2023). However, these models often face challenges in modality alignment, leading to issues such as hallucination in MLLMs(Bai et al.,2024).\n\n\nTo address these challenges, recent research has explored early-fusion approaches through unified representations, proposing token-based methods for multimodal learning(Team,2024; Lu et al.,2024; Zheng et al.,2024). These methods typically utilize Vector Quantization (VQ) models(Van Den Oord et al.,2017; Razavi et al.,2019; Esser et al.,2021)to convert images into discrete tokens. The concept of token-based multimodal learning was initially explored in studies such as BEiT(Bao et al.,2021), which introduced a self-supervised method for acquiring visual representations based on tokenized image patches. This idea was further developed in works like Cm3(Aghajanyan et al.,2022)and CM3Leon(Yu et al.,2023), which enabled joint reasoning across modalities and scaled up to autoregressive text-to-image generation. More recent models like Gemini(Team et al.,2023), Unicode(Zheng et al.,2024), and Chameleon(Team,2024)have adopted end-to-end token-based approaches for multimodal learning.\n\n\nWhile these token-based MLLMs demonstrate enhanced reasoning and generation capabilities across various modalities without requiring modality-specific components, they still face challenges in representation learning and alignment(Baltrušaitis et al.,2018). Our proposed BPE Image Tokenizer paradigm addresses these challenges by adhering more closely to the learning method of text-only LLMs. Unlike current token-based approaches, our method directly incorporates crucial structural prior information into the tokens through explicit merging. This approach enables Transformer models to learn input data more effectively, as supported by both our theoretical analysis and experimental results.\n\n\nThe key distinction of our work lies in its focus on optimizing the tokenization process itself, rather than relying solely on pre-trained visual encoders or simple quantization. While traditional visual encoders can achieve high compression rates, their encoded embeddings often depend on specialized decoders for interpretation. In contrast, our BPE Image Tokenizer creates tokens that are directly meaningful to the Transformer model, facilitating more effective learning of visual information without the need for specialized decoders. This approach bridges the gap between visual and textual modalities more seamlessly, potentially leading to more robust and versatile MLLMs.\n\n",
    "origin_citations": [
      "A survey on multimodal large language models.",
      "Chameleon: Mixed-modal early-fusion foundation models.",
      "Learning transferable visual models from natural language supervision.",
      "Eva: Exploring the limits of masked visual representation learning at scale.",
      "Llama: Open and efficient foundation language models.",
      "Instruction tuning with gpt-4.",
      "Flamingo: a visual language model for few-shot learning.",
      "Visual instruction tuning.",
      "Improved baselines with visual instruction tuning.",
      "Obelics: An open web-scale filtered dataset of interleaved image-text documents.",
      "Generative pretraining in multimodality.",
      "Hallucination of multimodal large language models: A survey.",
      "Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action.",
      "Unicode: Learning a unified codebook for multimodal large language models.",
      "Neural discrete representation learning.",
      "Generating diverse high-fidelity images with vq-vae-2.",
      "Taming transformers for high-resolution image synthesis.",
      "Beit: Bert pre-training of image transformers.",
      "Cm3: A causal masked multimodal model of the internet.",
      "Scaling autoregressive multi-modal models: Pretraining and instruction tuning.",
      "Gemini: a family of highly capable multimodal models.",
      "Multimodal machine learning: A survey and taxonomy."
    ],
    "citations": [
      2965,
      810882,
      1639,
      32005,
      1552,
      2216,
      7565,
      2238,
      322710,
      49276,
      185641,
      816131,
      346280,
      287895,
      3018,
      7999,
      4765,
      9149,
      57784,
      129685,
      8028,
      1030
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 90,
      "Synthesis_Analysis": 75,
      "Final_Score": 84.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 90\nSynthesis_Analysis: 75\nFinal_Score: 83"
    },
    "overall_score": 84.03999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "3TnLGGHhNx",
        "forum": "3TnLGGHhNx",
        "number": 3823,
        "title": "From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities",
        "normalized_title": "from pixels to tokens byte pair encoding on quantized visual modalities",
        "abstract": "Multimodal Large Language Models have made significant strides in integrating visual and textual information, yet they often struggle with effectively aligning these modalities. We introduce a novel image tokenizer that bridges this gap by applying the principle of Byte-Pair Encoding (BPE) to visual data. Unlike conventional approaches that rely on separate visual encoders, our method directly incorporates structural prior information into image tokens, mirroring the successful tokenization strategies used in text-only Large Language Models. This innovative approach enables Transformer models to more effectively learn and reason across modalities. Through theoretical analysis and extensive experiments, we demonstrate that our BPE Image Tokenizer significantly enhances MLLMs' multimodal understanding capabilities, even with limited training data. Leveraging this method, we develop Being-VL-0, a model that demonstrates superior performance across various benchmarks and shows promising scalability, potentially paving the way for more efficient and capable multimodal foundation models. For further details, visit our website https://github.com/BeingBeyond/Being-VL-0.",
        "authors": [
          "Wanpeng Zhang",
          "Zilong Xie",
          "Yicheng Feng",
          "Yijiang Li",
          "Xingrun Xing",
          "Sipeng Zheng",
          "Zongqing Lu"
        ],
        "decision": null,
        "pdate": 1737562333247,
        "cdate": 1727195047272
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "3TnLGGHhNx",
          "forum": "3TnLGGHhNx",
          "number": 3823,
          "title": "From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities",
          "normalized_title": "from pixels to tokens byte pair encoding on quantized visual modalities",
          "abstract": "Multimodal Large Language Models have made significant strides in integrating visual and textual information, yet they often struggle with effectively aligning these modalities. We introduce a novel image tokenizer that bridges this gap by applying the principle of Byte-Pair Encoding (BPE) to visual data. Unlike conventional approaches that rely on separate visual encoders, our method directly incorporates structural prior information into image tokens, mirroring the successful tokenization strategies used in text-only Large Language Models. This innovative approach enables Transformer models to more effectively learn and reason across modalities. Through theoretical analysis and extensive experiments, we demonstrate that our BPE Image Tokenizer significantly enhances MLLMs' multimodal understanding capabilities, even with limited training data. Leveraging this method, we develop Being-VL-0, a model that demonstrates superior performance across various benchmarks and shows promising scalability, potentially paving the way for more efficient and capable multimodal foundation models. For further details, visit our website https://github.com/BeingBeyond/Being-VL-0.",
          "authors": [
            "Wanpeng Zhang",
            "Zilong Xie",
            "Yicheng Feng",
            "Yijiang Li",
            "Xingrun Xing",
            "Sipeng Zheng",
            "Zongqing Lu"
          ],
          "decision": null,
          "pdate": 1737562333247,
          "cdate": 1727195047272
        }
      ]
    }
  },
  {
    "paper_id": 2097,
    "title": "Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach",
    "abstract": "Not all positive pairs are beneficial to time series contrastive learning.\nIn this paper, we study two types of bad positive pairs that can impair the quality of time series representation learned through contrastive learning: the noisy positive pair and the faulty positive pair. We observe that, with the presence of noisy positive pairs, the model tends to simply learn the pattern of noise (Noisy Alignment). Meanwhile, when faulty positive pairs arise, the model wastes considerable amount of effort aligning non-representative patterns (Faulty Alignment). To address this problem, we propose a Dynamic Bad Pair Mining (DBPM) algorithm, which reliably identifies and suppresses bad positive pairs in time series contrastive learning. Specifically, DBPM utilizes a memory module to dynamically track the training behavior of each positive pair along training process. This allows us to identify potential bad positive pairs at each epoch based on their historical training behaviors. The identified bad pairs are subsequently down-weighted through a transformation module, thereby mitigating their negative impact on the representation learning process. DBPM is a simple algorithm designed as a lightweightplug-inwithout learnable parameters to enhance the performance of existing state-of-the-art methods. Through extensive experiments conducted on four large-scale, real-world time series datasets, we demonstrate DBPM’s efficacy in mitigating the adverse effects of bad positive pairs.††Codes are available atGitHubFigure 1:Motivation: Two types of bad positive pair identified in real-world time series contrastive learning:Noisy Positive Pair: The presence of excessive noise in the original signal and the augmented view leads to the contrastive model predominantly learning patterns from noise (i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e ., Noisy Alignment).Faulty Positive Pair: The augmented view no longer has the same semantic meaning as the original signal due to the destruction of important temporal patterns during augmentation (the red part highlighted in the original signal), causing the contrastive model learns to align non-representative patterns (i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e ., Faulty Alignment).",
    "introduction": "Self-supervised contrastive learning has shown remarkable efficacy in learning meaningful representations from unlabeled time series data, thereby improving the performance of downstream tasks (e.g.formulae-sequence𝑒𝑔e.g.italic_e . italic_g ., time series classification)(Oord et al.,2018; Tonekaboni et al.,2021; Eldele et al.,2021; Yèche et al.,2021; Lan et al.,2022; Yue et al.,2022; Yang & Hong,2022; Zhang et al.,2022b; Ozyurt et al.,2023). The key concept of time series contrastive learning is to capture the meaningful underlying information shared between views (usually generated by random data augmentation), such that the learned representations are discriminative among time series. To accomplish this objective, existing contrastive methods comply with an assumption that the augmented views from the same instance (i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e ., positive pair) share meaningful semantic information(Chen et al.,2020; He et al.,2020a; Chen & He,2021a; Wen & Li,2021).\n\n\nBut what happens if this assumption is violated? Some recent studies(Morgado et al.,2021; Chuang et al.,2022)have investigated the effects of the faulty view problem in image contrastive learning. It was discovered that the effectiveness of the contrastive learning models will indeed suffer when two views generated from the same image actually do not share meaningful information (e.g.formulae-sequence𝑒𝑔e.g.italic_e . italic_g ., an unrelated cat-dog pair cropped from the same image). In other words, not all positive pairs are beneficial to contrastive learning. We extend this line of work to time series by investigating whether time series contrastive learning encounters similar challenges, and explore what kind of positive pairs are detrimental to the learning process. Specifically, when applying commonly used contrastive learning methods to time series applications, we observe two extreme cases of positive pairs that violate the assumption and consequently vitiate the quality of the learned time series representation:\n\n\nNoisy Positive Pairs: As shown in Figure1(a), a real-world example from the Sleep-EDF dataset(Goldberger et al.,2000)(EEG signal). These pairs can arise when the original signal exhibits substantial noise, which could be attributed to improper data collection in real-world scenarios. This results in noisy contrasting views, where the shared information between views is predominantly noise. In the given example, the low voltage EEG is interfered by other electrical signals such as the electrooculogram (EOG) from eye movement, which makes some subtle EEG spikes indicative of early stage epilepsy being overwhelmed by noise. As a result, a noisy alignment occurs, in which the model simply learns the pattern of noise rather than the true signal.\n\n\nFaulty Positive Pairs: As shown in Figure1(b), a real-world example from the PTB-XL dataset(Wagner et al.,2020)(ECG signal). Due to the sensitivity of time series, data augmentation may inadvertently impair sophisticated temporal patterns contained in the original signal and thus produce faulty views. In the example, the unique temporal pattern exhibited by ventricular premature beat (red part) for diagnosing ECG abnormality has been destroyed after augmentation. Consequently, the augmented view is no longer has the same semantic meaning as the original ECG, leading to a faulty alignment where the model learns to align non-representative patterns shared between views.\n\n\nFor clarity, we refer to these two extreme cases as bad positive pair problem in time series contrastive learning. An intuitive solution is to suppress bad positive pairs during contrastive training, however, directly identifying bad positive pairs in real-world datasets is challenging for two reasons. First, for noisy positive pairs, it is often infeasible to measure the noise level given a signal from a real-world dataset. Second, for faulty positive pairs, one will not be able to identify if the semantic meaning of augmented views are changed without the ground truth.\n\n\nTherefore, in this work, we start by investigating the training behavior of contrastive learning models with the presence of bad positive pairs using simulated data. As shown in Figure3, we find noisy positive pairs exhibit relatively small losses throughout the training process. In contrast, faulty positive pairs are often associated with large losses during training.\n\n\nInspired by these observations, we design a dynamic bad pair mining (DBPM) algorithm. The proposed DBPM algorithm addresses the bad positive pair problem with a simple concept:identify and suppress. Specifically, DBPM aims for reliable bad positive pair mining and down-weighting in the contrastive training. To this end, DBPM first utilizes amemory moduleto track the training behavior of each pair along training process. The memory module allows us to dynamically identify potential bad positive pairs at each epoch based on their historical training behaviors, which makes the identification more reliable. In addition, we design atransformation moduleto estimate suppressing weights for bad positive pairs. With this design, DBPM reliably reduces negative effects of bad positive pairs in the learning process, which in turn improves the quality of learned representations.\n\n\nOverall, the contributions of this work can be summarized in three aspects.First, to the best of our knowledge, this is the first study to investigate the bad positive pair problem exists in time series contrastive learning. Our study contributes to a deeper understanding of an important issue that has not been thoroughly explored.Second, we propose DBPM, a simple yet effective algorithm designed as a lightweight plug-in that dynamically deals with potential bad positive pairs along the contrastive learning process, thereby improving the quality of learned time series representations.Third, extensive experiments over four real-world large-scale time series datasets demonstrate the efficacy of DBPM in enhancing the performance of existing state-of-the-art methods.\n\n",
    "related_work": "2.1Self-Supervised Contrastive LearningContrastive learning aims to maximize the agreement between different yet related views from the same instance (i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e ., positive pairs), since they are assumed to share the meaningful underlying semantics(Chen et al.,2020; He et al.,2020a; Zbontar et al.,2021; Grill et al.,2020). By minimizing the InfoNCE loss(Oord et al.,2018), the contrastive model forces positive pairs to align in representation space. Therefore, the success of contrastive learning heavily relies on the view design, while training on inappropriate views can be detrimental to model performance(Wang & Qi,2022a; Tian et al.,2020; Wen & Li,2021).A few prior works in vision domain explored potential faulty positive pairs problem. RINCE(Chuang et al.,2022)alleviated this problem by designing a symmetrical InfoNCE loss function that is robust to faulty views. Weighted xID(Morgado et al.,2021)then down-weighted suspicious audio-visual pairs in audio-visual correspondence learning. Our DBPM differs from image-based solutions in following aspects. First, DBPM is designed for time series contrastive learning, in which not only faulty positive pairs arise, but also noisy positive pairs exist. The particular challenge of addressing two different types of bad positive pairs is unique to time series data and distinguishes them from image data. Adapting methods directly from image data may not effectively address the challenges posed by these pairs in the time series context. Our DBPM is designed based on the unique characteristics of these pairs within time series data, thus capable of handling both two types of bad pair simultaneously. Furthermore, DBPM identifies bad positive pairs based on their historical training behaviors in a dynamic manner, which allows more reliable identification of potential bad positive pairs.\n\n\n2.2Time Series Contrastive LearningSeveral recent studies have demonstrated that contrastive learning is a prominent self-supervised approach for time series representation learning. For example, TSTCC(Eldele et al.,2021)proposed temporal and contextual contrasting that worked with a weak-strong augmentation strategy to learn discriminative time series representation. BTSF(Yang & Hong,2022)presented a bilinear temporal-spectral fusion module that utilized the temporal-spectral affinities to improve the expressiveness of the representations. CoST(Woo et al.,2022)applied contrastive learning to learn disentangled seasonal-trend representations for long sequence time series forecasting. TS2Vec(Yue et al.,2022)utilized hierarchical contrasting over augmented context views such that the learned contextual representation for each timestamp is robust. TF-C(Zhang et al.,2022b)leveraged time-frequency consistency as a mechanism in the pre-training to facilitate knowledge transfer between datasets.\n\n\n2.3Learning with Label ErrorBased on the fact that contrasting views provide supervision information to each other(Wang et al.,2022), we relate the bad positive pair problem to the label error problem in supervised learning. Works inSwayamdipta et al. (2020); Shen & Sanghavi (2019)showed that data with label error exhibits different training behaviors compared to normal data. Inspired by these observations, we analyzed the training behavior (e.g.formulae-sequence𝑒𝑔e.g.italic_e . italic_g ., individual pair loss and its variance during training) of bad positive pair using simulated data in the context of time series contrastive learning, and further design DBPM to dynamically deal with bad positive pairs based on observations.Figure 2:Graphical illustration of time series contrastive learning with DBPM. DBPM consists of amemory moduleMand atransformation moduleT.Mrecords the training behaviors of each positive pair along training procedure (i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e .,m(i,e)subscript𝑚𝑖𝑒m_{(i,e)}italic_m start_POSTSUBSCRIPT ( italic_i , italic_e ) end_POSTSUBSCRIPT), and generates a global statisticℳesubscriptℳ𝑒\\mathcal{M}_{e}caligraphic_M start_POSTSUBSCRIPT italic_e end_POSTSUBSCRIPTfor identifying potential bad positive pairs at each epoch. These pairs are then down-weighted usingw(i,e)subscript𝑤𝑖𝑒w_{(i,e)}italic_w start_POSTSUBSCRIPT ( italic_i , italic_e ) end_POSTSUBSCRIPTestimated fromT. DBPM is simple yet effective, and can be easily integrated into existing frameworks to improve their performance (as shown in upper orange box).\n\n",
    "origin_citations": [
      "A simple framework for contrastive learning of visual representations.",
      "Momentum contrast for unsupervised visual representation learning.",
      "Barlow twins: Self-supervised learning via redundancy reduction.",
      "Bootstrap your own latent-a new approach to self-supervised learning.",
      "Representation learning with contrastive predictive coding.",
      "Contrastive learning with stronger augmentations.",
      "What makes for good views for contrastive learning?",
      "Toward understanding the feature learning process of self-supervised contrastive learning.",
      "Robust contrastive learning against noisy views.",
      "Robust audio-visual instance discrimination.",
      "Time-series representation learning via temporal and contextual contrasting.",
      "Unsupervised time-series representation learning with iterative bilinear temporal-spectral fusion.",
      "Cost: Contrastive learning of disentangled seasonal-trend representations for time series forecasting.",
      "Ts2vec: Towards universal representation of time series.",
      "Self-supervised contrastive pre-training for time series via time-frequency consistency.",
      "Rethinking minimal sufficient representation in contrastive learning.",
      "Dataset cartography: Mapping and diagnosing datasets with training dynamics.",
      "Learning with bad training data via iterative trimmed loss minimization."
    ],
    "citations": [
      7000,
      122,
      217616,
      2517,
      134,
      716211,
      2528,
      902598,
      254842,
      506225,
      392543,
      89577,
      90638,
      394151,
      82153,
      516650,
      309793,
      4125
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 83,
      "Final_Score": 82.3,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 85\nSynthesis_Analysis: 83\nFinal_Score: 82"
    },
    "overall_score": 83.97999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "K2c04ulKXn",
        "forum": "K2c04ulKXn",
        "number": 2562,
        "title": "Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach",
        "normalized_title": "towards enhancing time series contrastive learning a dynamic bad pair mining approach",
        "abstract": "*Not all positive pairs are beneficial to time series contrastive learning*. In this paper, we study two types of bad positive pairs that can impair the quality of time series representation learned through contrastive learning: the noisy positive pair and the faulty positive pair. We observe that, with the presence of noisy positive pairs, the model tends to simply learn the pattern of noise (Noisy Alignment). Meanwhile, when faulty positive pairs arise, the model wastes considerable amount of effort aligning non-representative patterns (Faulty Alignment). To address this problem, we propose a Dynamic Bad Pair Mining (DBPM) algorithm, which reliably identifies and suppresses bad positive pairs in time series contrastive learning. Specifically, DBPM utilizes a memory module to dynamically track the training behavior of each positive pair along training process. This allows us to identify potential bad positive pairs at each epoch based on their historical training behaviors. The identified bad pairs are subsequently down-weighted through a transformation module, thereby mitigating their negative impact on the representation learning process. DBPM is a simple algorithm designed as a lightweight **plug-in** without learnable parameters to enhance the performance of existing state-of-the-art methods. Through extensive experiments conducted on four large-scale, real-world time series datasets, we demonstrate DBPM's efficacy in mitigating the adverse effects of bad positive pairs.",
        "authors": [
          "Xiang Lan",
          "Hanshu Yan",
          "Shenda Hong",
          "Mengling Feng"
        ],
        "decision": null,
        "pdate": 1705410850294,
        "cdate": 1695215590765
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "K2c04ulKXn",
          "forum": "K2c04ulKXn",
          "number": 2562,
          "title": "Towards Enhancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach",
          "normalized_title": "towards enhancing time series contrastive learning a dynamic bad pair mining approach",
          "abstract": "*Not all positive pairs are beneficial to time series contrastive learning*. In this paper, we study two types of bad positive pairs that can impair the quality of time series representation learned through contrastive learning: the noisy positive pair and the faulty positive pair. We observe that, with the presence of noisy positive pairs, the model tends to simply learn the pattern of noise (Noisy Alignment). Meanwhile, when faulty positive pairs arise, the model wastes considerable amount of effort aligning non-representative patterns (Faulty Alignment). To address this problem, we propose a Dynamic Bad Pair Mining (DBPM) algorithm, which reliably identifies and suppresses bad positive pairs in time series contrastive learning. Specifically, DBPM utilizes a memory module to dynamically track the training behavior of each positive pair along training process. This allows us to identify potential bad positive pairs at each epoch based on their historical training behaviors. The identified bad pairs are subsequently down-weighted through a transformation module, thereby mitigating their negative impact on the representation learning process. DBPM is a simple algorithm designed as a lightweight **plug-in** without learnable parameters to enhance the performance of existing state-of-the-art methods. Through extensive experiments conducted on four large-scale, real-world time series datasets, we demonstrate DBPM's efficacy in mitigating the adverse effects of bad positive pairs.",
          "authors": [
            "Xiang Lan",
            "Hanshu Yan",
            "Shenda Hong",
            "Mengling Feng"
          ],
          "decision": null,
          "pdate": 1705410850294,
          "cdate": 1695215590765
        }
      ]
    }
  },
  {
    "paper_id": 9691,
    "title": "DQ-LoRe: Dual Queries with Low Rank Approximation Re-ranking for In-Context Learning",
    "abstract": "Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning.\nDual Queries first query\nLLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question’s knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5% to 94.2%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundary of in-context learning and opens up new avenues for addressing complex reasoning challenges. Our code is released athttps://github.com/AI4fun/DQ-LoRe.",
    "introduction": "Recently, significant advancements in natural language processing (NLP) have been driven by large language models (LLMs)(Chen et al.,2021; Chowdhery et al.,2022; Ouyang et al.,2022; Touvron et al.,2023a;b; Anil et al.,2023; OpenAI,2023). With the increasing capabilities of LLMs, in-context learning (ICL) has emerged as a new paradigm, where LLMs make predictions based on contexts augmented with a few exemplars(Brown et al.,2020). An important question in the field of in-context learning is how to improve the selection of in-context exemplars to enhance the performance of LLMs(Liu et al.,2022).\n\n\nSelecting exemplars for ICL poses challenges due to their instability(Zhao et al.,2021). Even minor changes in the order of samples within exemplars can affect the output(Lu et al.,2022; Su et al.,2023a). The selection of exemplars for LLMs is currently a community-wide trial and error effort, as it is difficult to extract generalizable regularity from empirical observations to form effective selection criteria(Fu et al.,2022; Zhang et al.,2022b). One exception is retrieval-based exemplar acquisition methods(Rubin et al.,2021; Liu et al.,2022; Ye et al.,2023; Li et al.,2023), where a retriever is used to select similar exemplars based on input questions during inference.\n\n\nHowever, these methods primarily focus on the similarity between input questions and examples in the training set, without fully exploiting the relationship between intermediate reasoning steps of the given question and other exemplars in the pool. Previous studies have shown that considering such chain-of-thought (CoT) can further improve the performance of LLMs on multi-step reasoning tasks(Wei et al.,2022b; Fu et al.,2022; Gao et al.,2023). Furthermore, some work has also observed that the transfer of knowledge between LLMs and retrievers can effectively enhance the common sense reasoning capabilities of LLMsXu et al. (2023). Additionally, we observed that prior efforts(Ye et al.,2023; Rubin et al.,2021)struggle to distinguish exemplars in high-dimensional embedding spaces. These observations suggest that exemplar selection based solely on trained question embeddings may suffer from redundant information within the “universal” representations, and may not effectively capture inherent relevance. Removing these redundant information often leads to improved speed and effectiveness(Wang et al.,2023b). The sentence embeddings within the retrieved exemplars often contain similar information, which commonly results in a dense and non-uniform distribution in the vector space. We posit that this is typically due to the embeddings encoding a significant amount of redundant information and exhibiting anisotropy. Employing Principal Component Analysis (PCA;Wold et al.1987) for dimensionality reduction can assist in filtering out this redundant information and distinguishing between different exemplars, effectively facilitating a more uniform distribution of representations within the vector space.\n\n\nTo address these challenges, we propose a framework that leverages Dual Queries with Low-rank approximation Re-ranking (DQ-LoRe) to incorporate CoTs beyond the input questions, improving the exemplar selection process for in-context learning. DQ-LoRe first queries LLM to generate CoT for a given question. We then concatenate CoT with the question to query the retriever and obtain exemplars from the training pool. We further apply PCA for dimensionality reduction to filter out redundant information and differentiate between different exemplars, improving the selection process.\n\n\nWe conduct extensive experiments on various multi-step reasoning benchmarks to evaluate the performance of DQ-LoRe. The results demonstrate that DQ-LoRe effectively and efficiently selects exemplars, outperforming existing methods. Furthermore, DQ-LoRe exhibits robustness and adaptability in the distribution shift setting, highlighting its versatility across different scenarios. These findings have implications for the use of low-rank constraints in the LLMs paradigm. Our contributions can be summarized as follows:•We introduce DQ-LoRe, a method that queries supplementary information from Large Language Models (LLMs) to subsequently re-query a smaller-scale retrieval model. Upon acquiring re-ranked exemplars from the low-rank small model, DQ-LoRe then supplies these exemplars to the LLMs for inference, thereby effectively tackling the challenge associated with the selection of exemplars.•We employ straightforward and efficient dimensionality reduction techniques to extract crucial reasoning information from the high-dimensional representations of CoTs and questions. This enables the differentiation between various exemplars, particularly distinguishing between exemplars characterized by word co-occurrence and spurious question-related associations and those exemplars that exhibit genuine logical relevance.•We demonstrate that DQ-LoRe achieves superior performance compared to existing methods and is particularly effective in the distribution shift setting, showcasing its robustness and adaptability across various scenarios.\n\n\nFigure 1:The overall pipeline of DQ-LoRe. It consists of three parts:Dual Queriesfirst query LLM to obtain CoTy𝑦yitalic_y, then query the retriever to obtain the final exemplars via both question and LLM-generated knowledge.LoReleverages PCA to approximate the low-rank embedding of retrieved exemplars, enabling us to better distinguish them.Retrieverobtains exemplars with similar CoT, through training with positive and negative sets constructed based on CoT similarity produced by BM25 and LLM.\n\n",
    "related_work": "In-Context LearningLLMs have demonstrated their in-context learning ability with the scaling of model size and corpus size(Brown et al.,2020; Chowdhery et al.,2022; OpenAI,2023). This ability allows language models to learn tasks with only a few exemplars. Several studies have shown that LLMs can successfully perform various complex tasks using in-context learning, including natural language understanding and multi-step reasoning(Shin et al.,2020; Sanh et al.,2022; Liu et al.,2023). In addition to in-context exemplars,Wei et al. (2022b)have explored augmenting the learning process with CoT. CoT involves providing a sequence of intermediate reasoning steps along with the in-context exemplars. Further studies show that the effectiveness of CoT can be enhanced through various approaches. These approaches include breaking down complex questions(Zhou et al.,2022), planning before inference(Wang et al.,2023a), and employing the CoT paradigm for multiple rounds of voting and reasoning(Wang et al.,2022; Zheng et al.,2023). Notably, in the case of multi-step reasoning, in-context learning with CoT has been found to outperform fine-tuning conducted on the same large model with the full training set(Lewkowycz et al.,2022; Wei et al.,2022a).\n\n\nExemplar SelectionThe selection of exemplars for in-context learning is a fundamental question. However, previous studies have highlighted the challenges and instability of exemplar selection(Zhao et al.,2021; Lu et al.,2022; Su et al.,2023a). Even slight changes in the order of samples within exemplars can affect the model’s output. The acquisition of exemplars is crucial for enhancing multi-step reasoning capabilities(Liu et al.,2022). Existing efforts mainly focus on the human-designed approach, the vanilla CoT(Wei et al.,2022b)utilizes eight manually written examples, while PAL(Gao et al.,2023)repurposes these exemplars by converting them into programming language statements. Complex-CoT(Fu et al.,2022)selects exemplars with the most complex CoTs from the training set, resulting in improved performance on multi-step reasoning tasks. Auto-CoTZhang et al. (2022b)clusters training instances intok𝑘kitalic_kcategories and selectsk𝑘kitalic_ksamples closest to the cluster center.Other efforts adopt a retrieval-based method that leverages encoders to encode exemplars and input questions during training(Liu et al.,2022; Rubin et al.,2021; Ye et al.,2023). This enables the selection of exemplars that are close to the vector representation of the input questions. For example, Efficient Prompt Retrieval (EPR;Rubin et al.2021) models the interaction between input questions and in-context exemplars and optimizes it through a contrastive learning objective to obtain preferred exemplars. Compositional Exemplars for In-context Learning (CEIL;Ye et al.2023) utilizes Determinantal Point Processes to model the interplay between the provided input and in-context exemplars. This modeling is further enhanced through a meticulously designed contrastive learning objective, with the goal of extracting preferences from language models.Li et al. (2023)proposes a unified retriever to retrieve exemplars for a wide range of tasks. Unlike these methods, we propose to model the relationship between the reasoning process through re-ranking in the representation space after projecting the original representation, enabling better exemplar selection.\n\n",
    "origin_citations": [
      "Language models are few-shot learners.",
      "Palm: Scaling language modeling with pathways.",
      "GPT-4 technical report.",
      "Autoprompt: Eliciting knowledge from language models with automatically generated prompts.",
      "Multitask prompted training enables zero-shot task generalization.",
      "Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing.",
      "Chain-of-thought prompting elicits reasoning in large language models.",
      "Least-to-most prompting enables complex reasoning in large language models.",
      "Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models.",
      "Self-consistency improves chain of thought reasoning in language models.",
      "Progressive-hint prompting improves reasoning in large language models.",
      "Solving quantitative reasoning problems with language models.",
      "Emergent abilities of large language models.",
      "Calibrate before use: Improving few-shot performance of language models.",
      "Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity.",
      "Selective annotation makes language models better few-shot learners.",
      "What makes good in-context examples for gpt-3?",
      "PAL: program-aided language models.",
      "Complexity-based prompting for multi-step reasoning.",
      "Automatic chain of thought prompting in large language models.",
      "Learning to retrieve prompts for in-context learning.",
      "Compositional exemplars for in-context learning.",
      "Unified demonstration retriever for in-context learning."
    ],
    "citations": [
      679,
      1554,
      9115,
      7787,
      8469,
      408717,
      1578,
      424,
      116223,
      2191,
      53971,
      5563,
      8556,
      1594,
      8470,
      3374,
      2189,
      7957,
      5953,
      2467,
      3376,
      3378,
      3375
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 85,
      "Final_Score": 82.9,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 85\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 91,
      "Synthesis_Analysis": 78,
      "Final_Score": 84.69999999999999,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 91\nSynthesis_Analysis: 78\nFinal_Score: 84"
    },
    "overall_score": 83.97999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "qAoxvePSlq",
        "forum": "qAoxvePSlq",
        "number": 4566,
        "title": "DQ-LoRe: Dual Queries with Low Rank Approximation Re-ranking for In-Context Learning",
        "normalized_title": "dq lore dual queries with low rank approximation re ranking for in context learning",
        "abstract": "Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning. Dual Queries first query LLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question's knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5\\% to 94.2\\%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundaries of in-context learning and opens up new avenues for addressing complex reasoning challenges.",
        "authors": [
          "Jing Xiong",
          "Zixuan Li",
          "Chuanyang Zheng",
          "Zhijiang Guo",
          "Yichun Yin",
          "Enze Xie",
          "Zhicheng YANG",
          "Qingxing Cao",
          "Haiming Wang",
          "Xiongwei Han",
          "Jing Tang",
          "Chengming Li",
          "Xiaodan Liang"
        ],
        "decision": null,
        "pdate": 1705410910451,
        "cdate": 1695353097661
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "qAoxvePSlq",
          "forum": "qAoxvePSlq",
          "number": 4566,
          "title": "DQ-LoRe: Dual Queries with Low Rank Approximation Re-ranking for In-Context Learning",
          "normalized_title": "dq lore dual queries with low rank approximation re ranking for in context learning",
          "abstract": "Recent advances in natural language processing, primarily propelled by Large Language Models (LLMs), have showcased their remarkable capabilities grounded in in-context learning. A promising avenue for guiding LLMs in intricate reasoning tasks involves the utilization of intermediate reasoning steps within the Chain-of-Thought (CoT) paradigm. Nevertheless, the central challenge lies in the effective selection of exemplars for facilitating in-context learning. In this study, we introduce a framework that leverages Dual Queries and Low-rank approximation Re-ranking (DQ-LoRe) to automatically select exemplars for in-context learning. Dual Queries first query LLM to obtain LLM-generated knowledge such as CoT, then query the retriever to obtain the final exemplars via both question and the knowledge. Moreover, for the second query, LoRe employs dimensionality reduction techniques to refine exemplar selection, ensuring close alignment with the input question's knowledge. Through extensive experiments, we demonstrate that DQ-LoRe significantly outperforms prior state-of-the-art methods in the automatic selection of exemplars for GPT-4, enhancing performance from 92.5\\% to 94.2\\%. Our comprehensive analysis further reveals that DQ-LoRe consistently outperforms retrieval-based approaches in terms of both performance and adaptability, especially in scenarios characterized by distribution shifts. DQ-LoRe pushes the boundaries of in-context learning and opens up new avenues for addressing complex reasoning challenges.",
          "authors": [
            "Jing Xiong",
            "Zixuan Li",
            "Chuanyang Zheng",
            "Zhijiang Guo",
            "Yichun Yin",
            "Enze Xie",
            "Zhicheng YANG",
            "Qingxing Cao",
            "Haiming Wang",
            "Xiongwei Han",
            "Jing Tang",
            "Chengming Li",
            "Xiaodan Liang"
          ],
          "decision": null,
          "pdate": 1705410910451,
          "cdate": 1695353097661
        }
      ]
    }
  },
  {
    "paper_id": 43404,
    "title": "Is Multiple Object Tracking a Matter of Specialization?",
    "abstract": "End-to-end transformer-based trackers have achieved remarkable performance on most human-related datasets. However, training these trackers in heterogeneous scenarios poses significant challenges, including negative interference – where the model learns conflicting scene-specific parameters – and limited domain generalization, which often necessitates expensive fine-tuning to adapt the models to new domains. In response to these challenges, we introduceParameter-efficientScenario-specificTrackingArchitecture (PASTA), a novel framework that combines Parameter-Efficient Fine-Tuning (PEFT) and Modular Deep Learning (MDL). Specifically, we define key scenario attributes (e.g., camera-viewpoint, lighting condition) and train specialized PEFT modules for each attribute. These expert modules are combined in parameter space, enabling systematic generalization to new domains without increasing inference time. Extensive experiments on MOTSynth, along with zero-shot evaluations on MOT17 and PersonPath22 demonstrate that a neural tracker built from carefully selected modules surpasses its monolithic counterpart. We release models and code.",
    "introduction": "Video Surveillance is essential for enhancing security, supporting law enforcement, improving safety, and increasing operational efficiency across various sectors. In this respect, Multiple Object Tracking (MOT) is a widely studied topic due to its inherent complexity. Nowadays, MOT is commonly tackled with two main paradigms: tracking-by-detection (TbD)[3,51,61,28,43,38,29]or query-based tracking[56,58,63,12](i.e., tracking-by-attention). Although tracking-by-detection methods have proven effective across multiple datasets, their performance struggles to scale on larger datasets due to the non-differentiable mechanism used for linking new detections to existing tracks. To this end, query-based methods are being employed to unify the detection and association phase.\n\n\nNevertheless, training such end-to-end transformer-based methods presents significant challenges, as they tend to overfit specific scenario settings[37,55](e.g., camera viewpoint, indoorvs.outdoor environments), require vast amounts of data[63], and incur substantial computational costs. Moreover, these methods degrade under domain shifts, struggling to outperform traditional TbD methods.\n\n\nIn light of these challenges, we propose a novel framework,Parameter-efficientScenario-specificTrackingArchitecture (PASTA), aimed at reducing the computational costs and enhancing the transfer capabilities of such models. Leveraging Parameter Efficient Fine-Tuning (PEFT) techniques[16,34]can significantly decrease computational expenses and training time, starting with a frozen backbone pre-trained on synthetic data. However, the model may still experiencenegative interference[49,50,37,55], a phenomenon for which training on multiple tasks (or scenarios) causes the model to learn task-specific parameters that may conflict. For instance, if the model learns parameters tailored for an indoor sports activity, it could detrimentally affect its performance on a novel outdoor scene depicting people walking. To this end, inspired by Modular Deep Learning (MDL)[35], we employ a lightweight expert module for each attribute, learn them separately, and finally compose them efficiently[17]. This approach – depicted inFig.1– is akin to a chef preparing a pasta dish. Each ingredient (i.e., module) is prepared individually to preserve its unique flavor and then combined harmoniously to create a balanced dish. Moreover, as pasta must be perfectlyal denteto serve as the ideal base for various sauces, the pre-trained backbone should be robust and well-tuned to serve as the foundation for the modules. These modules must be combined effectively to ensure the model performs well across diverse scenarios. Conversely, the result will be sub-optimal if incompatible modules are mixed – analogous to combining ingredients that do not complement each other. Indeed, combining contrasting modules can lead to ineffective handling of diverse tasks.\n\n\nFigure 1:Given a scene, we select the modules corresponding to its attributes, such as lighting and indoor/outdoor. These modules are composed and then deployed, yielding a specialized model.\n\n\nNotably, such a modular framework brings two advantages: it avoids negative interference and enhances generalization by leveraging domain-specific knowledge. Firstly, starting from a pre-trained backbone, we train each module independently to prevent parameter conflicts, ensuring that gradient updates are confined to the relevant module for the specific scenario. This assures that parameters learned for one attribute do not negatively impact the performance of another. Secondly, the modular approach allows us to exploit domain knowledge fully, even when encountering a novel attribute combination. Indeed, as shown inSec.5.5, our approach is effective even in a zero-shot setting (i.e., without further fine-tuning on the target dataset). Moreover, the selection of the modules may be done automatically or in a more realistic production environment by video surveillance operators.\n\n\nTo evaluate our approach, we conduct extensive experiments on the synthetic MOTSynth[10]and the real-world MOT17[8]and PersonPath22[44]datasets. The results show that PASTA can effectively leverage the knowledge learned by the modules to improve tracking performance on both the source dataset and in zero-shot scenarios. To summarize, we highlight the following main contributions:•We propose PASTA, a novel framework for Multiple Object Tracking built on Modular Deep Learning, enabling the fine-tuning of query-based trackers with PEFT techniques.•By incorporating expert modules, we improve domain transfer and prevent negative interference while fine-tuning MOT models.•Comprehensive evaluation confirms the validity of our approach and its effectiveness in zero-shot tracking scenarios.\n\n",
    "related_work": "Multiple Object Tracking.The most widely adopted paradigm for Multiple Object Tracking (MOT) istracking-by-detection(TbD)[3,51,61,28,43,38]. First, an object detector (e.g., YOLOX[14]) localizes objects in the current frame. Next, the association step matches detections to tracks from the previous frame by solving a minimum-cost bipartite matching problem, with the association cost defined in various forms (e.g., IoU[3,61], GIoU[40], or geometrical cues[29,33]).\nThis pairing typically occurs immediately after propagating the previous tracks to the current frame using a motion model (e.g., Kalman Filter[19]). Notably, methods following such paradigm have succeeded on complex human-related MOT benchmarks[8,9,47,44]. In TbD, the detection and data-association steps are equally crucial to accurately localizing and tracking objects.\nRecent works[65,1]have attempted to unify these steps; however, progress toward a fully unified algorithm was constrained by a significant limitation – the data association process (e.g., the Hungarian algorithm[20]) is inherently non-differentiable. An initial effort was made by Xuet al.[53]that proposed a differentiable version of the Hungarian algorithm, later advanced by end-to-end transformer-based trackers[31,58,63,56,13].\n\n\nHowever, transformer-based trackers (also known as tracking-by-attention) require large amounts of data to achieve decent generalization capabilities[58,31]. Due to the data scarcity in MOT, these models often overfit to the specific domain they were trained on, which hampers their ability to generalize to different domains[17,21,37].\n\n\nModular Deep Learning (MDL).Considering recent trends in the field of deep learning, state-of-the-art models have become increasingly larger. Consequentially, fine-tuning these models has become expensive; concurrently, they still struggle with tasks like symbolic reasoning and temporal understanding[35]. Recent learning paradigms based onModular Deep Learning(MDL)[35]can address these challenges by disentangling core pre-training knowledge from domain-specific capabilities. By applying modularity principles, deep models can be easily edited, allowing for the seamless integration of new capabilities and the selective removal of existing ones[26,36].\n\n\nSpecifically, lightweight computation functions namedmodulesare employed to adapt a pre-trained neural network. To do so, several fine-tuning techniques could be used to realize these modules, such as LoRA[16], (IA)3[25], and SSF[23]. These multiple modules can be learned on different tasks such that they can specialize in different concepts[32]. At inference time, not all modules have to be active at the same time. Instead, they can be selectively utilized as needed, either based on prior knowledge of the domain or dynamically in response to the current input. To establish which modules to activate, it is common practice to rely on arouting function, which can be either learned or fixed. Finally, the outputs of the selected modules are combined using anaggregation function. To minimize inference costs, this process is usually performed in the parameter space rather than the output space, an activity often referred to asmodel merging[54]. Specifically, a single forward pass is performed using weights generated by a linear combination of those selected by the routing function.\n\n\nDomain adaptation and open-vocabulary approaches in MOT.Currently, domain adaptation techniques have only been applied to tracking-by-detection methods, with GHOST[43]and DARTH[42]serving as notable examples. In particular, GHOST adapts the visual encoder employed to feed the appearance model by updating the sufficient statistics of the Batch Normalization layers during inference. In contrast, our approach regards tracking-by-attention approaches and adapts the entire network. Moreover, DARTH employs test-time adaptation (TTA)[24]and Knowledge Distillation, requiring multiple forward passes and entire sequences, making it computationally heavy and less practical for real-time use. In contrast, our method is entirely online and requires only basic target scene attributes, with no further training during deployment.\n\n\nRecent advances in zero-shot tracking have focused onopen-vocabulary tracking, where the model can track novel object categories by prompting it with the corresponding textual representation. In this respect, methods like OVTrack[22]and Z-GMOT[48]leverage CLIP[39]and language-based pre-training, while OVTracktor[7]extends tracking to any category. Our method does not use open-vocabulary models but emphasizes domain knowledge transfer in end-to-end trackers.\n\n",
    "origin_citations": [
      "Simple online and realtime tracking.",
      "Simple online and realtime tracking with a deep association metric.",
      "Bytetrack: Multi-object tracking by associating every detection box.",
      "Rethinking pseudo-lidar representation.",
      "Simple cues lead to a strong multi-object tracker.",
      "Motiontrack: Learning robust short-term and long-term motions for multi-object tracking.",
      "Yolox: Exceeding yolo series in 2021.",
      "Generalized intersection over union: A metric and a loss for bounding box regression.",
      "Trackflow: Multi-object tracking with normalizing flows.",
      "Distformer: Enhancing local and global features for monocular per-object distance estimation.",
      "A new approach to linear filtering and prediction problems.",
      "Motchallenge: A benchmark for single-camera multiple target tracking.",
      "Mot20: A benchmark for multi object tracking in crowded scenes.",
      "Dancetrack: Multi-object tracking in uniform appearance and diverse motion.",
      "Large scale real-world multi-person tracking.",
      "Tracking objects as points.",
      "Tracking without bells and whistles.",
      "The hungarian method for the assignment problem.",
      "How to train your deep multi-object tracker.",
      "Trackformer: Multi-object tracking with transformers.",
      "Motr: End-to-end multiple-object tracking with transformer.",
      "Motrv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors.",
      "Motrv3: Release-fetch supervision for end-to-end multi-object tracking.",
      "Multiple object tracking as id prediction.",
      "Compositionality decomposed: how do neural networks generalise?",
      "Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks.",
      "Towards generalizable multi-object tracking.",
      "Modular deep learning.",
      "Tangent model composition for ensembling and continual fine-tuning.",
      "A second-order perspective on model compositionality and incremental learning.",
      "Lora: Low-rank adaptation of large language models.",
      "Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning.",
      "Scaling & shifting your features: A new baseline for efficient model tuning.",
      "Task arithmetic in the tangent space: Improved editing of pre-trained models.",
      "Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities.",
      "Darth: Holistic test-time adaptation for multiple object tracking.",
      "A comprehensive survey on test-time adaptation under distribution shifts.",
      "Ovtrack: Open-vocabulary multiple object tracking.",
      "Z-GMOT: Zero-shot generic multiple object tracking.",
      "Learning transferable visual models from natural language supervision.",
      "Zero-shot open-vocabulary tracking with large pre-trained models."
    ],
    "citations": [
      4266,
      184177,
      340913,
      161177,
      412168,
      495570,
      410719,
      8216,
      405514,
      484979,
      575535,
      253560,
      337305,
      34395,
      207187,
      218892,
      1040545,
      39390,
      250446,
      245284,
      259390,
      424378,
      42150,
      974196,
      1040587,
      226613,
      470927,
      893365,
      1589,
      377481,
      336958,
      260249,
      777205,
      323643,
      493155,
      54576,
      257430,
      1639,
      321105
    ],
    "extract_rate": 0.9512195121951219,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 75,
      "Final_Score": 82.3,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 75\nFinal_Score: 82"
    },
    "overall_score": 83.97999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "aujnNnIiiM",
        "forum": "aujnNnIiiM",
        "number": 9252,
        "title": "Is Multiple Object Tracking a Matter of Specialization?",
        "normalized_title": "is multiple object tracking a matter of specialization",
        "abstract": "End-to-end transformer-based trackers have achieved remarkable performance on most human-related datasets. However, training these trackers in heterogeneous scenarios poses significant challenges, including negative interference - where the model learns conflicting scene-specific parameters - and limited domain generalization, which often necessitates expensive fine-tuning to adapt the models to new domains. In response to these challenges, we introduce Parameter-efficient Scenario-specific Tracking Architecture (PASTA), a novel framework that combines Parameter-Efficient Fine-Tuning (PEFT) and Modular Deep Learning (MDL). Specifically, we define key scenario attributes (e.g, camera-viewpoint, lighting condition) and train specialized PEFT modules for each attribute. These expert modules are combined in parameter space, enabling systematic generalization to new domains without increasing inference time. Extensive experiments on MOTSynth, along with zero-shot evaluations on MOT17 and PersonPath22 demonstrate that a neural tracker built from carefully selected modules surpasses its monolithic counterpart. We release models and code.",
        "authors": [
          "Gianluca Mancusi",
          "Mattia Bernardi",
          "Aniello Panariello",
          "Angelo Porrello",
          "Rita Cucchiara",
          "Simone Calderara"
        ],
        "decision": null,
        "pdate": 1727287905214,
        "cdate": 1715677377742
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "aujnNnIiiM",
          "forum": "aujnNnIiiM",
          "number": 9252,
          "title": "Is Multiple Object Tracking a Matter of Specialization?",
          "normalized_title": "is multiple object tracking a matter of specialization",
          "abstract": "End-to-end transformer-based trackers have achieved remarkable performance on most human-related datasets. However, training these trackers in heterogeneous scenarios poses significant challenges, including negative interference - where the model learns conflicting scene-specific parameters - and limited domain generalization, which often necessitates expensive fine-tuning to adapt the models to new domains. In response to these challenges, we introduce Parameter-efficient Scenario-specific Tracking Architecture (PASTA), a novel framework that combines Parameter-Efficient Fine-Tuning (PEFT) and Modular Deep Learning (MDL). Specifically, we define key scenario attributes (e.g, camera-viewpoint, lighting condition) and train specialized PEFT modules for each attribute. These expert modules are combined in parameter space, enabling systematic generalization to new domains without increasing inference time. Extensive experiments on MOTSynth, along with zero-shot evaluations on MOT17 and PersonPath22 demonstrate that a neural tracker built from carefully selected modules surpasses its monolithic counterpart. We release models and code.",
          "authors": [
            "Gianluca Mancusi",
            "Mattia Bernardi",
            "Aniello Panariello",
            "Angelo Porrello",
            "Rita Cucchiara",
            "Simone Calderara"
          ],
          "decision": null,
          "pdate": 1727287905214,
          "cdate": 1715677377742
        }
      ]
    }
  },
  {
    "paper_id": 50208,
    "title": "See What You Are Told: Visual Attention Sink in Large Multimodal Models",
    "abstract": "Large multimodal models (LMMs) “see” images by leveraging the attention mechanism between text and visual tokens in the transformer decoder. Ideally, these models should focus on key visual information relevant to the text token. However, recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens, even when these tokens are irrelevant to the corresponding text. In this study, we investigate the property behind the appearance of these irrelevant visual tokens and examine their characteristics. Our findings show that this behavior arises due to the massive activation of certain hidden state dimensions, which resembles the attention sink found in language models. Hence, we refer to this phenomenon as thevisual attention sink. In particular, our analysis reveals that removing the irrelevant visual sink tokens does not impact model performance, despite receiving high attention weights. Consequently, we recycle the attention to these tokens as surplus resources, redistributing the attention budget to enhance focus on the image. To achieve this, we introduce Visual Attention Redistribution (VAR), a method that redistributes attention in image-centric heads, which we identify as innately focusing on visual information. VAR can be seamlessly applied across different LMMs to improve performance on a wide range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks, all without the need for additional training, models, or inference steps. Experimental results demonstrate that VAR enables LMMs to process visual information more effectively by adjusting their internal attention mechanisms, offering a new direction to enhancing the multimodal capabilities of LMMs.",
    "introduction": "Large multimodal models (LMMs) have been actively expanding the capabilities of large language models to multimodal tasks(Liu et al.,2024c;a;b; Li et al.,2023b; Bai et al.,2023). In particular, the LMMs leverage a pre-trained visual encoder(Radford et al.,2021)to process image data and the transformer decoder of a large language model to generate text responses(OpenAI,2023; Touvron et al.,2023; Yang et al.,2024). This straightforward yet powerful architecture has proven highly effective in utilizing visual information from images for vision-language tasks such as visual question answering, image captioning, and visual reasoning(Peng et al.,2023; Alayrac et al.,2022; Tsimpoukelli et al.,2021).\n\n\nTo incorporate visual information into text responses, LMMs rely on the attention mechanism(Vaswani et al.,2017)within the transformer decoder. Specifically, when processing multimodal inputs, the attention weights between visual and text tokens determine how much each text token focuses on the corresponding visual information. For instance, as illustrated in the top-left corner of Fig.1, when the text token is ‘bird’, the model concentrates on visual tokens associated with the bird in the image. Intuitively, LMMs should primarily attend to the visual tokens that are relevant to each text token.\n\n\nHowever, in practice, not all attention is directed toward the relevant visual tokens. As shown in Fig.1, the model also allocates high attention weights to visual tokens which are unrelated to the corresponding text. This phenomenon is widespread in LMMs(Woo et al.,2024; An et al.,2024), and a notable pattern emerges where irrelevant visual tokens consistently appear in fixed locations across different text tokens. For example, in each case illustrated in Fig.1, irrelevant visual tokens (highlighted in red boxes) consistently occupy the same positions regardless of the text tokens, indicating an underlying pattern. The cause and meaning of this phenomenon remain open questions, motivating this study.\n\n\nIn this work, we explore the underlying property and the characteristics of the irrelevant visual tokens. We find that these tokens in the visual attention map arise from the massive activation of specific dimensions in the hidden states. This mechanism is analogous to the attention sink observed in language models, where the model consistently assigns large attention weights to tokens with limited semantic meaning (e.g., ‘BOS’, ‘.’, ‘\\n’, etc.)(Xiao et al.,2023; Sun et al.,2024a). Irrelevant visual tokens can be identified by the extreme magnitudes in a few specific dimensions, which we refer to asvisual sink tokens, as they also have limited semantic information from the image. Furthermore, we demonstrate that removing these visual sink tokens does not significantly impact the quality of the model’s response, despite the model assigning high attention weights to them.\n\n\nFigure 1:Visual attention maps of LLaVA-1.5-7B between specified text tokens and visual tokens.Attention map visualizes where the model “see” when processing the text token. The model is expected to focus only on the visual tokens related to each text token. However, the model also attends to irrelevant visual tokens (red boxes) that are unrelated to the corresponding text token. Although we visualize the attention maps only for a few specified text tokens, these irrelevant tokens consistently occur in fixed locations across the entire text tokens, including the instructions and the generated responses (see Fig.13in Appendix for more examples).\n\n\nBased on these experiments, we propose that the attention weights assigned to sink tokens can be recycled as an “attention budget”. Since recent studies have reported that attention allocated to images is often insufficient compared to that given to text(Chen et al.,2024; Liu et al.,2024d), we redistribute the excess attention from sink tokens to image. Also, considering each attention head serves a distinct function(Zheng et al.,2024), we identify the heads that are primarily responsible for focusing on visual information, namely, image-centric heads, based on the presence of visual attention sinks. Finally, we introduce Visual Attention Redistribution (VAR), a two-step method: first, selecting the image-centric heads, and second, redistributing the attention budget to strengthen image focus within these selected heads.\n\n\nIn summary, we uncover the underlying properties of irrelevant visual tokens and demonstrate that, much like sink tokens in language models, they are unnecessary for the model’s functioning. To address this, we propose VAR, which reallocates attention from sink tokens to enhance focus on the image. Experimental results show that VAR improves the overall performance of LMMs across a range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks. Notably, VAR can be applied to various models without additional training, models, or inference steps. This suggests that the existing LMMs can readily benefit from our approach to further enhance their multimodal capabilities by intensifying their attention to images. Our work presents an effective method to address the issue of insufficient image attention and offers a new perspective on understanding the attention mechanisms within LMMs.\n\n",
    "related_work": "Visual attention in large multimodal models.In large multimodal models (LMMs), the attention mechanism between text and images plays a pivotal role in incorporating visual information into the text responses. As such, the model’s focus on images is typically represented as a visual attention map(Aflalo et al.,2022; Stan et al.,2024). However, recent findings suggest that LMMs exhibit certain unintuitive behaviors in their visual attention patterns. Specifically, LMMs tend to disproportionately focus on a few visual tokens(Woo et al.,2024; Arif et al.,2024), with some tokens receiving high attention weights regardless of the corresponding text token(An et al.,2024). Additionally, recent works have shown that LMMs often fail to adequately attend to visual information overall(Chen et al.,2024; Liu et al.,2024d). To address this issue, visual contrastive decoding(Leng et al.,2024; Favero et al.,2024)has been proposed, which contrasts the outputs of two models—one with and one without visual input to encourage greater reliance on visual cues. Further, other approaches(Zhang et al.,2024b; Zhu et al.,2024)enhance this by increasing the attention weights assigned to images, ensuring that visual information receives sufficient focus.\n\n\nAttention sink in language models.Attention sink is an intriguing phenomenon in language models, where certainsink tokenswith limited semantic meaning (e.g., ‘BOS’, ‘.’, ‘,’, ‘\\n’, etc.) receive disproportionately high attention weights(Xiao et al.,2023; Ferrando & Voita,2024). Background tokens, which contain little information, in vision transformers also exhibit similar behavior(Darcet et al.,2023), suggesting that attention sink is a common phenomenon across different modalities. Although sink tokens receive substantial attention weights, they contribute minimally to the model’s overall predictions(Kobayashi et al.,2020; Bondarenko et al.,2023). Recent research suggests that attention sink arises from the massive activation of specific dimensions within the hidden states of sink tokens, which occurs prior to the high attention allocation(Sun et al.,2024a; Cancedda,2024).Gu et al. (2024)further investigated the factors that contribute to the emergence of attention sink. Additionally,Yu et al. (2024)recalibrated the attention weights assigned to sink tokens in specific attention heads to elicit more accurate responses from language models. We extend the concept of attention sink to the multimodal domain by introducing the idea of avisual attention sinkin LMMs.\n\n",
    "origin_citations": [
      "Vl-interpret: An interactive visualization tool for interpreting vision-language transformers.",
      "Lvlm-interpret: An interpretability tool for large vision-language models, 2024.",
      "Don’t miss the forest for the trees: Attentional vision calibration for large vision language models.",
      "Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments, 2024.",
      "Agla: Mitigating object hallucinations in large vision-language models with assembly of global and local attention.",
      "An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models.",
      "Paying more attention to image: A training-free method for alleviating hallucination in lvlms.",
      "Mitigating object hallucinations in large vision-language models through visual contrastive decoding.",
      "Multi-modal hallucination control by visual information grounding.",
      "Prompt highlighter: Interactive control for multi-modal llms.",
      "Ibd: Alleviating hallucinations in large vision-language models via image-biased decoding.",
      "Efficient streaming language models with attention sinks.",
      "Information flow routes: Automatically interpreting language models at scale.",
      "Vision transformers need registers.",
      "Attention is not only a weight: Analyzing transformers with vector norms.",
      "Quantizable transformers: Removing outliers by helping attention heads do nothing.",
      "Massive activations in large language models.",
      "Spectral filters, dark signals, and attention sinks.",
      "When attention sink emerges in language models: An empirical view.",
      "Unveiling and harnessing hidden attention sinks: Enhancing large language models without training through attention calibration."
    ],
    "citations": [
      5348,
      742535,
      705079,
      1038163,
      848667,
      289046,
      830729,
      22573,
      425887,
      272790,
      462888,
      470,
      463663,
      477709,
      192138,
      48986,
      463273,
      97272,
      969713,
      875279
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 75,
      "Final_Score": 82.3,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 88\nSynthesis_Analysis: 75\nFinal_Score: 82"
    },
    "overall_score": 83.97999999999999,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "7uDI7w5RQA",
        "forum": "7uDI7w5RQA",
        "number": 1085,
        "title": "See What You Are Told: Visual Attention Sink in Large Multimodal Models",
        "normalized_title": "see what you are told visual attention sink in large multimodal models",
        "abstract": "Large multimodal models (LMMs) \"see\" images by leveraging the attention mechanism between text and visual tokens in the transformer decoder. Ideally, these models should focus on key visual information relevant to the text token. However, recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens, even when these tokens are irrelevant to the corresponding text. In this study, we investigate the property behind the appearance of these irrelevant visual tokens and examine their characteristics. Our findings show that this behavior arises due to the massive activation of certain hidden state dimensions, which resembles the attention sink found in language models. Hence, we refer to this phenomenon as the visual attention sink. In particular, our analysis reveals that removing the irrelevant visual sink tokens does not impact model performance, despite receiving high attention weights. Consequently, we recycle the attention to these tokens as surplus resources, redistributing the attention budget to enhance focus on the image. To achieve this, we introduce Visual Attention Redistribution (VAR), a method that redistributes attention in image-centric heads, which we identify as innately focusing on visual information. VAR can be seamlessly applied across different LMMs to improve performance on a wide range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks, all without the need for additional training, models, or inference steps. Experimental results demonstrate that VAR enables LMMs to process visual information more effectively by adjusting their internal attention mechanisms, offering a new direction to enhancing the multimodal capabilities of LMMs.",
        "authors": [
          "Seil Kang",
          "Jinyeong Kim",
          "Junhyeok Kim",
          "Seong Jae Hwang"
        ],
        "decision": null,
        "pdate": 1737562217273,
        "cdate": 1726481950949
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "7uDI7w5RQA",
          "forum": "7uDI7w5RQA",
          "number": 1085,
          "title": "See What You Are Told: Visual Attention Sink in Large Multimodal Models",
          "normalized_title": "see what you are told visual attention sink in large multimodal models",
          "abstract": "Large multimodal models (LMMs) \"see\" images by leveraging the attention mechanism between text and visual tokens in the transformer decoder. Ideally, these models should focus on key visual information relevant to the text token. However, recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens, even when these tokens are irrelevant to the corresponding text. In this study, we investigate the property behind the appearance of these irrelevant visual tokens and examine their characteristics. Our findings show that this behavior arises due to the massive activation of certain hidden state dimensions, which resembles the attention sink found in language models. Hence, we refer to this phenomenon as the visual attention sink. In particular, our analysis reveals that removing the irrelevant visual sink tokens does not impact model performance, despite receiving high attention weights. Consequently, we recycle the attention to these tokens as surplus resources, redistributing the attention budget to enhance focus on the image. To achieve this, we introduce Visual Attention Redistribution (VAR), a method that redistributes attention in image-centric heads, which we identify as innately focusing on visual information. VAR can be seamlessly applied across different LMMs to improve performance on a wide range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks, all without the need for additional training, models, or inference steps. Experimental results demonstrate that VAR enables LMMs to process visual information more effectively by adjusting their internal attention mechanisms, offering a new direction to enhancing the multimodal capabilities of LMMs.",
          "authors": [
            "Seil Kang",
            "Jinyeong Kim",
            "Junhyeok Kim",
            "Seong Jae Hwang"
          ],
          "decision": null,
          "pdate": 1737562217273,
          "cdate": 1726481950949
        }
      ]
    }
  },
  {
    "paper_id": 45401,
    "title": "Boosting Adversarial Transferability with Spatial Adversarial Alignment",
    "abstract": "Deep neural networks are vulnerable to adversarial examples that exhibit transferability across various models. Numerous approaches are proposed to enhance the transferability of adversarial examples, including advanced optimization, data augmentation, and model modifications. However, these methods still show limited transferability, particularly in cross-architecture scenarios, such as from CNN to ViT. To achieve high transferability, we propose a technique termed Spatial Adversarial Alignment (SAA), which employs an alignment loss and leverages a witness model to fine-tune the surrogate model. Specifically, SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. First, we minimize the divergences of features between the two models in both global and local regions, facilitating spatial alignment. Second, we introduce a self-adversarial strategy that leverages adversarial examples to impose further constraints, aligning features from an adversarial perspective. Through this alignment, the surrogate model is trained to concentrate on the common features extracted by the witness model. This facilitates adversarial attacks on these shared features, thereby yielding perturbations that exhibit enhanced transferability. Extensive experiments on various architectures on ImageNet show that aligned surrogate models based on SAA can provide higher transferable adversarial examples, especially in cross-architecture attacks.",
    "introduction": "Deep neural networks (DNNs) have been successfully and extensively deployed across security-sensitive applications, including autonomous driving[46,51,50], facial verification[53,52,34], and video surveillance[17,54,18,14]. However, DNNs exhibit considerable vulnerability to adversarial examples[13,27,12,3,4,6,5], where imperceptible perturbations are introduced into natural images, leading models to produce incorrect predictions. In real-world applications, DNNs are typically concealed from user access, necessitating adversaries to generate adversarial examples within a black-box setting, where no knowledge of the target model’s parameters or architecture is available. Adversarial transferability plays a crucial role in black-box settings as it allows adversaries to effectively compromise target models by employing adversarial examples generated on surrogate models.\nIn black-box settings, adversarial transferability plays a crucial role, which enables adversaries to leverage adversarial examples crafted on surrogate models to effectively attack target models. Thus, generating highly transferable adversarial examples is instrumental in uncovering and understanding the vulnerabilities within DNNs, drawing substantial attention in recent research.\n\n\nCross-model transferability has been extensively studied for CNNs[9,45,10]. Highly transferable adversarial examples are usually based on advanced optimization[9,22,37]and data augmentation[45,10,25]. The principle is to alleviate the overfitting of adversarial examples on surrogate models, determining whether the attack can be successfully transferred to the target models. In addition, some model modification methods[44,15], such as amplifying the gradient on skip connections (the structure in ResNet[16]), can also improve transferability. However, few works explore adversarial transferability on Vision Transformer (ViT)[11]and the performance of existing work extending CNN to ViT is poor due to significant structural differences. Specifically, ViT flattens the image into a sequence of patch tokens and employs multi-head self-attention to capture global relationships among the patches. In contrast, CNNs typically consist of stacked convolutional layers that learn feature relationships progressively through downsampling. Therefore,[43]first empirically analyzes the structure of ViT and propose PNA and PatchOut[43], but there is still much room for improvement in cross-architecture transferability.\n\n\nIn this paper, we argue that unique structural features are critical to cross-architecture adversarial transferability.\nGiven a dataset, various models tend to exhibit analogous decision boundaries[23], arising from their ability to learn similar features. If we can obtain a surrogate model whose features are similar to those of models with different architectures, then the resulting adversarial perturbation can be transferable across different models. A recent technique known as Model Alignment (MA)[26]employs an alignment loss to minimize prediction divergences between surrogate models and witness models, thereby indirectly facilitating the extraction of features that are similarly represented by the witness model. However, directly applying MA to black-box attacks may lead to the degradation of cross-architecture transferability. The main reasons are:(i)Features are not aligned in space. MA only uses the final prediction of the model, but in fact, the spatial features of ViT and CNN are different. It is difficult to directly constrain the similarity of features only by the final logits.(ii)Features are not aligned from the perspective of adversarial features. In addition to the features of clean images, the features of adversarial examples also have similarities across different models and need to be considered.\n\n\nTo overcome these challenges and enhance transferability, we propose a technique called Spatial Adversarial Alignment (SAA), which utilizes an alignment loss from the perspective of spatial and adversarial features and incorporates a witness model to refine the surrogate model. SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. In the spatial-aware alignment, in addition to aligning on the final global features, we also focus on the features of local regions. We make local features of CNNs by position to align ViTs’ embeddings at the same position. In the adversarial-aware alignment, we introduce a self-adversarial strategy, which constructs adversarial examples so that the model can learn the differences between different architectures in adversarial features, thereby enabling the model to further capture more common features. Aligned surrogate models by SAA provide promising adversarial transferability and can be seamlessly integrated with existing transfer attacks. Our contributions can be summarized as follows:•We reveal for the first time the importance of spatial and adversarial features for cross-architecture transferability.•We propose Spatial Adversarial Alignment (SAA), which leverages a witness model to fine-tune the surrogate model via spatial-aware and adversarial-aware alignment to generate highly transferable adversarial examples.•Experiments on 6 CNNs and 4 ViTs show that SAA has state-of-the-art adversarial transferability, especially in cross-model transferability. Compared with MA, on ResNet50, the transferability from CNN to ViT is improved by 25.5-39.1%.\n\n",
    "related_work": "2.1Transfer attacks on CNNsEarly transfer attacks are mainly conducted between CNNs, and the most popular methods were advanced optimization[9,22,37], data augmentation[45,10,25], and model modification[44,15,43].Advanced Optimization.[22]compare adversarial attacks to model training: better optimization methods can obtain models with better generalization, and therefore also generate adversarial examples with higher transferability. FGSM[13]is the earliest gradient-based transfer attack, which was then extended to I-FGSM[20]. The subsequent advance optimization further improves the transferability by introducing momentum[9,22,39,37]and smoothness[31].Data Augmentation.Data augmentation serves as an effective strategy to prevent model overfitting, achieving state-of-the-art performance in model generalization[48,8]. Building on this principle, numerous adversarial attacks incorporate various transformations to enhance adversarial transferability, including modifications in size[45], scale[22], mixup[40], and frequency domain[25]adjustments. This integration aims to mitigate the overfitting of adversarial examples to the surrogate model, thereby increasing their effectiveness across different models.Model Modification.According to certain characteristics of the model, modifying the parameters of the surrogate model or changing the forward or backward propagation can also improve the transferability. Skip Gradient Method (SGM)[44]using more gradients from the skip connections rather than the residual modules, allows one to craft adversarial examples with high transferability. Similarly, Linear Backpropagation (LinBP)[15]and Backward Propagation Attack (BPA)[41]concentrate on non-linear activations by modifying the ReLU derivatives to enhance attack transferability. Model Alignment (MA)[26]promotes alignment of model predictions through an alignment loss relative to a witness model, with the aim of capturing shared features across models. However, MA overlooks spatial and adversarial feature alignment across architectures, limiting its effectiveness. Unlike these methods, our SAA requires no modifications to the forward or backpropagation processes, enabling the efficient generation of highly transferable adversarial examples with minimal training overhead. In contrast, LinBP and BPA, involve altering backpropagation or even full model retraining, incurring significantly higher computational costs.\n\n\n2.2Transfer attacks on ViTsCurrent transfer attacks for ViTs largely adapt methods developed for CNNs. Pay No Attention (PNA)[43]method extends Skip Gradient Method (SGM) to ViTs by omitting the gradient computation of attention blocks during back-propagation, thereby enhancing adversarial transferability. PatchOut[43]strategy selects a random subset of image patches to compute the gradient at each iteration, functioning as an image transformation technique to increase transferability.\nThen, Self-Ensemble (SE)[29]approach employs the class token at each layer with a shared classification head to create an ensemble model, facilitating optimized perturbation; however, many ViTs, such as Visformer[2]and CaiT[36], lack sufficient class tokens to build this ensemble. Additionally, Token Refinement (TR)[29]module fine-tunes class tokens to further boost transferability.\nRecently, Token Gradient Regularization (TGR)[49]works from the perspective of variance reduction, stabilizing the gradient direction to prevent adversarial examples from getting stuck in poor local optima.\nDistinct from these approaches, SAA is the first method to specifically analyze architectural differences across models. By leveraging shared features between different architectures, SAA enables the creation of more generalized surrogate models that integrate seamlessly with optimization and data augmentation methods, ultimately achieving state-of-the-art transferability.\n\n",
    "origin_citations": [
      "Boosting adversarial attacks with momentum.",
      "Nesterov accelerated gradient and scale invariance for adversarial attacks.",
      "Boosting the transferability of adversarial attacks with global momentum initialization.",
      "Improving transferability of adversarial examples with input diversity.",
      "Evading defenses to transferable adversarial examples by translation-invariant attacks.",
      "Frequency domain model augmentation for adversarial attack.",
      "Skip connections matter: On the transferability of adversarial examples generated with resnets.",
      "Backpropagating linearly improves transferability of adversarial examples.",
      "Towards transferable adversarial attacks on vision transformers.",
      "Explaining and harnessing adversarial examples.",
      "Adversarial examples in the physical world.",
      "Enhancing the transferability of adversarial attacks through variance tuning.",
      "Boosting the transferability of adversarial attacks with reverse adversarial perturbation.",
      "mixup: Beyond empirical risk minimization.",
      "Improved regularization of convolutional neural networks with cutout.",
      "Admix: Enhancing the transferability of adversarial attacks.",
      "Rethinking the backward propagation for adversarial transferability.",
      "Improving adversarial transferability via model alignment.",
      "On improving adversarial transferability of vision transformers.",
      "Visformer: The vision-friendly transformer.",
      "Going deeper with image transformers.",
      "Transferable adversarial attacks on vision transformers with token gradient regularization."
    ],
    "citations": [
      919,
      42845,
      244597,
      6146,
      963,
      12902,
      130935,
      110080,
      6676,
      892,
      902,
      506334,
      459227,
      7191,
      8319,
      818516,
      49099,
      776735,
      5772,
      7849,
      8434,
      492985
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 84,
      "Publication_Potential": 78,
      "Final_Score": 82.19999999999999,
      "raw_text": "Content_Quality: 84\nPublication_Potential: 78\nFinal_Score: 82"
    },
    "related_work_scores": {
      "Content_Coherence": 83,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 78,
      "Final_Score": 85.1,
      "raw_text": "Content_Coherence: 83\nCitation_Quality: 92\nSynthesis_Analysis: 78\nFinal_Score: 85"
    },
    "overall_score": 83.94,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2025/Conference",
        "venue": "NeurIPS",
        "year": "2025",
        "openreview_id": "seCBUZYs5c",
        "forum": "seCBUZYs5c",
        "number": 27785,
        "title": "Boosting Adversarial Transferability with Spatial Adversarial Alignment",
        "normalized_title": "boosting adversarial transferability with spatial adversarial alignment",
        "abstract": "Deep neural networks are vulnerable to adversarial examples that exhibit transferability across various models. Numerous approaches are proposed to enhance the transferability of adversarial examples, including advanced optimization, data augmentation, and model modifications. However, these methods still show limited transferability, partiovovocularly in cross-architecture scenarios, such as from CNN to ViT. To achieve high transferability, we propose a technique termed Spatial Adversarial Alignment (SAA), which employs an alignment loss and leverages a witness model to fine-tune the surrogate model. Specifically, SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. First, we minimize the divergences of features between the two models in both global and local regions, facilitating spatial alignment. Second, we introduce a self-adversarial strategy that leverages adversarial examples to impose further constraints, aligning features from an adversarial perspective. Through this alignment, the surrogate model is trained to concentrate on the common features extracted by the witness model. This facilitates adversarial attacks on these shared features, thereby yielding perturbations that exhibit enhanced transferability. Extensive experiments on various architectures on ImageNet show that aligned surrogate models based on SAA can provide higher transferable adversarial examples, especially in cross-architecture attacks.",
        "authors": [
          "Zhaoyu Chen",
          "HaiJing Guo",
          "Kaixun Jiang",
          "Jiyuan Fu",
          "Xinyu Zhou",
          "Dingkang Yang",
          "Hao Tang",
          "Bo Li",
          "Wenqiang Zhang"
        ],
        "decision": null,
        "pdate": 1758217571067,
        "cdate": 1747037969298
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2025/Conference",
          "venue": "NeurIPS",
          "year": "2025",
          "openreview_id": "seCBUZYs5c",
          "forum": "seCBUZYs5c",
          "number": 27785,
          "title": "Boosting Adversarial Transferability with Spatial Adversarial Alignment",
          "normalized_title": "boosting adversarial transferability with spatial adversarial alignment",
          "abstract": "Deep neural networks are vulnerable to adversarial examples that exhibit transferability across various models. Numerous approaches are proposed to enhance the transferability of adversarial examples, including advanced optimization, data augmentation, and model modifications. However, these methods still show limited transferability, partiovovocularly in cross-architecture scenarios, such as from CNN to ViT. To achieve high transferability, we propose a technique termed Spatial Adversarial Alignment (SAA), which employs an alignment loss and leverages a witness model to fine-tune the surrogate model. Specifically, SAA consists of two key parts: spatial-aware alignment and adversarial-aware alignment. First, we minimize the divergences of features between the two models in both global and local regions, facilitating spatial alignment. Second, we introduce a self-adversarial strategy that leverages adversarial examples to impose further constraints, aligning features from an adversarial perspective. Through this alignment, the surrogate model is trained to concentrate on the common features extracted by the witness model. This facilitates adversarial attacks on these shared features, thereby yielding perturbations that exhibit enhanced transferability. Extensive experiments on various architectures on ImageNet show that aligned surrogate models based on SAA can provide higher transferable adversarial examples, especially in cross-architecture attacks.",
          "authors": [
            "Zhaoyu Chen",
            "HaiJing Guo",
            "Kaixun Jiang",
            "Jiyuan Fu",
            "Xinyu Zhou",
            "Dingkang Yang",
            "Hao Tang",
            "Bo Li",
            "Wenqiang Zhang"
          ],
          "decision": null,
          "pdate": 1758217571067,
          "cdate": 1747037969298
        }
      ]
    }
  },
  {
    "paper_id": 16561,
    "title": "Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization",
    "abstract": "In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available athttps://video-lavit.github.io.",
    "introduction": "Recently, the significant breakthrough of Large Language Models (LLMs)(Brown et al.,2020; Touvron et al.,2023a)has brought a surge in building general-purpose multimodal AI assistants(OpenAI,2023b; Gemini Team,2023)that can follow both textual and visual instructions. Drawing on the remarkable reasoning abilities of LLMs and knowledge in massive alignment corpus (e.g., image-text pairs), they showcase the great potential of accurately comprehending and generating visual content(Sun et al.,2024; Jin et al.,2024; Dong et al.,2024). Despite their success, these multimodal LLMs(Alayrac et al.,2022; Liu et al.,2023c)predominantly concentrate on the image-text data, leaving the adaptation for video modality less explored. In contrast to static images, video serves as a dynamic media form that is more in line with human visual perception. Learning effectively from video is particularly essential for enhancing machine intelligence to comprehend the real world.\n\n\nFigure 1:The key observation in this work is: most video parts have a high degree of temporal redundancy that may be described by motion vectors. By exploiting these motion vectors, the video can be efficiently tokenized for pre-training of multimodal LLMs.\n\n\nTo this end, several approaches have made attempts at harnessing the generative capabilities of LLMs for handling video data. Inheriting the successful paradigm from the image domain, they represent video as a sequence of visual tokens that aligns with LLMs’ semantic space by utilizing a pre-trained 2D image model(Li et al.,2023d; Zhang et al.,2023)or a 3D video backbone(Kondratyuk et al.,2023). Nevertheless, the existing designs are still not competent for effectively encoding videos. Compared to images, videos pose unique challenges associated with higher demands for learning complex spatiotemporal clues, such as time-varying actions and scene changes. In this regard, encoding individual video frames separately by the 2D visual encoder falls short of capturing the temporal motion information, which plays a vital role in identifying distinct behaviors and events within the video content. Although the recent concurrent work VideoPoet(Kondratyuk et al.,2023)crafts a 3D video tokenizer for video generation with LLM, its applicability is constrained to short video clips due to the use of long token sequences (e.g., 1280 tokens for a 2.2s clip).When it comes to understanding or generating long videos, inputting excessive numbers of tokens into LLMs is deemed unacceptable in terms of computational resources.\n\n\nThis work addresses the limitation in video-language pre-training by exploring an efficient video representation that decomposes video into keyframes and temporal motions.\nOur motivation is built upon the natural characteristics of video data itself. As illustrated inFigure1, a video is typically divided into several shots, where video frames within each shot often exhibit substantial information redundancy. It is superfluous to encode all of these frames as tokens and incorporate them into the generative pre-training of LLMs. This fact strongly spurs us to decompose each video into alternating keyframes and motion vectors, where the former encapsulate the primary visual semantics and the latter depict the dynamic evolution of its corresponding keyframe over time. There are several benefits to such decomposed representation: (1) Compared to processing consecutive video frames utilizing 3D encoders, the combination of a single keyframe and motion vectors requires fewer tokens to represent video temporal dynamics, which is more efficient for large-scale pre-training. (2) The model can inherit the acquired visual knowledge from an off-the-shelf image-only LLM and focus solely on modeling temporal information without learning from scratch.\n\n\nBased on the above motivations, we presentVideo-LaVIT(Language-VIsionTransformer), a new multimodal pre-training approach that effectively empowers LLMs to comprehend and generate video content in a unified framework. Specifically, Video-LaVIT incorporates two core components: atokenizerand adetokenizerto handle video modality. The video tokenizer aims to transform the continuous video data into a sequence of compact discrete tokens akin to a foreign language, where the keyframes are processed by utilizing an established image tokenizer(Jin et al.,2024). For converting the temporal motions into the compatible discrete format, a spatiotemporal motion encoder is devised. It can capture the time-varying contextual information contained in extracted motion vectors, thereby significantly enhancing LLMs’ ability to comprehend the intricate actions in video. The video detokenizer is responsible for mapping the discretized video token generated by LLMs back into its original continuous pixel space. During training, video is represented as an alternating discrete visual-motion token sequence, and thus can be optimized under the same next-token prediction objective together with different modalities. Since video is inherently a time series, this joint autoregressive pre-training contributes to learning the sequential relationships of different video clips. We found that Video-LaVIT, is capable of serving as a multimodal generalist to achieve promising results in both understanding and generation tasks without further fine-tuning. The key contributions of this work are summarized as:\n\n\n•We introduce Video-LaVIT, a multimodal pre-training method that pushes the limit of LLMs’ unified understanding and generation capability towards video.•To efficiently model visual and temporal information in video, Video-LaVIT incorporates a novel video tokenizer and detokenizer that operates on the decomposed representations of keyframes and motion vectors.•Experiments on 13 multimodal benchmarks demonstrate that Video-LaVIT achieves very competitive performance, ranging from image and video comprehension to zero-shot text-to-image and text-to-video generation.\n\n",
    "related_work": "Vision-language pre-training. Following the success of using large-scale image-text pairs for contrastive learning of vision-language models(Radford et al.,2021), a similar idea has been exploited in generative pre-training, where visual and language data are jointly modeled under an autoregressive process. In practice, this is typically achieved by adapting visual image inputs to pre-trained LLMs(Raffel et al.,2020; Brown et al.,2020; Touvron et al.,2023a)via an intermediate module like cross-attention(Alayrac et al.,2022), Q-Former(Li et al.,2023c), or linear projection(Liu et al.,2023c). More recent approaches such as CM3Leon(Yu et al.,2023a)and LaVIT(Jin et al.,2024)advocate the use of discrete visual tokenizers(van den Oord et al.,2017; Esser et al.,2021)to form a unified next token prediction objective. However, these methods are primarily focused on image-text data and cannot be directly extended to videos due to the significantly higher computational cost.\n\n\nFigure 2:For each video-text pair, Video-LaVIT decomposes the video into keyframes and motion vectors for efficient tokenization. The tokenizers are learned by maximally reconstructing original inputs (e.g., the motion tokenizer is shown on the right). Finally, the encoded tokens are concatenated with text tokens to form a multimodal sequence, allowing for unified generative pre-training of the LLM (left).\n\n\nVideo understanding and generation. By unifying videos in the above pre-training framework, remarkable progress has been made in video comprehension with masked(Yang et al.,2022)and autoregressive language models(Li et al.,2023d; Zhang et al.,2023; Maaz et al.,2023). However, for video generation, the mainstream approaches are still based on diffusion models(Sohl-Dickstein et al.,2015; Song & Ermon,2019; Ho et al.,2020), which enhance existing image pre-trained models with better temporal consistency(Ho et al.,2022; Singer et al.,2023; Blattmann et al.,2023b; Esser et al.,2023; Blattmann et al.,2023a). Language model based counterparts(Yan et al.,2021; Hong et al.,2023; Kondratyuk et al.,2023), on the other hand, face the critical challenge of efficiently encoding video temporal dynamics with limited context windows and computational resources. In response, our work leverages motion vectors, a classic and effective cue in video modeling(Zhang et al.,2016; Wang et al.,2023b; Shen et al.,2024), for improving the efficacy of LLM-based video comprehension and generation.\n\n",
    "origin_citations": [
      "Learning transferable visual models from natural language supervision.",
      "Exploring the limits of transfer learning with a unified text-to-text transformer.",
      "Language models are few-shot learners.",
      "LLaMA: Open and efficient foundation language models.",
      "Flamingo: A visual language model for few-shot learning.",
      "BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.",
      "LLaVA-Plus: Learning to use tools for creating multimodal agents.",
      "Scaling autoregressive multi-modal models: Pretraining and instruction tuning.",
      "Unified language-vision pretraining in LLM with dynamic discrete visual tokenization.",
      "Neural discrete representation learning.",
      "Taming transformers for high-resolution image synthesis.",
      "Zero-shot video question answering via frozen bidirectional language models.",
      "VideoChat: Chat-centric video understanding.",
      "Video-LLaMA: An instruction-tuned audio-visual language model for video understanding.",
      "Video-ChatGPT: Towards detailed video understanding via large vision and language models.",
      "Deep unsupervised learning using nonequilibrium thermodynamics.",
      "Generative modeling by estimating gradients of the data distribution.",
      "Denoising diffusion probabilistic models.",
      "Video diffusion models.",
      "Make-A-Video: Text-to-video generation without text-video data.",
      "Align your latents: High-resolution video synthesis with latent diffusion models.",
      "Structure and content-guided video synthesis with diffusion models.",
      "Stable video diffusion: Scaling latent video diffusion models to large datasets.",
      "VideoGPT: Video generation using VQ-VAE and transformers.",
      "CogVideo: Large-scale pretraining for text-to-video generation via transformers.",
      "VideoPoet: A large language model for zero-shot video generation.",
      "Real-time action recognition with enhanced motion vector CNNs.",
      "VideoComposer: Compositional video synthesis with motion controllability.",
      "Decouple content and motion for conditional image-to-video generation."
    ],
    "citations": [
      1639,
      9,
      679,
      1552,
      7565,
      212324,
      294565,
      129685,
      128753,
      3018,
      4765,
      82318,
      2242,
      2234,
      305761,
      5918,
      5921,
      4827,
      5927,
      176197,
      54397,
      490753,
      23455,
      789395,
      145593,
      347735,
      504110,
      307529,
      23823
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 85,
      "Final_Score": 82.9,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 85\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 87,
      "Synthesis_Analysis": 84,
      "Final_Score": 84.60000000000001,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 87\nSynthesis_Analysis: 84\nFinal_Score: 85"
    },
    "overall_score": 83.92000000000002,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "S9lk6dk4LL",
        "forum": "S9lk6dk4LL",
        "number": 2513,
        "title": "Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization",
        "normalized_title": "video lavit unified video language pre training with decoupled visual motional tokenization",
        "abstract": "In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.",
        "authors": [
          "Yang Jin",
          "Zhicheng Sun",
          "Kun Xu",
          "Kun Xu",
          "Liwei Chen",
          "Hao Jiang",
          "Quzhe Huang",
          "Chengru Song",
          "Yuliang Liu",
          "Di ZHANG",
          "Yang Song",
          "Kun Gai",
          "Yadong MU"
        ],
        "decision": null,
        "pdate": 1714610313709,
        "cdate": 1706467398754
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "S9lk6dk4LL",
          "forum": "S9lk6dk4LL",
          "number": 2513,
          "title": "Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization",
          "normalized_title": "video lavit unified video language pre training with decoupled visual motional tokenization",
          "abstract": "In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.",
          "authors": [
            "Yang Jin",
            "Zhicheng Sun",
            "Kun Xu",
            "Kun Xu",
            "Liwei Chen",
            "Hao Jiang",
            "Quzhe Huang",
            "Chengru Song",
            "Yuliang Liu",
            "Di ZHANG",
            "Yang Song",
            "Kun Gai",
            "Yadong MU"
          ],
          "decision": null,
          "pdate": 1714610313709,
          "cdate": 1706467398754
        }
      ]
    }
  },
  {
    "paper_id": 27968,
    "title": "MVGamba: Unify 3D Content Generation as State Space Sequence Modeling",
    "abstract": "Recent 3D large reconstruction models (LRMs) can generate high-quality 3D content in sub-seconds by integrating multi-view diffusion models with scalable multi-view reconstructors. Current works further leverage 3D Gaussian Splatting as 3D representation for improved visual quality and rendering efficiency. However, we observe that existing Gaussian reconstruction models often suffer from multi-view inconsistency and blurred textures. We attribute this to the compromise of multi-view information propagation in favor of adopting powerful yet computationally intensive architectures (e.g., Transformers).\nTo address this issue, we introduce MVGamba, a general and lightweight Gaussian reconstruction model featuring a multi-view Gaussian reconstructor based on the RNN-like State Space Model (SSM). Our Gaussian reconstructor propagates causal context containing multi-view information for cross-view self-refinement while generating a long sequence of Gaussians for fine-detail modeling with linear complexity.\nWith off-the-shelf multi-view diffusion models integrated, MVGamba unifies 3D generation tasks from a single image, sparse images, or text prompts. Extensive experiments demonstrate that MVGamba outperforms state-of-the-art baselines in all 3D content generation scenarios with approximately only0.1×0.1\\times0.1 ×of the model size. The codes are available athttps://github.com/SkyworkAI/MVGamba.",
    "introduction": "We address the challenge of crafting 3D content from a single image, sparse-view images, or text input, which can facilitate a broad range of applications,e.g., Virtual Reality, immersive filming, digital gaming and animation. Previous research on 3D generation has investigated distilling 2D diffusion priors into 3D representations via score distillation sampling (SDS)[1]. Although these optimization-based approaches exhibit strong zero-shot generation capability with high-fidelity rendering quality[2,3,4,5], they are extremely time- and memory-intensive, often requiring hours to produce a single 3D asset, thus not practical for a real-world scenario.\n\n\nWith the advent of large-scale open-world 3D datasets[6,7,8], recent 3D large reconstruction models (LRMs)[9,10,11,12]integrate multi-view diffusion models[13,14,15]with scalable multi-view 3D reconstructor to regress a certain 3D representation (e.g.Triplane-NeRF[16,17], mesh) in a feed-forward manner. Specifically, current LRMs[18,19,20]adopt aone image (or text)→→\\rightarrow→mulit-view images→→\\rightarrow→3Ddiagram to predict 3D Gaussian Splatting (3DGS)[21]parameters, thereby ensuring the rendering efficiency while preserving fine details. Given a single image or text prompt, they first generate a set of images using multi-view diffusion models, which are then fed into a multi-view reconstructor (e.g., U-Net[22]or Transformer[23]), mapping image tokens to 3D Gaussians with superior generation speed and unprecedented quality.\n\n\nHowever, we observe that existing feed-forward Gaussian reconstruction models typically adopt powerful yet computationally intensive architectures[23,24]to generate long sequences of Gaussians for intricate 3D modeling. Such approaches inevitably compromise the integrity of multi-view information propagation to manage computational costs. For instance, they use local[18]or mixed[19]attention on limited multi-view image tokens or even deal each view separately and simply merge the predicted Gaussians afterwards[20]. Consequently, the generated 3D models often suffer from multi-view inconsistency and blurred textures, as illustrated in Figure2(a). These issues indicate that current compromise strategies fail to translate into coherent, high-quality outputs in practice. This raises a crucial question:How can we preserve the integrity of multi-view information while efficiently generating a sufficiently long sequence of Gaussians?\n\n\nTo address this issue, in this paper, we introduceMulti-ViewGaussianMamba (MVGamba), a general and lightweight Gaussian reconstruction model. At its core, MVGamba features a multi-view Gaussian reconstructor based on the recently introduced RNN-like architecture Mamba[25], which expands the given multi-view images into a long sequence of 3D Gaussian tokens and processes them recurrently in a causal manner.\nBy adopting causal context propagation, our approach efficiently maintains multi-view information integrity and further enables cross-view self-refinement from earlier to current views.\nAdditionally, our Gaussian reconstructor enables the fine-detailed generation of long Gaussian sequences with linear complexity[26,27]in a single forward process, eliminating the need for any post hoc operations used in previous work.\n\n\nMore concretely, we first patchify the multi-view images intoN𝑁Nitalic_Ntokens and rearrange them according to a cross-scan order[27,28], resulting in4×N4𝑁4\\times N4 × italic_Nimage tokens for selective scanning. These tokens are then processed through a series of Mamba blocks for state space sequence modeling. Subsequently, we feed the output Gaussian sequence into a lightweight Multi-Layer Perceptron (MLP) for channel-wise knowledge selection, followed by a set of linear decoders to obtain the Gaussian parameters representing high-quality 3D content (Sec.3.2). Compared to previous LRMs[29,11,30], our MVGamba features many computationally efficient components: a single-layer 2D convolution image tokenizer replaces the pre-trained DINO[31]transformer encoder, a lightweight MLP combined with linear decoders replaces the deep MLP decoder, and most importantly, linear complexity Mamba blocks replace quadratic complexity Transformer blocks (Figure2(b)). Together, these designs ensure efficient training and inference while achieving higher generation quality (Sec.4). Moreover, to directly convert the generated Gaussians into smooth textured polygonal meshes, we alternatively incorporate a 3DGS variant — 2DGS[32]— for accurate geometric modeling and mesh extraction.\n\n\nFigure 2:(a)Previous Gaussian reconstruction models sacrifice the integrity of multi-view information for computationally intensive architectures, resulting in multi-view inconsistency and blurred textures.(b)Comparison of FLOPs between self-attention in Transformers and SSM in Mamba. Detailed FLOPs data are provided in Table3.\n\n\nWe conducted comprehensive qualitative and quantitative experiments to verify the efficacy of our proposed MVGamba. The experimental results demonstrate that MVGamba (49M parameters) outperforms other latest LRMs[19,29,33]and even optimization-based methods[34,35]on the task of text-to-3D generation, single-view reconstruction and sparse-view reconstruction with roughly only0.1×0.1\\times0.1 ×of the model size.\nThe contributions and novelties of our paper are summarized as follows:•We point out that directly generating a sufficiently long sequence of Gaussians with full multi-view information is crucial for consistent and fine-detailed 3D generation.•We introduce MVGamba, a novel feed-forward pipeline that incorporates causal context propagation for cross-view self-refinement, allowing the efficient generation of long sequences of 2D/3D Gaussians for high-quality 3D content modeling.•Extensive experiments demonstrate that MVGamba is a potentiallygeneralsolution for 3D content generation, including text-to-3D, image-to-3D and sparse-view reconstruction task.\n\n",
    "related_work": "3D Generation.Previous approaches for generating high-fidelity 3D models predominantly used SDS-based optimization techniques[36,1]and their variants[37,5,3,4,34,2]. These methods yield high-quality 3D generations but require hours for the per-instance optimization process to converge. Pioneered by the large reconstruction model (LRM)[29], recent works[38,30,11,10]show that image tokens can be directly mapped to 3D representations, typically triplane-NeRF, in a feed-forward manner via a scalable transformer-based architecture[23]with large-scale 3D training data[6,7,8]. Among them, Instant3D[11]integrates LRM with multi-view image diffusion models[39,15,13,14,40], using four generated images for better quality. To avoid inefficient volume rendering and limited triplane resolution, some concurrent works[19,20,18]follow Instant3D and introduce 3D Gaussian Splatting[21]into sparse-view LRM variants. Specifically, GRM[18]and GS-LRM[20]use pixel-aligned Gaussian with a pure transformer-based reconstruction model, increasing the number of Gaussians through image feature upsampling and per-pixel merge operations. LGM[19]combines the 3D Gaussians from different views using a convolution-based asymmetric U-Net[22]. Our MVGamba, on the other hand, directly processes multi-view conditions causally, recurrently generating a long sequence of Gaussians for coherent and high-fidelity 3D modeling.\n\n\nMamba model for visual applications.Recent advancements in State Space Models (SSMs)[17,41,42], notably Mamba[25], have gained prominence in long sequence modeling for harmonizing computational efficiency and model versatility[43,44,45,46]. Following Mamba’s progress, there has been a surge in applying this framework to critical vision domains, including generic vision backbones[26,27,47,48], multi-modal streams[49,50], and vertical applications, especially in medical image processing[51,52,53,54,55,56]. Specifically, VMamba[27]pioneers a purely Mamba-based backbone to handle intensive prediction tasks. Similarly, Vim[26]leverages bidirectional SSMs for data-dependent global visual context without image-specific biases. Subsequent works progress with advanced selective scanning algorithms[47,48], integration with other networks[57,58], and adapted structural designs[59,60,59]. Concurrently, Gamba[61]marries Mamba with 3DGS for single-view reconstruction with limited texture quality and generalization capacity. In this paper, we explore and demonstrate the efficiency and long-sequence modeling capacity of Mamba in various 3D generation tasks with large-scale pre-training.\n\n",
    "origin_citations": [
      "Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022.",
      "Dreamfusion: Text-to-3d using 2d diffusion.",
      "Dreamtime: An improved optimization strategy for text-to-3d content creation.",
      "Diffusion time-step curriculum for one image to 3d generation.",
      "Consistent3d: Towards consistent high-fidelity text-to-3d generation with deterministic sampling prior.",
      "Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation.",
      "Dreamgaussian: Generative gaussian splatting for efficient 3d content creation.",
      "Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior.",
      "Lrm: Large reconstruction model for single image to 3d.",
      "Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction.",
      "Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model.",
      "Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.",
      "Meshlrm: Large reconstruction model for high-quality mesh.",
      "Attention is all you need.",
      "Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation.",
      "Objaverse: A universe of annotated 3d objects.",
      "Mvimgnet: A large-scale dataset of multi-view images.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Zero123++: a single image to consistent multi-view diffusion base model.",
      "Imagedream: Image-prompt multi-view diffusion for 3d generation.",
      "Mvdream: Multi-view diffusion for 3d generation.",
      "Direct2. 5: Diverse text-to-3d generation via multi-view 2.5 d diffusion.",
      "Lgm: Large multi-view gaussian model for high-resolution 3d content creation.",
      "Gs-lrm: Large reconstruction model for 3d gaussian splatting.",
      "Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "U-net: Convolutional networks for biomedical image segmentation.",
      "Efficient geometry-aware 3d generative adversarial networks.",
      "Combining recurrent, convolutional, and continuous-time models with linear state space layers.",
      "On the parameterization and initialization of diagonal state space models.",
      "Mamba: Linear-time sequence modeling with selective state spaces.",
      "Mambaout: Do we really need mamba for vision?",
      "A survey on vision mamba: Models, applications and challenges.",
      "Vision mamba: A comprehensive survey and taxonomy.",
      "Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges.",
      "Vision mamba: Efficient visual representation learning with bidirectional state space model.",
      "Vmamba: Visual state space model.",
      "Localmamba: Visual state space model with windowed selective scan.",
      "Plainmamba: Improving non-hierarchical mamba in visual recognition.",
      "Mambatalk: Efficient holistic gesture synthesis with selective state space models.",
      "Vl-mamba: Exploring state space models for multimodal learning.",
      "U-mamba: Enhancing long-range dependency for biomedical image segmentation.",
      "H-vmunet: High-order vision mamba unet for medical image segmentation.",
      "Weak-mamba-unet: Visual mamba makes cnn and vit work better for scribble-based medical image segmentation.",
      "Fusion-mamba for cross-modality object detection.",
      "Changemamba: Remote sensing change detection with spatio-temporal state space model.",
      "Pan-mamba: Effective pan-sharpening with state space model.",
      "Cobra: Extending mamba to multi-modal large language model for efficient inference.",
      "Jamba: A hybrid transformer-mamba language model.",
      "Res-vmamba: Fine-grained food category visual classification using selective state space models with deep residual learning.",
      "Mambaad: Exploring state space models for multi-class unsupervised anomaly detection.",
      "Gamba: Marry gaussian splatting with mamba for single view 3d reconstruction."
    ],
    "citations": [
      176110,
      49207,
      448108,
      126708,
      258271,
      477674,
      312653,
      295035,
      24846,
      294218,
      135724,
      38,
      214429,
      923081,
      76585,
      495077,
      313538,
      274286,
      1087300,
      22999,
      99387,
      903158,
      425520,
      180969,
      825,
      884251,
      512258,
      80609,
      274507,
      1046552,
      1089390,
      893218,
      133932,
      126520,
      126226,
      287742,
      423937,
      287675,
      426052,
      484281,
      426032,
      466975,
      445858,
      448637,
      466234,
      425574,
      422836,
      464358,
      447071,
      423346
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 78,
      "Final_Score": 80.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 78\nFinal_Score: 81"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 95,
      "Synthesis_Analysis": 75,
      "Final_Score": 86.0,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 95\nSynthesis_Analysis: 75\nFinal_Score: 86"
    },
    "overall_score": 83.92,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "AprsVxrwXT",
        "forum": "AprsVxrwXT",
        "number": 2178,
        "title": "MVGamba: Unify 3D Content Generation as State Space Sequence Modeling",
        "normalized_title": "mvgamba unify 3d content generation as state space sequence modeling",
        "abstract": "Recent 3D large reconstruction models (LRMs) can generate high-quality 3D content in sub-seconds by integrating multi-view diffusion models with scalable multi-view reconstructors. Current works further leverage 3D Gaussian Splatting as 3D representation for improved visual quality and rendering efficiency. However, we observe that existing Gaussian reconstruction models often suffer from multi-view inconsistency and blurred textures. We attribute this to the compromise of multi-view information propagation in favor of adopting powerful yet computationally intensive architectures (\\eg, Transformers). \nTo address this issue, we introduce MVGamba, a general and lightweight Gaussian reconstruction model featuring a multi-view Gaussian reconstructor based on the RNN-like State Space Model (SSM). Our Gaussian reconstructor propagates causal context containing multi-view information for cross-view self-refinement while generating a long sequence of Gaussians for fine-detail modeling with linear complexity.\nWith off-the-shelf multi-view diffusion models integrated, MVGamba unifies 3D generation tasks from a single image, sparse images, or text prompts. Extensive experiments demonstrate that MVGamba outperforms state-of-the-art baselines in all 3D content generation scenarios with approximately only $0.1\\times$ of the model size. The codes are available at \\url{https://github.com/SkyworkAI/MVGamba}.",
        "authors": [
          "Xuanyu Yi",
          "Zike Wu",
          "Qiuhong Shen",
          "Qingshan Xu",
          "Pan Zhou",
          "Joo Hwee Lim",
          "Shuicheng YAN",
          "Xinchao Wang",
          "Hanwang Zhang"
        ],
        "decision": null,
        "pdate": 1727287682934,
        "cdate": 1714987622230
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "AprsVxrwXT",
          "forum": "AprsVxrwXT",
          "number": 2178,
          "title": "MVGamba: Unify 3D Content Generation as State Space Sequence Modeling",
          "normalized_title": "mvgamba unify 3d content generation as state space sequence modeling",
          "abstract": "Recent 3D large reconstruction models (LRMs) can generate high-quality 3D content in sub-seconds by integrating multi-view diffusion models with scalable multi-view reconstructors. Current works further leverage 3D Gaussian Splatting as 3D representation for improved visual quality and rendering efficiency. However, we observe that existing Gaussian reconstruction models often suffer from multi-view inconsistency and blurred textures. We attribute this to the compromise of multi-view information propagation in favor of adopting powerful yet computationally intensive architectures (\\eg, Transformers). \nTo address this issue, we introduce MVGamba, a general and lightweight Gaussian reconstruction model featuring a multi-view Gaussian reconstructor based on the RNN-like State Space Model (SSM). Our Gaussian reconstructor propagates causal context containing multi-view information for cross-view self-refinement while generating a long sequence of Gaussians for fine-detail modeling with linear complexity.\nWith off-the-shelf multi-view diffusion models integrated, MVGamba unifies 3D generation tasks from a single image, sparse images, or text prompts. Extensive experiments demonstrate that MVGamba outperforms state-of-the-art baselines in all 3D content generation scenarios with approximately only $0.1\\times$ of the model size. The codes are available at \\url{https://github.com/SkyworkAI/MVGamba}.",
          "authors": [
            "Xuanyu Yi",
            "Zike Wu",
            "Qiuhong Shen",
            "Qingshan Xu",
            "Pan Zhou",
            "Joo Hwee Lim",
            "Shuicheng YAN",
            "Xinchao Wang",
            "Hanwang Zhang"
          ],
          "decision": null,
          "pdate": 1727287682934,
          "cdate": 1714987622230
        }
      ]
    }
  },
  {
    "paper_id": 40237,
    "title": "Inference Scaling for Long-Context Retrieval Augmented Generation",
    "abstract": "The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge. However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring the combination of multiple strategies beyond simply increasing the quantity of knowledge, including in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs’ ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from thescaling of inference computationwhen optimally configured? (2) Can we predict the optimaltest-time compute allocationfor a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as theinference scaling laws for RAG. Building on this, we further develop thecomputation allocation modelto estimate RAG performance across different inference configurations. The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.",
    "introduction": "Long-context large language models (LLMs) are designed to handle extended input sequences, enabling them to process and understand longer context (e.g., Gemini 1.5 Pro with up to 2M tokens)(Achiam et al.,2023; Team et al.,2023; Reid et al.,2024). Combined with increased inference computation, long-context LLMs demonstrate improved performance across various downstream tasks(Agarwal et al.,; Snell et al.,2024). For example, many-shot in-context learning (ICL) can match the performance of supervised fine-tuning by providing extensive in-context examples(Bertsch et al.,2024). Particularly for knowledge-intensive tasks that leverage retrieval augmented generation (RAG), increasing the quantity or size of retrieved documents up to a certain threshold consistently enhances the performance(Ram et al.,2023; Xu et al.,2024; Jiang et al.,2024).\n\n\nFigure 1:Normalized performance vs. effective context lengths on MuSiQue. Each line represents a fixed configuration, scaled by adjusting the number of documents. Red dots and dash lines represent the optimal configurations and their fitting results. Standard RAG plateaus early at104superscript10410^{4}10 start_POSTSUPERSCRIPT 4 end_POSTSUPERSCRIPTtokens, in contrast, DRAG and IterDRAG show near-linear improvement as the effective context length grows.\n\n\nPrevious studies on inference scaling for RAG focus on expanding the retrieved knowledge by increasing the number or lengths of retrieved documents(Xu et al.,2024; Jiang et al.,2024; Shao et al.,2024). However, only emphasizing on the knowledge quantity without providing further guidance presents certain limitations. On one hand, current long-context LLMs still have limited ability to effectively locate relevant information in ultra-long sequences upon challenging tasks(Li et al.,2024; Kuratov et al.,2024). For instance, the optimal performance of long-context LLMs is often achieved without fully utilizing the maximum length(Agarwal et al.,). On the other hand, numerous studies show that retrieving over soft thresholds (e.g., top-10 documents) leads to a performance plateau and may even cause declines(Ram et al.,2023; Lee et al.,2024a; Kuratov et al.,2024). Such performance drops may be traced back to the increased noise within context, which causes distraction and adversely affects generation(Yoran et al.,2024; Zhang et al.,2024; Leng et al.,2024). As a result, inference scaling of long-context RAG remains challenging for existing methods.\n\n\nIn this work, we leverage a broader range of strategies to comprehensively explore how RAG benefits from the scaling of inference computation. A straightforward strategy isdemonstration-based RAG(DRAG), where multiple RAG examples are provided as demonstrations to utilize the long-context capabilities of LLMs(Brown et al.,2020). DRAG allows models to learn (in-context) how to locate relevant information and apply it to response generation111Different from in-context RAG that prepends documents / QA examples(Press et al.,2023; Ram et al.,2023), we leverage multiple examples comprising of documents, questions and answers to demonstrate the task.. Nevertheless, the quality of one-step retrieval varies across tasks and often fails to provide sufficient information. Inspired by iterative methods(Trivedi et al.,2023; Yoran et al.,2024), we developiterative demonstration-based RAG(IterDRAG). IterDRAG learns to decompose input queries into simpler sub-queries and answer them using interleaved retrieval. By iteratively retrieving and generating upon sub-queries, LLMs construct reasoning chains that bridge the compositionality gap for multi-hop queries. Together, these strategies provide additional flexibility in scaling inference computation for RAG, allowing long-context LLMs to more effectively address complex knowledge-intensive queries.\n\n\nBuilding on these strategies, we investigate multiple ways to scale up inference computation. Here, we measure computation by considering the total number of input tokens across all iterations, referred to as theeffective context length. In DRAG, scaling the effective context length can be done by increasing two inference parameters: the number of retrieved documents and in-context examples. In IterDRAG, test-time compute can be further extended by introducing additional generation steps. Since different combinations of inference parameters result in varied allocations of computational resources, our goal is to establish the relationship between RAGperformance, differentscalesandallocationsof inference computation. Through extensive experiments on benchmark QA datasets, we demonstrate an almost linear relationship between RAG performance and the scale of effective context length by combining both RAG strategies, as shown inFigure1(right). Moreover, our RAG strategies exhibit improved performance than merely scaling the number of documents, achieving state-of-the-art performance with the compact Gemini 1.5 Flash (See evaluation inFigure2).\n\n\nDrawing from our observations, we examine the relationship between RAG performance and inference computation, which we quantify as theinference scaling laws for RAG. These observed inference scaling laws reveal that RAG performance consistently improves with the expansion of the effective context length under optimal configurations. Consequently, we take a deeper dive into modeling RAG performance with respect to various inference computationallocations. Our goal is to predict the optimal set of inference parameters that maximize the performance across different RAG tasks. To achieve this, we quantitatively model the relationship between RAG performance and varying inference configurations with thecomputation allocation model for RAG. Using the estimated computation allocation model, the optimal configurations can be empirically determined and generalize well for various scenarios, thereby maximizing the utilization of the computation budget. We summarize our contributions as follows:•We systematically investigate inference scaling for long-context RAG, for which we introduce two scaling strategies, DRAG and IterDRAG, to effectively scale inference compute.•We comprehensively evaluate DRAG and IterDRAG, where they not only achieve state-of-the-art performance, but also exhibit superior scaling properties compared to solely increasing the quantity of documents.•Through extensive experiments on benchmark QA datasets, we demonstrate that when test-time compute is optimally allocated, long-context RAG performance can scale almost linearly with the increasing order of magnitude of the computation budget.•We quantitatively model the relationship between RAG performance and different inference parameters, deriving the computation allocation model. This model aligns closely with our experimental results and generalize well across scenarios, providing practical guidance for optimal computation allocation in long-context RAG.\n\n\nFigure 2:Evaluation accuracy of Gemini 1.5 Flash using different methods: zero-shot QA, many-shot QA, RAG (with an optimal number of documents), DRAG and IterDRAG on benchmark QA datasets. By scaling up inference compute (up to 5M tokens), DRAG consistently outperforms baselines, while IterDRAG improves upon DRAG through interleaving retrieval and iterative generation.\n\n",
    "related_work": "2.1Long-Context LLMsLong-context large language models (LLMs) are designed to utilize extensive context and thereby improve their generative capabilities. Early works in extending context lengths involve sparse / low-rank kernels to reduce memory requirements(Kitaev et al.,2019; Beltagy et al.,2020; Zaheer et al.,2020; Choromanski et al.,2020). In addition, recurrent and state space models (SSMs) are proposed as efficient substitutes for transformer-based models(Gu et al.,2021; Gu and Dao,2023; Peng et al.,2023a; Beck et al.,2024). For causal LLMs, extrapolation and interpolation methods have proven effective in expanding context window lengths(Press et al.,2021; Chen et al.,2023; Sun et al.,2023; Peng et al.,2023b). Recent advancements in efficient attention methods(Dao et al.,2022; Jacobs et al.,2023; Liu et al.,2023)further enable LLMs to train and infer upon input sequences comprising millions of tokens(Achiam et al.,2023; Team et al.,2023; Reid et al.,2024).\n\n\n2.2In-Context LearningIn-context learning (ICL) offers a computationally efficient approach to enhance model performance at inference time by conditioning on a few demonstrations of the task(Brown et al.,2020). To further improve ICL performance, existing works focuses on pretraining strategies that optimize the language models to learn in-context(Min et al.,2022; Wei et al.,2023; Gu et al.,2023). In addition, selective usage of few-shot examples are shown to be helpful for enhancing downstream task performance(Liu et al.,2022; Rubin et al.,2022; Wang et al.,2024). Notably, reformatting or finding optimal ordering of in-context examples also improves ICL performance effectiveness(Lu et al.,2022; Wu et al.,2023; Liu et al.,2024a). With the emergence of long-context LLMs(Achiam et al.,2023; Team et al.,2023; Reid et al.,2024), scaling the number of examples becomes possible in ICL(Li et al.,2023; Bertsch et al.,2024;Agarwal et al.,). For instance,Agarwal et al.show that many-shot ICL can mitigate pretraining biases within LLMs and thus improves ICL performance across various tasks.\n\n\n2.3Retrieval Augmented GenerationRetrieval augmented generation (RAG) improves language model performance by incorporating relevant knowledge from external sources(Lewis et al.,2020; Guu et al.,2020; Karpukhin et al.,2020). In contrast to naïve RAG, optimizing the retrieval stage can effectively enhance context relevance and improve generation performance(Ma et al.,2023; Trivedi et al.,2023; Jiang et al.,2023; Shi et al.,2024; Sarthi et al.,2024; Lin et al.,2024). An example is REPLUG, in whichShi et al. (2024)leverage LLM as supervision to learn a dense retriever model. In addition, encoding documents can increase knowledge retrieval and improve generation capabilities(Khandelwal et al.,2019; Izacard and Grave,2021; Borgeaud et al.,2022; Izacard et al.,2023). For instance,Izacard and Grave (2021)leverages fusion-in-decoder architecture to encode multiple question-passage pairs while maintaining the model efficiency. Alternatively, selectively utilizing knowledge from the documents improves the robustness of LLMs against irrelevant context(Yu et al.,2023; Yoran et al.,2024; Yan et al.,2024; Yue et al.,2024; Zhang et al.,2024). For example, RAFT proposes to train language models with negative documents to improve generation quality and relevance(Zhang et al.,2024). Concurrent to our work, long-document retrieval and datastore scaling are proposed to optimize RAG performance(Jiang et al.,2024; Shao et al.,2024). Despite such progress, inference scaling remains under-explored for long-context RAG methods. As such, we investigate how variations in inference computation impact RAG performance, with the goal of optimizing test-time compute allocation.\n\n",
    "origin_citations": [
      "Reformer: The efficient transformer.",
      "Longformer: The long-document transformer.",
      "Big bird: Transformers for longer sequences.",
      "Rethinking attention with performers.",
      "Efficiently modeling long sequences with structured state spaces.",
      "Mamba: Linear-time sequence modeling with selective state spaces.",
      "RWKV: Reinventing rnns for the transformer era.",
      "xLSTM: Extended long short-term memory.",
      "Train short, test long: Attention with linear biases enables input length extrapolation.",
      "Extending context window of large language models via positional interpolation.",
      "A length-extrapolatable transformer.",
      "Yarn: Efficient context window extension of large language models.",
      "Flashattention: Fast and memory-efficient exact attention with IO-awareness.",
      "DeepSpeed Ulysses: System optimizations for enabling training of extreme long sequence transformer models.",
      "Ring attention with blockwise transformers for near-infinite context.",
      "GPT-4 technical report.",
      "Gemini: a family of highly capable multimodal models.",
      "Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.",
      "Language models are few-shot learners.",
      "MetaICL: Learning to learn in context.",
      "Symbol tuning improves in-context learning in language models.",
      "Pre-training to learn in context.",
      "What makes good in-context examples for GPT-3?",
      "Learning to retrieve prompts for in-context learning.",
      "Large language models are latent variable models: Explaining and finding good demonstrations for in-context learning.",
      "Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity.",
      "Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering.",
      "In-context vectors: Making in context learning more effective and controllable through latent space steering.",
      "In-context learning with many demonstration examples.",
      "In-context learning with long-context models: An in-depth exploration.",
      "Many-shot in-context learning.",
      "Retrieval-augmented generation for knowledge-intensive NLP tasks.",
      "Retrieval augmented language model pre-training.",
      "Dense passage retrieval for open-domain question answering.",
      "Query rewriting in retrieval-augmented large language models.",
      "Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.",
      "Active retrieval augmented generation.",
      "REPLUG: Retrieval-augmented black-box language models.",
      "RAPTOR: Recursive abstractive processing for tree-organized retrieval.",
      "RA-DIT: Retrieval-augmented dual instruction tuning.",
      "Generalization through memorization: Nearest neighbor language models.",
      "Leveraging passage retrieval with generative models for open domain question answering.",
      "Improving language models by retrieving from trillions of tokens.",
      "Atlas: Few-shot learning with retrieval augmented language models.",
      "Chain-of-note: Enhancing robustness in retrieval-augmented language models.",
      "Making retrieval-augmented language models robust to irrelevant context.",
      "Corrective retrieval augmented generation.",
      "Evidence-driven retrieval augmented response generation for online misinformation.",
      "RAFT: Adapting language model to domain specific rag.",
      "LongRAG: Enhancing retrieval-augmented generation with long-context LLMs.",
      "Scaling retrieval-based language models with a trillion-token datastore."
    ],
    "citations": [
      794,
      7298,
      1499,
      8384,
      8835,
      274507,
      8027,
      875511,
      452715,
      47815,
      139677,
      962709,
      7562,
      478697,
      323667,
      9115,
      8028,
      5914,
      679,
      511436,
      8649,
      7136,
      2189,
      3376,
      212677,
      8470,
      139767,
      294031,
      489887,
      819712,
      7716,
      7225,
      429,
      386,
      379,
      7228,
      8351,
      377,
      452245,
      5579,
      365,
      5557,
      729378,
      415,
      419,
      425394,
      8356,
      729248,
      729737
    ],
    "extract_rate": 0.9607843137254902,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 90,
      "Final_Score": 87.9,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 90\nFinal_Score: 88"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 75,
      "Final_Score": 81.1,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 75\nFinal_Score: 81"
    },
    "overall_score": 83.82,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "FSjIrOm1vz",
        "forum": "FSjIrOm1vz",
        "number": 12199,
        "title": "Inference Scaling for Long-Context Retrieval Augmented Generation",
        "normalized_title": "inference scaling for long context retrieval augmented generation",
        "abstract": "The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge.  However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring the combination of multiple strategies beyond simply increasing the quantity of knowledge, including in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs’ ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from the scaling of inference computation when optimally configured? (2) Can we predict the optimal test-time compute allocation for a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as the inference scaling laws for RAG. Building on this,  we further develop the computation allocation model to estimate RAG performance across different inference configurations.  The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.",
        "authors": [
          "Zhenrui Yue",
          "Honglei Zhuang",
          "Aijun Bai",
          "Kai Hui",
          "Rolf Jagerman",
          "Hansi Zeng",
          "Zhen Qin",
          "Dong Wang",
          "Xuanhui Wang",
          "Michael Bendersky"
        ],
        "decision": null,
        "pdate": 1737562848889,
        "cdate": 1727473624297
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "FSjIrOm1vz",
          "forum": "FSjIrOm1vz",
          "number": 12199,
          "title": "Inference Scaling for Long-Context Retrieval Augmented Generation",
          "normalized_title": "inference scaling for long context retrieval augmented generation",
          "abstract": "The scaling of inference computation has unlocked the potential of long-context large language models (LLMs) across diverse settings. For knowledge-intensive tasks, the increased compute is often allocated to incorporate more external knowledge.  However, without effectively utilizing such knowledge, solely expanding context does not always enhance performance. In this work, we investigate inference scaling for retrieval augmented generation (RAG), exploring the combination of multiple strategies beyond simply increasing the quantity of knowledge, including in-context learning and iterative prompting. These strategies provide additional flexibility to scale test-time computation (e.g., by increasing retrieved documents or generation steps), thereby enhancing LLMs’ ability to effectively acquire and utilize contextual information. We address two key questions: (1) How does RAG performance benefit from the scaling of inference computation when optimally configured? (2) Can we predict the optimal test-time compute allocation for a given budget by modeling the relationship between RAG performance and inference parameters? Our observations reveal that increasing inference computation leads to nearly linear gains in RAG performance when optimally allocated, a relationship we describe as the inference scaling laws for RAG. Building on this,  we further develop the computation allocation model to estimate RAG performance across different inference configurations.  The model predicts optimal inference parameters under various computation constraints, which align closely with the experimental results. By applying these optimal configurations, we demonstrate that scaling inference compute on long-context LLMs achieves up to 58.9% gains on benchmark datasets compared to standard RAG.",
          "authors": [
            "Zhenrui Yue",
            "Honglei Zhuang",
            "Aijun Bai",
            "Kai Hui",
            "Rolf Jagerman",
            "Hansi Zeng",
            "Zhen Qin",
            "Dong Wang",
            "Xuanhui Wang",
            "Michael Bendersky"
          ],
          "decision": null,
          "pdate": 1737562848889,
          "cdate": 1727473624297
        }
      ]
    }
  },
  {
    "paper_id": 16954,
    "title": "Premier-TACO is a Few-Shot Policy Learner: Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss",
    "abstract": "We present Premier-TACO, a multitask feature representation learning approach designed to improve few-shot policy learning efficiency in sequential decision-making tasks. Premier-TACO leverages a subset of multitask offline datasets for pretraining ageneral feature representation, which captures critical environmental dynamics and is fine-tuned using minimal expert demonstrations. It advances the temporal action contrastive learning (TACO) objective, known for state-of-the-art results in visual control tasks, by incorporating a novel negative example sampling strategy. This strategy is crucial in significantly boosting TACO’s computational efficiency, making large-scale multitask offline pretraining feasible.\nOur extensive empirical evaluation in a diverse set of continuous control benchmarks including Deepmind Control Suite, MetaWorld, and LIBERO demonstrate Premier-TACO’s effectiveness in pretraining visual representations, significantly enhancing few-shot imitation learning of novel tasks.\nOur code, pretraining data, as well as pretrained model checkpoints will be released athttps://github.com/PremierTACO/premier-taco.",
    "introduction": "In the dynamic and ever-changing world we inhabit, the importance of sequential decision-making (SDM) in machine learning cannot be overstated.\nUnlike static tasks, sequential decisions reflect the fluidity of real-world scenarios, from robotic manipulations to evolving healthcare treatments.\nJust as foundation models in language, such as BERT(Devlin et al.,2019)and GPT(Radford et al.,2019; Brown et al.,2020), have revolutionized natural language processing by leveraging vast amounts of textual data to understand linguistic nuances,pretrained foundation modelshold similar promise for sequential decision-making (SDM).\nIn language, these models capture the essence of syntax, semantics, and context, serving as a robust starting point for a myriad of downstream tasks. Analogously, in SDM, where decisions are influenced by a complex interplay\n\n\nFigure 1:Performance of Premier-TACO pretrained visual representation for few-shot imitation learning on downstream unseen tasks from Deepmind Control Suite, MetaWorld, and LIBERO. LfS here representslearning from scratch.\n\n\nof past actions, current states, and future possibilities, a pretrained foundation model can provide a rich, generalized understanding of decision sequences. This foundational knowledge, built upon diverse decision-making scenarios, can then be fine-tuned to specific tasks, much like how language models are adapted to specific linguistic tasks.\n\n\nThe followingchallengesare unique to sequential decision-making, setting it apart from existing vision and language pretraining paradigms.(C1) Data Distribution Shift: Training data usually consists of specific behavior-policy-generated trajectories. This leads to vastly different data distributions at various stages—pretraining, finetuning, and deployment—resulting in compromised performance(Lee et al.,2021).(C2) Task Heterogeneity: Unlike language and vision tasks, which often share semantic features, decision-making tasks vary widely in configurations, transition dynamics, and state and action spaces. This makes it difficult to develop a universally applicable representation.(C3) Data Quality and Supervision: Effective representation learning often relies on high-quality data and expert guidance. However, these resources are either absent or too costly to obtain in many real-world decision-making tasks(Brohan et al.,2023; Stooke et al.,2021b).\nOuraspirational criteriafor foundation model for sequential decision-making encompass several key features:(W1) Versatilitythat allows the model to generalize across a wide array of tasks, even those not previously encountered, such as new embodiments viewed or observations from novel camera angles;(W2) Efficiencyin adapting to downstream tasks, requiring minimal data through few-shot learning techniques;(W3) Robustnessto pretraining data of fluctuating quality, ensuring a resilient foundation; and(W4) Compatibilitywith existing large pretrained models such asNair et al. (2022).\n\n\nIn light of these challenges and desirables in building foundation models for SDM, our approach to develop foundational models for sequential decision-making focuses on creating a universal and transferable encoder using a reward-free, dynamics based, temporal contrastive pretraining objective.\nThis encoder would be tailored to manage tasks with complex observation spaces, such as visual inputs.\nBy excluding reward signals during the pretraining stage, the model will be better poised to generalize across a broad array of downstream tasks that may have divergent objectives.\nLeveraging a world-model approach ensures that the encoder learns a compact representation that can capture universal transition dynamics, akin to the laws of physics, thereby making it adaptable for multiple scenarios.\nThis encoder enables the transfer of knowledge to downstream control tasks, even when such tasks were not part of the original pretraining data set.\n\n\nExisting works apply self-supervised pre-training from rich vision data such as ImageNet(Deng et al.,2009)or Ego4D datasets(Grauman et al.,2022)to build foundation models(Nair et al.,2022; Majumdar et al.,2023; Ma et al.,2023).\nHowever, applying these approaches to sequential decision-making tasks is challenging.\nSpecifically, they often overlook control-relevant considerations and suffer from a domain gap between pre-training datasets and downstream control tasks.\nIn this paper, rather than focusing on leveraging large vision datasets, we propose a novel control-centric objective function for pretraining.\nOur approach, calledPremier-TACO(pretrainingmultitask representation viatemporalaction-drivencontrastive loss), employs a temporal action-driven contrastive loss function for pretraining.\nThis control-centric objective learns a state representation by optimizing the mutual information between representations of current states paired with action sequences and representations of the corresponding future states.\n\n\nPremier-TACO markedly enhances the effectiveness and efficiency of the temporal action contrastive learning (TACO) objective, as detailed inZheng et al. (2023), which delivers state-of-the-art outcomes in visual control tasks within asingle-task setting. It extends these capabilities to efficient, large-scalemultitask offline pretraining, broadening its applicability and performance.\nSpecifically, while TACO considers every data point in a batch as a potential negative example, Premier-TACO strategically samples a single negative example from a proximate window of the subsequent state. This method ensures the negative example is visually akin to the positive one, necessitating that the latent representation captures control-relevant information, rather than relying on extraneous features like visual appearance. This efficient negative example sampling strategy adds no computational burden and is compatible with smaller batch sizes. In particular, on MetaWorld, using a batch size of1818\\frac{1}{8}divide start_ARG 1 end_ARG start_ARG 8 end_ARGfor TACO, Premier-TACO achieves a 25% relative performance improvement.\nPremier-TACO can be seamlessly scaled for multitask offline pretraining, enhancing its usability and effectiveness.\n\n\nBelow we list our key contributions:⊳contains-as-subgroup\\rhd⊳(1)We introduce Premier-TACO, a new framework designed for the multi-task offline visual representation pretraining of sequential decision-making problems. In particular, we develop a new temporal contrastive learning objective within the Premier-TACO framework. Compared with other temporal contrastive learning objectives such as TACO, Premier-TACO employs a simple yet efficient negative example sampling strategy, making it computationally feasible for multi-task representation learning.⊳contains-as-subgroup\\rhd⊳(2) [(W1) Versatility (W2) Efficiency]Through extensive empirical evaluation, we verify the effectiveness of Premier-TACO’s pretrained visual representations for few-shot learning on unseen tasks.\nOn MetaWorld(Yu et al.,2019)and LIBERO(Liu et al.,2023), with 5 expert trajectories, Premier-TACO outperforms the best baseline pretraining method by 37% and 17% respectively.\nRemarkably, in LIBERO, we are the first method to demonstrate benefits from pretraining.\nOn Deepmind Control Suite (DMC)(Tassa et al.,2018), using only 20 trajectories, which is considerably fewer demonstrations than(Sun et al.,2023; Majumdar et al.,2023), Premier-TACO achieves the best performance across 10 challenging tasks, including the hard Dog and Humanoid tasks.\nThis versatility extends even to unseen embodiments in DMC as well as unseen tasks with unseen camera views in MetaWorld.⊳contains-as-subgroup\\rhd⊳(3) [(W3) Robustness (W4) Compatability]Furthermore, we demonstrate that Premier-TACO is not only resilient to data of lower quality but also compatible with exisiting large pretrained models. In DMC, Premier-TACO works well with the pretraining dataset collected randomly.\nAdditionally, we showcase the capability of the temporal contrastive learning objective of Premier-TACO to finetune a generalized visual encoder such as R3M(Nair et al.,2022), resulting in an averaged performance enhancement of around 50% across the assessed tasks on Deepmind Control Suite and MetaWorld.\n\n",
    "related_work": "Existing works, including R3M(Nair et al.,2022), VIP(Ma et al.,2023), MVP(Xiao et al.,2022), PIE-G(Yuan et al.,2022), and VC-1(Majumdar et al.,2023), focus on self-supervised pre-training for building foundation models but struggle with the domain gap in sequential decision-making tasks.\nRecent studies, such as one byHansen et al. (2022a), indicate that models trained from scratch often outperform pre-trained representations.\nApproaches like SMART(Sun et al.,2023)and DualMind(Wei et al.,2023)offer control-centric pre-training, but at the cost of extensive fine-tuning or task sets.\nContrastive learning techniques like CURL(Laskin et al.,2020), CPC(Henaff,2020), ST-DIM(Anand et al.,2019), and ATC(Stooke et al.,2021a)have succeeded in visual RL, but mainly focus on high-level features and temporal dynamics without a holistic consideration of state-action interactions, a gap partially filled by TACO(Zheng et al.,2023).\nOur work builds upon these efforts but eliminates the need for extensive task sets and fine-tuning, efficiently capturing control-relevant features.\nThis positions our method as a distinct advancement over DRIML(Mazoure et al.,2020)and Homer(Misra et al.,2019), which require more computational or empirical resources.\n\n\nA detailed discussion of related work is in AppendixA.\n\n",
    "origin_citations": [
      "R3m: A universal visual representation for robot manipulation.",
      "VIP: Towards universal visual reward and representation via value-implicit pre-training.",
      "Masked visual pre-training for motor control, 2022.",
      "Pre-trained image encoder for generalizable visual reinforcement learning.",
      "Where are we in the search for an artificial visual cortex for embodied intelligence?, 2023.",
      "On pre-training for visuo-motor control: Revisiting a learning-from-scratch baseline.",
      "SMART: Self-supervised multi-task pretraining with control transformers.",
      "Is imitation all you need? generalized decision-making with dual-phase training.",
      "CURL: Contrastive unsupervised representations for reinforcement learning.",
      "Data-efficient image recognition with contrastive predictive coding.",
      "Unsupervised state representation learning in atari.",
      "Decoupling representation learning from reinforcement learning.",
      "TACO: Temporal latent action-driven contrastive loss for visual reinforcement learning.",
      "Deep reinforcement and infomax learning.",
      "Kinematic state abstraction and provably efficient rich-observation reinforcement learning."
    ],
    "citations": [
      2603,
      739554,
      2593,
      140396,
      847764,
      221536,
      213584,
      471054,
      2600,
      2508,
      439644,
      188073,
      48882,
      349976,
      450174
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 82,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 85\nSynthesis_Analysis: 82\nFinal_Score: 82"
    },
    "overall_score": 83.8,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "KSNl7VgeVr",
        "forum": "KSNl7VgeVr",
        "number": 6356,
        "title": "Premier-TACO is a Few-Shot Policy Learner: Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss",
        "normalized_title": "premier taco is a few shot policy learner pretraining multitask representation via temporal action driven contrastive loss",
        "abstract": "We present Premier-TACO, a multitask feature representation learning approach designed to improve few-shot policy learning efficiency in sequential decision-making tasks. Premier-TACO leverages a subset of multitask offline datasets for pretraining a general feature representation, which captures critical environmental dynamics and is fine-tuned using minimal expert demonstrations. It advances the temporal action contrastive learning (TACO) objective, known for state-of-the-art results in visual control tasks, by incorporating a novel negative example sampling strategy. This strategy is crucial in significantly boosting TACO’s computational efficiency, making large-scale multitask offline pretraining feasible. Our extensive empirical evaluation in a diverse set of continuous control benchmarks including Deepmind Control Suite, MetaWorld, and LIBERO demonstrate Premier-TACO’s effective- ness in pretraining visual representations, significantly enhancing few-shot imitation learning of novel tasks.",
        "authors": [
          "Ruijie Zheng",
          "Yongyuan Liang",
          "Xiyao Wang",
          "Shuang Ma",
          "Hal Daumé III",
          "Huazhe Xu",
          "John Langford",
          "Praveen Palanisamy",
          "Kalyan Shankar Basu",
          "Furong Huang"
        ],
        "decision": null,
        "pdate": 1714610413839,
        "cdate": 1706818839366
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "KSNl7VgeVr",
          "forum": "KSNl7VgeVr",
          "number": 6356,
          "title": "Premier-TACO is a Few-Shot Policy Learner: Pretraining Multitask Representation via Temporal Action-Driven Contrastive Loss",
          "normalized_title": "premier taco is a few shot policy learner pretraining multitask representation via temporal action driven contrastive loss",
          "abstract": "We present Premier-TACO, a multitask feature representation learning approach designed to improve few-shot policy learning efficiency in sequential decision-making tasks. Premier-TACO leverages a subset of multitask offline datasets for pretraining a general feature representation, which captures critical environmental dynamics and is fine-tuned using minimal expert demonstrations. It advances the temporal action contrastive learning (TACO) objective, known for state-of-the-art results in visual control tasks, by incorporating a novel negative example sampling strategy. This strategy is crucial in significantly boosting TACO’s computational efficiency, making large-scale multitask offline pretraining feasible. Our extensive empirical evaluation in a diverse set of continuous control benchmarks including Deepmind Control Suite, MetaWorld, and LIBERO demonstrate Premier-TACO’s effective- ness in pretraining visual representations, significantly enhancing few-shot imitation learning of novel tasks.",
          "authors": [
            "Ruijie Zheng",
            "Yongyuan Liang",
            "Xiyao Wang",
            "Shuang Ma",
            "Hal Daumé III",
            "Huazhe Xu",
            "John Langford",
            "Praveen Palanisamy",
            "Kalyan Shankar Basu",
            "Furong Huang"
          ],
          "decision": null,
          "pdate": 1714610413839,
          "cdate": 1706818839366
        }
      ]
    }
  },
  {
    "paper_id": 40018,
    "title": "Dynamic Diffusion Transformer",
    "abstract": "Diffusion Transformer (DiT), an emerging diffusion model for image generation, has demonstrated superior performance but suffers from substantial computational costs. Our investigations reveal that these costs stem from thestaticinference paradigm, which inevitably introduces redundant computation in certaindiffusion timestepsandspatial regions. To address this inefficiency, we proposeDynamicDiffusionTransformer (DyDiT), an architecture thatdynamicallyadjusts its computation along bothtimestepandspatialdimensions during generation. Specifically, we introduce aTimestep-wise Dynamic Width(TDW) approach that adapts model width conditioned on the generation timesteps. In addition, we design aSpatial-wise Dynamic Token(SDT) strategy to avoid redundant computation at unnecessary spatial locations. Extensive experiments on various datasets and different-sized models verify the superiority of DyDiT. Notably, with <3% additional fine-tuning iterations, our method reduces the FLOPs of DiT-XL by 51%, accelerates generation by 1.73×\\times×, and achieves a competitive FID score of 2.07 on ImageNet. The code is publicly available athttps://github.com/NUS-HPC-AI-Lab/Dynamic-Diffusion-Transformer.",
    "introduction": "Diffusion models(Ho et al.,2020; Dhariwal & Nichol,2021; Rombach et al.,2022; Blattmann et al.,2023)have demonstrated significant superiority in visual generation tasks. Recently, the remarkable scalability of Transformers(Vaswani et al.,2017; Dosovitskiy et al.,2020)has led to the growing prominence of Diffusion Transformer (DiT)(Peebles & Xie,2023). DiT has shown strong potential across a variety of generation tasks(Chen et al.,2023; Ma et al.,2024; Chen et al.,2024)and is considered a foundational component in the development of Sora(Brooks et al.,2024), a pioneering model for video generation. Like Transformers in other vision and language domains(Dosovitskiy et al.,2020; Brown et al.,2020), DiT faces significant efficiency challenges during generation.\n\n\nExisting approaches to improving DiT’s efficiency include efficient diffusion samplers(Song et al.,2020a;2023; Salimans & Ho,2022; Meng et al.,2023; Luo et al.,2023)and global acceleration techniques(Ma et al.,2023; Pan et al.,2024). In addition, reducing computational redundancy within the DiT architecture using model compression techniques, such as structural pruning(Fang et al.,2024; Molchanov et al.,2016; He et al.,2017), also shows significant promise.\n\n\nHowever, pruning methods typically retain astaticarchitecture across both thetimestepandspatialdimensions throughout the diffusion process. As shown in Figure1(c), both the original DiT and the pruned DiT employ a fixed model width across all diffusion timesteps and allocate the same computational cost to every image patch. This static inference paradigm overlooks the varying complexities associated with different timesteps and spatial regions, leading to significant computational inefficiency. To explore this redundancy in more detail, we analyze the training process of DiT, during which it is optimized for a noise prediction task. Our analysis yields two key insights:\n\n\na) Timestep perspective:We plot the loss value differences between a pre-trained small model (DiT-S) and a larger model (DiT-XL) in Figure1(a). The results show that the loss differences diminish substantially fort>t^𝑡^𝑡t>\\hat{t}italic_t > over^ start_ARG italic_t end_ARG, and even approach negligible levels ast𝑡titalic_tnears the prior distribution (t→T→𝑡𝑇t\\to Titalic_t → italic_T). This indicates that the prediction task becomesprogressively easier at later timestepsand could be managed effectively even bya smaller model. However, DiT applies the same architecture across all timesteps, leading toexcessive computational costs at timesteps where the task complexity is low.\n\n\nb) Spatial perspective:We visualize the loss maps in Figure1(b) and observe a noticeable imbalance in loss values across different spatial regions of the image. Loss values are higher in patches corresponding to the main object, while patches representing background regions exhibit relatively lower loss. This suggests that the difficulty of noise prediction varies across spatial regions. Consequently,uniform computational treatment of all patches introduces redundancy and is likely suboptimal.\n\n\nBased on the above insights, a promising approach to improve DiT’s computational efficiency isdynamic computation. To this end, we proposeDynamicDiffusionTransformer (DyDiT), which adaptively allocates computational resources during the generation process, as illustrated in Figure1(c). Specifically, from the timestep perspective, we introduce aTimestep-wise Dynamic Width(TDW) mechanism, where the model learns to adjust the width of the attention and MLP blocks based on the currenttimestep. From a spatial perspective, we develop aSpatial-wise Dynamic Token(SDT) strategy, which identifies image patches where noise prediction is relatively “easy”, allowing them to bypass computationally intensive blocks, thus reducing unnecessary computation.\n\n\nNotably, both TWD and SDT are plug-and-play modules that can be easily implemented on DiT to build DyDiT. Moreover, our method contributes to significant speedup due to the hardware-friendly design: 1) the model architecture at each timestep can be pre-determined offline, eliminating additional overhead for width adjustments and enabling efficient batch processing (Section3.2); and 2) the token gathering and scattering operations incur minimal overhead and are straightforward to implement (Section3.3). Such hardware efficiency distinguishes our approach from traditional dynamic networks(Herrmann et al.,2020; Meng et al.,2022; Han et al.,2023b), which adapt their inference graphs for each sample and struggle to improve practical efficiency in batched inference.\n\n\nWe conduct extensive experiments across multiple datasets and model scales to validate the effectiveness of the proposed method. For example, compared to the static counterpart DiT-XL, our DyDiT-XL reduces FLOPs by 51% and accelerates the generation by 1.73 times, with less than 3% fine-tuning iterations, while maintaining a competitive FID score of 2.07 on ImageNet (256×\\times×256)(Deng et al.,2009). Our method shows potential for further efficiency gains when combined with efficient samplers, such as DDIM(Song et al.,2020a)and DPM Solver++(Lu et al.,2022), or global acceleration techniques like DeepCache(Ma et al.,2023). We anticipate that DyDiT will inspire future research in the development of more efficient diffusion Transformers.\n\n\nFigure 1:(a) The loss difference between DiT-S and DiT-XL across all diffusion timesteps (T=1000𝑇1000T=1000italic_T = 1000). The difference is slight at most timesteps. (b) Loss maps (normalized to the range [0, 1]) at different timesteps, show that the noise in different patches has varying levels of difficulty to predict. (c) Difference of the inference paradigm between the static DiT and the proposed DyDiT.\n\n",
    "related_work": "Efficient Diffusion Models.Although diffusion models(Ho et al.,2020; Rombach et al.,2022)have achieved remarkable performance in generation tasks, their generation speed has always hindered their further applications primarily due to long sampling steps and high computational costs. Existing attempts to make diffusion models efficient can be roughly categorized into sampler-based methods, model-based methods, and global acceleration methods. The sampler-based methods(Song et al.,2020a;2023; Salimans & Ho,2022; Meng et al.,2023; Luo et al.,2023)aim to reduce the sampling steps. Model-based approaches(Fang et al.,2024; So et al.,2024; Shang et al.,2023; Yang et al.,2023)attempt to compress the size of diffusion models via pruning(Fang et al.,2024; Shang et al.,2023)or quantization(Li et al.,2023; Shang et al.,2023). Global acceleration methods like Deepcache(Ma et al.,2023)tend to reuse or share some features across different timesteps.Our DyDiT is mostly relates to the model-based approaches and orthoganal to the other two lines of work. However, unlike the pruning methods yieldingstatic architectures, DyDiT performsdynamic computationfor different diffusion timesteps and image tokens.\n\n\nDynamic Neural Networks.Compared to static models, dynamic neural networks(Han et al.,2021)can adapt their computational graph based on inputs, enabling superior trade-off between performance and efficiency. They generally realize dynamic architectures by varying the network depth(Teerapittayanon et al.,2016; Bolukbasi et al.,2017; Yang et al.,2020; Han et al.,2022;2023a)or width(Herrmann et al.,2020; Li et al.,2021; Han et al.,2023b)during inference. Some works explore the spatial redundancy in imagerecognition(Wang et al.,2021; Song et al.,2021; Rao et al.,2021; Liang et al.,2022; Meng et al.,2022). Despite their promising theoretical efficiency, existing dynamic networks usually struggle in achievingpractical efficiencyduring batched inference(Han et al.,2023b)due to the per-sample inference graph. Moreover, the potential of dynamic architectures in diffusion models, where atimestepdimension is introduced, remains unexplored.This work extends the research of dynamic networks to the imagegenerationfield. More importantly, our TDW adjusts the network structure only conditioned on thetimesteps, avoiding the sample-conditioned tensor shapes in batched inference. Together with the efficient token gathering and scattering mechanism of SDT, DyDiT shows preferable realistic efficiency.\n\n",
    "origin_citations": [
      "Denoising diffusion probabilistic models.",
      "High-resolution image synthesis with latent diffusion models.",
      "Denoising diffusion implicit models.",
      "Consistency models.",
      "Progressive distillation for fast sampling of diffusion models.",
      "On distillation of guided diffusion models.",
      "Latent consistency models: Synthesizing high-resolution images with few-step inference.",
      "Structural pruning for diffusion models.",
      "Temporal dynamic quantization for diffusion models.",
      "Post-training quantization on diffusion models.",
      "Diffusion probabilistic model made slim.",
      "Q-diffusion: Quantizing diffusion models.",
      "Deepcache: Accelerating diffusion models for free.",
      "Dynamic neural networks: A survey.",
      "Branchynet: Fast inference via early exiting from deep neural networks.",
      "Adaptive neural networks for efficient inference.",
      "Resolution adaptive networks for efficient inference.",
      "Learning to weight samples for dynamic early-exiting networks.",
      "Dynamic perceiver for efficient visual recognition.",
      "Channel selection using gumbel softmax.",
      "Dynamic slimmable network.",
      "Latency-aware unified dynamic networks for efficient image recognition.",
      "Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition.",
      "Dynamic grained encoder for vision transformers.",
      "Dynamicvit: Efficient vision transformers with dynamic token sparsification.",
      "Not all patches are what you need: Expediting vision transformers via token reorganizations.",
      "Adavit: Adaptive vision transformers for efficient image recognition."
    ],
    "citations": [
      4827,
      2240,
      356595,
      5919,
      91065,
      460643,
      322408,
      261332,
      307359,
      242559,
      242893,
      490179,
      274524,
      5119,
      689,
      698,
      7266,
      179048,
      49752,
      708,
      704,
      403397,
      947826,
      302763,
      787,
      175575,
      704850
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 82,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 85\nSynthesis_Analysis: 82\nFinal_Score: 82"
    },
    "overall_score": 83.8,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "taHwqSrbrb",
        "forum": "taHwqSrbrb",
        "number": 2019,
        "title": "Dynamic Diffusion Transformer",
        "normalized_title": "dynamic diffusion transformer",
        "abstract": "Diffusion Transformer (DiT), an emerging diffusion model for image generation,\nhas demonstrated superior performance but suffers from substantial computational\ncosts. Our investigations reveal that these costs stem from the static inference\nparadigm, which inevitably introduces redundant computation in certain diffusion\ntimesteps and spatial regions. To address this inefficiency, we propose Dynamic\nDiffusion Transformer (DyDiT), an architecture that dynamically adjusts its compu-\ntation along both timestep and spatial dimensions during generation. Specifically,\nwe introduce a Timestep-wise Dynamic Width (TDW) approach that adapts model\nwidth conditioned on the generation timesteps. In addition, we design a Spatial-\nwise Dynamic Token (SDT) strategy to avoid redundant computation at unnecessary\nspatial locations. Extensive experiments on various datasets and different-sized\nmodels verify the superiority of DyDiT. Notably, with <3% additional fine-tuning it-\nerations, our method reduces the FLOPs of DiT-XL by 51%, accelerates generation\nby 1.73×, and achieves a competitive FID score of 2.07 on ImageNet.",
        "authors": [
          "Wangbo Zhao",
          "Yizeng Han",
          "Jiasheng Tang",
          "Kai Wang",
          "Yibing Song",
          "Gao Huang",
          "Fan Wang",
          "Yang You"
        ],
        "decision": null,
        "pdate": 1737562252803,
        "cdate": 1726801841337
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "taHwqSrbrb",
          "forum": "taHwqSrbrb",
          "number": 2019,
          "title": "Dynamic Diffusion Transformer",
          "normalized_title": "dynamic diffusion transformer",
          "abstract": "Diffusion Transformer (DiT), an emerging diffusion model for image generation,\nhas demonstrated superior performance but suffers from substantial computational\ncosts. Our investigations reveal that these costs stem from the static inference\nparadigm, which inevitably introduces redundant computation in certain diffusion\ntimesteps and spatial regions. To address this inefficiency, we propose Dynamic\nDiffusion Transformer (DyDiT), an architecture that dynamically adjusts its compu-\ntation along both timestep and spatial dimensions during generation. Specifically,\nwe introduce a Timestep-wise Dynamic Width (TDW) approach that adapts model\nwidth conditioned on the generation timesteps. In addition, we design a Spatial-\nwise Dynamic Token (SDT) strategy to avoid redundant computation at unnecessary\nspatial locations. Extensive experiments on various datasets and different-sized\nmodels verify the superiority of DyDiT. Notably, with <3% additional fine-tuning it-\nerations, our method reduces the FLOPs of DiT-XL by 51%, accelerates generation\nby 1.73×, and achieves a competitive FID score of 2.07 on ImageNet.",
          "authors": [
            "Wangbo Zhao",
            "Yizeng Han",
            "Jiasheng Tang",
            "Kai Wang",
            "Yibing Song",
            "Gao Huang",
            "Fan Wang",
            "Yang You"
          ],
          "decision": null,
          "pdate": 1737562252803,
          "cdate": 1726801841337
        }
      ]
    }
  },
  {
    "paper_id": 41847,
    "title": "Latent Action Pretraining From Videos",
    "abstract": "We introduce Latent Action Pretraining for general Action models (LAPA), an unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain alatentVLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation model. We open-source the model checkpoints and code atlatentactionpretraining.github.io.",
    "introduction": "Vision-Language-Action Models (VLA) for robotics(Brohan et al.,2023; Kim et al.,2024)are trained by aligning large language models with vision encoders, and then finetuning it on on diverse robot datasets(Collaboration et al.,2023); this enables generalization to novel instructions, unseen objects, and distribution shifts(Michał et al.,2024). However, diverse real-world robot datasets mostly require human teleoperation, which makes scaling difficult. Internet video data, on the other hand, offers abundant examples of human behavior and physical interactions at scale, presenting a promising approach to overcome the limitations of small, specialized robotic datasets(Yang et al.,2024c). However, it is challenging to learn from internet video data for two major challenges: first, much of the raw data on the web lacks explicit action labels; second, the data distribution from the web is fundamentally different from the embodiments and environments of typical robotic systems(McCarthy et al.,2024). We proposeLatentActionPretraining for GeneralAction Models (LAPA), an unsupervised approach to pretraining a robotic foundation model without the need for ground-truth robot action labels (Figure1).\n\n\nLAPA has two pretraining stages, followed by a fine-tuning stage to map the latent actions to real robot actions. In the first pretraining stage, we use a VQ-VAE-based objective(Van Den Oord et al.,2017)to learn quantized latent actions between raw image frames. Analogous to Byte Pair Encoding(Sennrich et al.,2016)used for language modeling, this can be seen as learning to tokenize atomic actions without requiring predefined action priors (e.g., end-effector positions, joint positions). In the second stage, we perform behavior cloning by pretraining a Vision-Language Model to predict latent actions derived from the first stage based on video observations and task descriptions.\nFinally, we fine-tune the model on a small-scale robot manipulation dataset with robot actions to learn the mapping from the latent actions to robot actions. In this work, we refer to both the proposed method and the resulting VLA models as LAPA.\n\n\nFigure 1:Problem Formulation.We investigate building a generalist robotic foundation model from human motion videos without action labels.\n\n\nWe measure performance on diverse manipulation videos, including existing robot video datasets (without utilizing ground-truth actions) and human manipulation datasets. Our results show that the proposed method significantly outperforms baseline methods of training manipulation policies from actionless videos, particularly in cross-environment and cross-embodiment scenarios. Furthermore, on real-world manipulation tasks, our method leads to a new monolithic VLA model, outperformingOpenVLA(Kim et al.,2024), the current state-of-the-art model Vision Language Action (VLA) model trained on a diverse mixture of datasets with ground-truth actions. These results demonstrate the effectiveness of learning unified quantized latent action representations across diverse robotic datasets featuring different embodiments (shown in Section5.2). We further demonstrate that LAPA remains effective even when pretrained ononlyhuman manipulation video, outperforming models pretrained on Bridgev2(Walke et al.,2023), one of the largest open-sourced robotic datasets. We observe that LAPA effectively captures environment-centric actions, including object and camera movements, which could be beneficial for downstream tasks like navigation or dynamic, non-quasistatic tasks. We expect that our method opens up the potential for building foundation models for robotics by pretraining on much larger web-scale video data.\n\n\nWe summarize our main contributions and findings below:•We propose Latent Action Pretraining for general Action models (LAPA), an unsupervised approach to pretraining a robotic foundation model to encode robotic skills from web-scale video data.•Experiments on simulation and real-world robot tasks show that our method not only significantly outperforms baseline methods for training robotic manipulation policies from actionless video, but also leads to a VLA model that outperforms the current state-of-the-art VLA model trained with ground-truth actions (by +6.22%), while achieving over 30x greater pretraining efficiency.•With LAPA, we qualitatively demonstrate that it is possible to use the learned world and action prediction models to simulate full trajectories in diverse environments, effectively building a neural simulation capable of performing closed-loop evaluations entirely through neural inference.\n\n",
    "related_work": "2.1Vision-Language-Action ModelsVision-Language Models (VLMs), which are trained on extensive internet-scale datasets encompassing text, image, and video, have demonstrated the ability to understand and generate both text and multimodal data(Liu et al.,2023; Team,2024; Liu et al.,2024; Abdin et al.,2024). Building on the generalization power of VLMs, recent work has introduced Vision-Language-Action Models (VLAs), which extend VLMs by fine-tuning them on robotic action data to improve physical grounding(Brohan et al.,2023; Kim et al.,2024; Team et al.,2024; Collaboration et al.,2023). Recently, incorporating auxiliary objectives, such as visual traces(Niu et al.,2024), language reasoning paths(Michał et al.,2024), or constructing a conversational-style instruction dataset using robot trajectory data(Li et al.,2024a)during VLA training has been shown to improve performance. However, these methods still heavily rely on labeled action data, which limits the scalability of developing general VLAs, since they will be bounded by the amount of robotic data made available through human teleoperation. In contrast, our approach only requires labeled actions for finetuning, significantly reducing the burden of requiring a vast amount of human teleoperated robot data.Training Robot Policies From VideosVideos contain rich information about dynamics and behavior which can be potentially beneficial for robot learning. However, most raw videos do not contain any action labels(McCarthy et al.,2024). One line of work attempts to leverage egocentric human videos(Grauman et al.,2022)to improve visual representations for visuomotor robotic tasks(Nair et al.,2022; Dasari et al.,2023). Some recent works explore the benefits of video generative models pretrained on human videos for downstream robotic tasks(Wu et al.,2024; Liang et al.,2024). Another line of work aims to learn useful information from human videos by learning from interactions(Zeng et al.,2024), affordances(Bahl et al.,2023; Kannan et al.,2023; Srirama et al.,2024; Shaw et al.,2023), or visual traces(Wen et al.,2023; Bharadhwaj et al.,2024b)extracted from human videos. However, all these works learn useful visual priors whereas our approach allows learning the mapping directly from perception to control during pretraining.Another line of work aims to learn robot manipulation policies by retargeting human motions to robot motions. These works rely on off-the-shelf models such as hand pose estimators(Wang et al.,2023; Zhu et al.,2024; Shaw et al.,2023; Bharadhwaj et al.,2023; Ye et al.,2023; Qin et al.,2022)or motion capture systems(Yang et al.,2024a)to retarget the human motions directly to robot motions. However, these works either learn only task-specific policies or require large in-domain perfectly aligned human-robot data and do not leverage the diverse, web-scale human motion data to train a single generalist policy.The last line of work assumes one has access to a small-scaled robotic dataset with action labels to train an inverse dynamic model (IDM), optical flow, or reinforcement learning models that predicts actions from future state rollouts generated by world models(Du et al.,2023; Ko et al.,2024; Yang et al.,2024b; Bharadhwaj et al.,2024a)trained on diverse web-scale human videos. Another related work(Baker et al.,2022)trains an IDM with a small amount of action-labeled data and uses the IDM to label large-scaled actionless web data and pretrain a visuomotor robot policy. We consider these methods as direct baselines to our proposed methods since they are scalable methods with the same experimental assumptions and learn both visual and control priors to train generalist robot policies.Latent ActionsPrevious works have employed latent actions across diverse scenarios. GENIE(Bruce et al.,2024)maps user inputs (ground-truth actions) to a latent space, allowing generative models to create interactive environments. We adopt a similar latent action model but apply it to label actionless data for training a monolithic VLA to solve robotic tasks. Similarly,Edwards et al. (2018)andSchmidt & Jiang (2024)use latent actions to pretrain and fine-tune policies for video games(Cobbe et al.,2019). In contrast, we focus on learning latent actions from real-world human motions for more complex, continuous robotic tasks. Unlike other work that leverages latent actions by converting ground-truth actions into latent to capture better multimodality and task semanticsLynch et al. (2020); Jiang et al. (2023); Lee et al. (2024); Mete et al. (2024), our approach derives latent actions directly from observations, not ground-truth actions.\n\n",
    "origin_citations": [
      "Visual instruction tuning.",
      "Chameleon: Mixed-modal early-fusion foundation models.",
      "World model on million-length video and language with ringattention.",
      "Phi-3 technical report: A highly capable language model locally on your phone.",
      "Rt-2: Vision-language-action models transfer web knowledge to robotic control.",
      "Openvla: An open-source vision-language-action model.",
      "Octo: An open-source generalist robot policy.",
      "Open x-embodiment: Robotic learning datasets and rt-x models.",
      "Llarva: Vision-action instruction tuning enhances robot learning.",
      "Robotic control via embodied chain-of-thought reasoning.",
      "Llara: Supercharging robot learning data for vision-language policy.",
      "Towards generalist robot learning from internet video: A survey.",
      "Ego4d: Around the world in 3,000 hours of egocentric video.",
      "R3m: A universal visual representation for robot manipulation.",
      "An unbiased look at datasets for visuo-motor pre-training.",
      "Unleashing large-scale video generative pre-training for visual robot manipulation.",
      "Dreamitate: Real-world visuomotor policy learning via video generation.",
      "Learning manipulation by predicting interaction.",
      "Affordances from human videos as a versatile representation for robotics.",
      "Deft: Dexterous fine-tuning for real-world hand policies.",
      "Hrp: Human affordances for robotic pre-training.",
      "Videodex: Learning dexterity from internet videos.",
      "Any-point trajectory modeling for policy learning.",
      "Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation.",
      "Mimicplay: Long-horizon imitation learning by watching human play.",
      "Vision-based manipulation from single human video with open-world object graphs.",
      "Zero-shot robot manipulation from passive human videos.",
      "Learning continuous grasping function with a dexterous hand from human demonstrations.",
      "Dexmv: Imitation learning for dexterous manipulation from human videos.",
      "Equibot: Sim (3)-equivariant diffusion policy for generalizable and data efficient learning.",
      "Learning universal policies via text-guided video generation.",
      "Learning to act from actionless videos through dense correspondences.",
      "Learning interactive real-world simulators.",
      "Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation.",
      "Video pretraining (vpt): Learning to act by watching unlabeled online videos.",
      "Genie: Generative interactive environments.",
      "Imitating latent policies from observation.",
      "Learning to act without actions.",
      "Leveraging procedural generation to benchmark reinforcement learning.",
      "Learning latent plans from play.",
      "H-gap: Humanoid control with a generalist planner.",
      "Behavior generation with latent actions.",
      "Quest: Self-supervised skill abstractions for learning continuous control."
    ],
    "citations": [
      2238,
      810882,
      1089662,
      134754,
      7672,
      947830,
      762427,
      320174,
      739412,
      859880,
      569585,
      762384,
      340823,
      2603,
      319966,
      348145,
      569582,
      54539,
      311207,
      762174,
      222128,
      346205,
      1088551,
      226227,
      996517,
      491249,
      13027,
      94238,
      911548,
      884906,
      320314,
      321520,
      1039257,
      7665,
      464547,
      74858,
      349237,
      5071,
      219976,
      273535,
      290939,
      1055105
    ],
    "extract_rate": 0.9767441860465116,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 82,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 85\nSynthesis_Analysis: 82\nFinal_Score: 82"
    },
    "overall_score": 83.8,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "VYOe2eBQeh",
        "forum": "VYOe2eBQeh",
        "number": 3433,
        "title": "Latent Action Pretraining from Videos",
        "normalized_title": "latent action pretraining from videos",
        "abstract": "We introduce Latent Action Pretraining for general Action models (LAPA), the first unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation models.",
        "authors": [
          "Seonghyeon Ye",
          "Joel Jang",
          "Byeongguk Jeon",
          "Se June Joo",
          "Jianwei Yang",
          "Baolin Peng",
          "Ajay Mandlekar",
          "Reuben Tan",
          "Yu-Wei Chao",
          "Bill Yuchen Lin",
          "Lars Liden",
          "Kimin Lee",
          "Jianfeng Gao",
          "Luke Zettlemoyer",
          "Dieter Fox",
          "Minjoon Seo"
        ],
        "decision": null,
        "pdate": 1737562317415,
        "cdate": 1727155185392
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "VYOe2eBQeh",
          "forum": "VYOe2eBQeh",
          "number": 3433,
          "title": "Latent Action Pretraining from Videos",
          "normalized_title": "latent action pretraining from videos",
          "abstract": "We introduce Latent Action Pretraining for general Action models (LAPA), the first unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation models.",
          "authors": [
            "Seonghyeon Ye",
            "Joel Jang",
            "Byeongguk Jeon",
            "Se June Joo",
            "Jianwei Yang",
            "Baolin Peng",
            "Ajay Mandlekar",
            "Reuben Tan",
            "Yu-Wei Chao",
            "Bill Yuchen Lin",
            "Lars Liden",
            "Kimin Lee",
            "Jianfeng Gao",
            "Luke Zettlemoyer",
            "Dieter Fox",
            "Minjoon Seo"
          ],
          "decision": null,
          "pdate": 1737562317415,
          "cdate": 1727155185392
        }
      ]
    }
  },
  {
    "paper_id": 42012,
    "title": "Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models",
    "abstract": "Content warning:This paper contains harmful images and texts!Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful\nactivation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared\nto text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs.TGAretrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show thatTGAnot only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks. Code is available111https://github.com/xsc1234/VLM_Safety_Transfer.",
    "introduction": "Vision-language alignment methods for Large Vision-Language Models (LVLMs) use a basic LLM, a lightweight vision encoder and projector to efficiently enable the LLM to understand visual input for various vision tasks with relatively low training costs(Liu et al.,2024c; Dai et al.,2023; Zhu et al.,2023). Recent studies indicate that the safety of LVLMs deserves attention(Liu et al.,2024a; Wang et al.,2023; Gong et al.,2023). Given that vision and language are aligned into a common space in LVLMs, the safety mechanism should be shared by both of them. However, this is not the case.\nWe find that compared to toxic text input, LVLMs are more vulnerable to toxic vision input. Existing studies on the safety of LVLMs(Zong et al.,2024; Wang et al.,2024)fall short of providing an essential explanation for the question:“Why can’t the safety mechanism for text be shared by vision after vision-language alignment?”. This is the core issue that this paper aims to address.\n\n\nFigure 1:(a): Hidden states at the specific local transformer layers in LVLMs play a crucial role in the successful activation of safety mechanism. (b): Current vision-language alignment methods cannot effectively align vision with its semantics in text modality at hidden states level. (c): Insufficient alignment at hidden states level shifts the semantics of image and misleads the layers for safety.\n\n\nSince the mainstream vision-language alignment methods for LVLMs (e.g., LLaVA(Liu et al.,2024c)) lack additional safety fine-tuning and the training data in this process contain few toxic samples(Wang et al.,2023), the safety mechanism of LVLMs is mostly inherited from the safety mechanism that has been established for text in their basic LLMs. So an intuitive view to explain the dilemma of LVLMs’ safety on vision is that the vision-language alignment cannot effectively transfer the safety mechanism for text in LLMs to vision. Based on this, this paper proposes a novel perspective calledCross-Modal Safety Mechanism Transferto rethink, explain and address the exacerbated vulnerability of LVLMs to toxic vision inputs compared to toxic text inputs. Cross-modal safety mechanism transfer means transferring the existing safety mechanism for text in LLMs to vision in vision-language alignment training without any additional safety fine-tuning on vision. Our experiments in §3reveal that current vision-language alignment methods fail to achieve an effective cross-modal safety mechanism transfer. LVLMs exhibit significantly different safety capabilities when handling toxicity with the same semantics but different modalities (as shown in Figure1).\n\n\nTo analyze the cause of this issue, which also provides an essential explanation for the vulnerability of LVLMs to toxic vision input, we first offer an insightful understanding of where and how the safety mechanism of LVLMs operates in §4.1. Specifically, we propose a novel method to locate the activation of safety mechanism in LVLMs and analyze the attention pattern on toxic tokens. We find that the hidden states at the specific local transformer layers play a crucial role in the successful activation of safety mechanism (Figure1a). Then our comparative analysis between text and vision modalities in §4.2reveals that current vision-language alignment methods in training LVLMs cannot effectively align the vision with the language at hidden states level (Figure1b). This misalignment causes the hidden states of the vision to shift from the hidden states of the corresponding text with the same semantics. When the hidden states of toxic image are input to the specific transformer layers responsible for safety mechanism activation, this shift prevents these layers from correctly capturing the semantics of images as they do for texts, thereby hindering their ability to accurately assess the toxicity in images (Figure1c). This is the key reason why the safety mechanism for text in basic LLMs cannot be transferred to vision in vision-language alignment training for LVLMs.\n\n\nTo address the above problem, we propose a novel vision-language alignment training method calledTGA, which uses retrieved texts related to the input vision to guide the projection of vision into the hidden states space in LLMs, thereby achieving alignment at the hidden-state level. Extensive experiments show thatTGAsuccessfully transfers the safety\nmechanism for text in basic LLMs to vision during vision-language alignment training for LVLMs without any safety fine-tuning on the visual modality. Besides, ourTGAmaintains general performance across various vision tasks compared with existing state-of-the-art (SoTA) LVLMs. This indicates ourTGAis a safe and effective vision-language alignment method for training LVLMs. The contributions of this paper are:\n\n\n∙∙\\bullet∙We propose a novel perspective calledCross-Modal Safety Mechanism Transfer, which aims to transfer the safety mechanism for text in LLMs to vision in vision-language alignment without any safety fine-tuning on the vision data. This rethinks, explains, and addresses the exacerbated vulnerability of LVLMs to toxic vision compared to toxic text.\n\n\n∙∙\\bullet∙We explain where and how safety mechanism in LVLMs operates, and reveal that current vision-language alignment methods fail to achieve effective cross-modal safety mechanism transfer because of insufficient alignment at hidden states level.\n\n\n∙∙\\bullet∙We propose a novel vision-alignment method calledTGAthat can not only transfer the safety mechanism against toxic text in basic LLMs to vision but also maintain the general performance on various vision tasks compared with existing SoTA LVLMs.\n\n",
    "related_work": "Vision-Language Alignment in LVLMs.Vision-language alignment in LVLMs equips basic LLMs with the ability to understand and process visual input by pre-training and instruction-tuning on large-scale text-image pairs such as works in LLaVA(Liu et al.,2024c), InstructBLip(Dai et al.,2023), Qwen-VL(Bai et al.,2023b), MiniGPT-4(Zhu et al.,2023), Flamingo(Awadalla et al.,2023), PaLM-E(Driess et al.,2023), etc. However, it remains unknown whether vision-language alignment can extend all the capabilities of LLM on text to vision. This paper unveils that the safety mechanism of LLMs on text cannot be transferred to vision in existing typical vision-language alignment methods. We give the essential explanation for the cause of this issue and the novel solution method.\n\n\nSafety of LVLMs.Recent studies have shown that LVLMs are still prone to generating toxic content when facing the toxic input (especially the toxic image), even when trained on carefully curated datasets containing few toxic samples(Wang et al.,2023; Liu et al.,2024a; Gong et al.,2023). Some studies try to address this problem by safety instruction-tuning on supervised toxic vision data(Wang et al.,2024; Zong et al.,2024). The limitations of previous studies are: (1) lacking the essential explanation for the phenomenon that safety mechanism cannot be shared by both vision and language that has been aligned. (2) collecting the multi-modal data for safety instruction-tuning is much more challenging and requires more significant human effort than only text, especially for the modalities such as audio, speech and video(Chakraborty et al.,2024). This paper gives a novel perspective that views the safety issue in LVLMs as a problem of transferring the safety mechanism for text in base LLMs to vision to rethink, explain and address the vulnerability of LVLMs to toxic vision.\n\n",
    "origin_citations": [
      "Visual instruction tuning.",
      "Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023.",
      "Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond.",
      "Minigpt-4: Enhancing vision-language understanding with advanced large language models.",
      "Openflamingo: An open-source framework for training large autoregressive vision-language models, 2023.",
      "Palm-e: An embodied multimodal language model, 2023.",
      "Tovilag: Your visual-language generative model is also an evildoer.",
      "A survey of attacks on large vision-language models: Resources, advances, and future trends.",
      "Figstep: Jailbreaking large vision-language models via typographic visual prompts.",
      "Cross-modality safety alignment.",
      "Safety fine-tuning at (almost) no cost: A baseline for vision large language models.",
      "Cross-modal safety alignment: Is textual unlearning all you need?"
    ],
    "citations": [
      2238,
      7564,
      2233,
      2243,
      182218,
      2963,
      270918,
      1038081,
      294490,
      768931,
      100986,
      1090542
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 70,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 90,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 70\nCitation_Quality: 85\nSynthesis_Analysis: 90\nFinal_Score: 83"
    },
    "overall_score": 83.8,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "45rvZkJbuX",
        "forum": "45rvZkJbuX",
        "number": 4056,
        "title": "Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models",
        "normalized_title": "cross modal safety mechanism transfer in large vision language models",
        "abstract": "Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that \\textbf{TGA} not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good). Code is in supplemental material and will be released on GitHub after acceptance.",
        "authors": [
          "Shicheng Xu",
          "Liang Pang",
          "Yunchang Zhu",
          "Huawei Shen",
          "Xueqi Cheng"
        ],
        "decision": null,
        "pdate": 1737562344372,
        "cdate": 1727229081424
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "45rvZkJbuX",
          "forum": "45rvZkJbuX",
          "number": 4056,
          "title": "Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models",
          "normalized_title": "cross modal safety mechanism transfer in large vision language models",
          "abstract": "Vision-language alignment in Large Vision-Language Models (LVLMs) successfully enables LLMs to understand visual input. However, we find that existing vision-language alignment methods fail to transfer the existing safety mechanism for text in LLMs to vision, which leads to vulnerabilities in toxic image. To explore the cause of this problem, we give the insightful explanation of where and how the safety mechanism of LVLMs operates and conduct comparative analysis between text and vision. We find that the hidden states at the specific transformer layers play a crucial role in the successful activation of safety mechanism, while the vision-language alignment at hidden states level in current methods is insufficient. This results in a semantic shift for input images compared to text in hidden states, therefore misleads the safety mechanism. To address this, we propose a novel Text-Guided vision-language Alignment method (TGA) for LVLMs. TGA retrieves the texts related to input vision and uses them to guide the projection of vision into the hidden states space in LLMs. Experiments show that \\textbf{TGA} not only successfully transfers the safety mechanism for text in basic LLMs to vision in vision-language alignment for LVLMs without any safety fine-tuning on the visual modality but also maintains the general performance on various vision tasks (Safe and Good). Code is in supplemental material and will be released on GitHub after acceptance.",
          "authors": [
            "Shicheng Xu",
            "Liang Pang",
            "Yunchang Zhu",
            "Huawei Shen",
            "Xueqi Cheng"
          ],
          "decision": null,
          "pdate": 1737562344372,
          "cdate": 1727229081424
        }
      ]
    }
  },
  {
    "paper_id": 16287,
    "title": "ALERT-Transformer: Bridging Asynchronous and Synchronous Machine Learning for Real-Time Event-based Spatio-Temporal Data",
    "abstract": "We seek to enable classic processing of continuous ultra-sparse spatiotemporal data generated by event-based sensors with dense machine learning models.\nWe propose a novel hybrid pipeline composed of asynchronous sensing and synchronous processing that combines several ideas: (1) an embedding based on PointNet models – the ALERT module – that can continuously integrate new and dismiss old events thanks to a leakage mechanism, (2) a flexible readout of the embedded data that allows to feed any downstream model with always up-to-date features at any sampling rate, (3) exploiting the input sparsity in a patch-based approach inspired by Vision Transformer to optimize the efficiency of the method.\nThese embeddings are then processed by a transformer model trained for object and gesture recognition.\nUsing this approach, we achieve performances at the state-of-the-art with a lower latency than competitors.\nWe also demonstrate that our asynchronous model can operate at any desired sampling rate.",
    "introduction": "Event-based sensors capture visual information in an event-driven, asynchronous manner(Finateu et al.,2020; Gallego et al.,2020).\nEfficiently exploiting their data has proven challenging as the vast majority of approaches published in the literature consist of either converting event-based data to dense representations, or deploying spiking neural networks (SNNs) on streams of events.\nThe former allows to exploit standard machine learning (ML) frameworks such asPyTorchandTensorflow, but does not leverage the inherent sparsity and other properties of event-based data(Gehrig et al.,2019).\nThe latter relies on SNNs, which are hard to train and usually exhibit lower accuracy than an equivalent dense neural network.\nFurthermore, while the neuromorphic community has argued in favor of their higher energy efficiency for decades, recent research and breakthroughs in edge AI accelerators indicate this is still an open question(Dampfhoffer et al.,2023; Garrett et al.,2023; Moosmann et al.,2023; Caccavella et al.,2023).\n\n\nNevertheless, considering the inherent advantages of event-based vision sensors, namely high dynamic range (HDR) and high temporal resolution – simultaneously, without any tradeoffs between the two –, we aim to find a way to leverage this sparse and low-latency data for real-world situations.\n\n\nFigure 1:Schematic representation of the system integrating our proposed asynchronous embedding module. The asynchronous part (left) processes all events as they come in an event driven manner, thus updating theF⁢e⁢a⁢t⁢u⁢r⁢e⁢s𝐹𝑒𝑎𝑡𝑢𝑟𝑒𝑠Featuresitalic_F italic_e italic_a italic_t italic_u italic_r italic_e italic_stensor continuously. The synchronous part (right) samples theF⁢e⁢a⁢t⁢u⁢r⁢e⁢s𝐹𝑒𝑎𝑡𝑢𝑟𝑒𝑠Featuresitalic_F italic_e italic_a italic_t italic_u italic_r italic_e italic_stensor on demand, thus allowing a seamless interface between asynchronous and synchronous processing.\n\n\nStandard ML relies on tensor-based processing.\nConverting the stream of events – represented as tuples of values (x𝑥xitalic_xandy𝑦yitalic_ypixel coordinates, polarities and timestamp) – to a multidimensional tensor is thus a crucial step.\nThe challenge involves (1) representing time in a reliable and continuous manner, allowing it to be processed similarly to the finite spatial and polarity dimensions, (2) continuously incorporating new events in the feature tensors which also requires forgetting previous events, (3) using limited computational resources to allow real-time processing.\nOur main contributions towards Event-Based ML are the following:\n\n\n•The ALERT module, an embedding based on PointNet which continuously integrates new events dismissing old ones via a leakage mechanism. This module introduces novel asynchronous embedding updates.•A flexible readout of the embedded data that can feed any downstream model with up-to-date features at different sampling rates, down to a per-event operation, allowing ultra-low latency decision making.•A patch-based approach inspired by Vision Transformer to exploit input sparsity and optimize efficiency.•A time encoding solution to represent continuous time as a bi-dimensional vector of bounded values, at the cost of negligible decrease of relative accuracy.•The ALERT-Transformer, a framework incorporating all of the above, which is trained on event-based data end-to-end. The model can then operate in synchronous regime for high accuracy on gesture recognition, or asynchronously for ultra-low latency.\n\n\nOur idea is that, even if the sensor outputs data at 1MHz or more(Finateu et al.,2020), the application/client side will not, and does not need to, run at this speed.\nFor instance, in an embedded product, such as a drone, that relies on event-based sensors for self-localization in space, this localization step would not operate at more than 100Hz(Kaufmann et al.,2023).\nTherefore, we consider that this hybrid (asynchronous-to-synchronous) conversion module would be operated as close to the sensor as possible, inside a 3D-stacked image sensor, as proposed in(Bouvier et al.,2021; Bonazzi et al.,2023).\n\n\n1.1Related Works1.1.1Event-Based Data RepresentationsEvent-based data is extremely sparse, with unpredictable sparsity patterns, making it unsuitable for raw processing with dense machine learning pipelines.\nAn extended approach consists in integrating events over fixed time windows, creating frame-like representations using methods such as histograms or event queues(Innocenti et al.,2021; Sabater et al.,2022; Maqueda et al.,2018; Zhu et al.,2018).\nTo maximize task accuracy from end-to-end,(Gehrig et al.,2019)proposed learning the kernel for convolving the stream of events into a discrete tensor, called “Event-Spike Tensor”.\nWhile the resulting networks perform better, this method still requires waiting to accumulate all events before starting processing and is thus not asynchronous.1.1.2PointNet ArchitectureThe PointNet(Qi et al.,2017a,b)is an architecture designed for processing three-dimensional point clouds.\nIt extracts features from individual points with a shared Multi-Layer Perceptron (MLP).\nIt then reduces the entire point cloud with a max-pooling operation into a single global feature vector.\nGiven the similarity in data representation, several works interpreted event streams as point clouds(Wang et al.,2019a; Chen et al.,2022; Zhao et al.,2021; Wang et al.,2019b).\nThese approaches mostly treat time as any other coordinate, making fast and efficient processing challenging.EventNet(Sekikawa et al.,2019)addresses this issue by processing time separately from other coordinates.\nThey propose a modification to them⁢a⁢x𝑚𝑎𝑥maxitalic_m italic_a italic_xoperator which becomes a recursive function capable of updating temporal and spatial information as new events arrive.\nOur solution, introduced in Section2, eliminates the need to modify them⁢a⁢x𝑚𝑎𝑥maxitalic_m italic_a italic_xoperator.\nThe structure of the PointNet remains unaltered and is thus entirely compatible with existing AI accelerators.1.1.3Event-Based TransformersOur work takes inspiration from PointBERT(Yu et al.,2022), which combines PointNets with a Transformer(Vaswani et al.,2017)for 3D classification.\nModels proposed by(Sabater et al.,2022,2023; Wang et al.,2022; Peng et al.,2023)generate frames of aggregated events using various strategies, and convert them into tokens with a patch-to-token strategy inspired by the Vision Transformer (ViT111In ViT, images are split into fixed-size patches, which are then linearly embedded to obtain a sequence of vectors.)(Dosovitskiy et al.,2021).\nThey leverage spatial sparsity by discarding input patches lacking sufficient events.\nIn(Peng et al.,2023), grouped convolutions are used to embed successive patches through time, each token thus handles information from different time steps.(Chen et al.,2022)attempts to input a reduced 3D event cloud to the Point Transformer(Zhao et al.,2021), a model originally designed for spatial 3D point clouds.\nFollowing a different approach,(Blegiers et al.,2023)employ a Video Transformer Network using event-frames.\nAll these approaches demonstrate high accuracy, motivating our choice of using Transformer models.\nHowever, most are not trained end-to-end, and none can process events asynchronously.1.1.4Asynchronous ProcessingBesides(Sekikawa et al.,2019), some works focus on asynchronous processing of event streams with standard network architectures.(Sironi et al.,2018)presented an architecture using local memory units shared by neighboring pixels.(Messikommer et al.,2020)introduced Asynchronous Sparse Convolutional Networks, a framework for converting models trained on synchronous image-like event representations into asynchronous models.\nEven though these solutions are built for event-driven processing, they add on complexity with respect to standard ML tools.\n\n",
    "related_work": "1.1.1Event-Based Data RepresentationsEvent-based data is extremely sparse, with unpredictable sparsity patterns, making it unsuitable for raw processing with dense machine learning pipelines.\nAn extended approach consists in integrating events over fixed time windows, creating frame-like representations using methods such as histograms or event queues(Innocenti et al.,2021; Sabater et al.,2022; Maqueda et al.,2018; Zhu et al.,2018).\nTo maximize task accuracy from end-to-end,(Gehrig et al.,2019)proposed learning the kernel for convolving the stream of events into a discrete tensor, called “Event-Spike Tensor”.\nWhile the resulting networks perform better, this method still requires waiting to accumulate all events before starting processing and is thus not asynchronous.\n\n\n1.1.2PointNet ArchitectureThe PointNet(Qi et al.,2017a,b)is an architecture designed for processing three-dimensional point clouds.\nIt extracts features from individual points with a shared Multi-Layer Perceptron (MLP).\nIt then reduces the entire point cloud with a max-pooling operation into a single global feature vector.\nGiven the similarity in data representation, several works interpreted event streams as point clouds(Wang et al.,2019a; Chen et al.,2022; Zhao et al.,2021; Wang et al.,2019b).\nThese approaches mostly treat time as any other coordinate, making fast and efficient processing challenging.EventNet(Sekikawa et al.,2019)addresses this issue by processing time separately from other coordinates.\nThey propose a modification to them⁢a⁢x𝑚𝑎𝑥maxitalic_m italic_a italic_xoperator which becomes a recursive function capable of updating temporal and spatial information as new events arrive.\nOur solution, introduced in Section2, eliminates the need to modify them⁢a⁢x𝑚𝑎𝑥maxitalic_m italic_a italic_xoperator.\nThe structure of the PointNet remains unaltered and is thus entirely compatible with existing AI accelerators.\n\n\n1.1.3Event-Based TransformersOur work takes inspiration from PointBERT(Yu et al.,2022), which combines PointNets with a Transformer(Vaswani et al.,2017)for 3D classification.\nModels proposed by(Sabater et al.,2022,2023; Wang et al.,2022; Peng et al.,2023)generate frames of aggregated events using various strategies, and convert them into tokens with a patch-to-token strategy inspired by the Vision Transformer (ViT111In ViT, images are split into fixed-size patches, which are then linearly embedded to obtain a sequence of vectors.)(Dosovitskiy et al.,2021).\nThey leverage spatial sparsity by discarding input patches lacking sufficient events.\nIn(Peng et al.,2023), grouped convolutions are used to embed successive patches through time, each token thus handles information from different time steps.(Chen et al.,2022)attempts to input a reduced 3D event cloud to the Point Transformer(Zhao et al.,2021), a model originally designed for spatial 3D point clouds.\nFollowing a different approach,(Blegiers et al.,2023)employ a Video Transformer Network using event-frames.\nAll these approaches demonstrate high accuracy, motivating our choice of using Transformer models.\nHowever, most are not trained end-to-end, and none can process events asynchronously.\n\n\n1.1.4Asynchronous ProcessingBesides(Sekikawa et al.,2019), some works focus on asynchronous processing of event streams with standard network architectures.(Sironi et al.,2018)presented an architecture using local memory units shared by neighboring pixels.(Messikommer et al.,2020)introduced Asynchronous Sparse Convolutional Networks, a framework for converting models trained on synchronous image-like event representations into asynchronous models.\nEven though these solutions are built for event-driven processing, they add on complexity with respect to standard ML tools.\n\n",
    "origin_citations": [
      "Temporal binary representation for event-based action recognition.",
      "Event transformer. a sparse-aware solution for efficient event data processing.",
      "Event-based vision meets deep learning on steering prediction for self-driving cars.",
      "Ev-flownet: Self-supervised optical flow estimation for event-based cameras.",
      "End-to-end learning of representations for asynchronous event-based data.",
      "Pointnet: Deep learning on point sets for 3d classification and segmentation.",
      "Pointnet++: Deep hierarchical feature learning on point sets in a metric space.",
      "Space-time event clouds for gesture recognition: From rgb cameras to event cameras.",
      "Efficient human pose estimation via 3d event point cloud.",
      "Point transformer.",
      "Dynamic graph cnn for learning on point clouds.",
      "Eventnet: Asynchronous recursive event processing.",
      "Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modeling.",
      "Attention is all you need.",
      "Event transformer+. a multi-purpose solution for efficient event data processing.",
      "Exploiting spatial sparsity for event cameras with visual transformers.",
      "Get: Group event transformer for event-based vision.",
      "An image is worth 16x16 words: Transformers for image recognition at scale.",
      "Eventtransact: A video transformer-based framework for event-camera based action recognition.",
      "Hats: Histograms of averaged time surfaces for robust event-based object classification.",
      "Event-based asynchronous sparse convolutional networks."
    ],
    "citations": [
      437660,
      64845,
      36451,
      388981,
      301187,
      155153,
      7385,
      412224,
      4823,
      278,
      359589,
      337259,
      38,
      244123,
      89087,
      323267,
      732,
      404526,
      531047,
      253497
    ],
    "extract_rate": 0.9523809523809523,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 78,
      "Final_Score": 81.5,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 78\nFinal_Score: 81"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 82,
      "Final_Score": 85.3,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 88\nSynthesis_Analysis: 82\nFinal_Score: 85"
    },
    "overall_score": 83.78,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "8ZDFn7BDaH",
        "forum": "8ZDFn7BDaH",
        "number": 3289,
        "title": "ALERT-Transformer: Bridging Asynchronous and Synchronous Machine Learning for Real-Time Event-based Spatio-Temporal Data",
        "normalized_title": "alert transformer bridging asynchronous and synchronous machine learning for real time event based spatio temporal data",
        "abstract": "We seek to enable classic processing of continuous ultra-sparse spatiotemporal data generated by event-based sensors with dense machine learning models. We propose a novel hybrid pipeline composed of asynchronous sensing and synchronous processing that combines several ideas: (1) an embedding based on PointNet models -- the ALERT module -- that can continuously integrate new and dismiss old events thanks to a leakage mechanism, (2) a flexible readout of the embedded data that allows to feed any downstream model with always up-to-date features at any sampling rate, (3) exploiting the input sparsity in a patch-based approach inspired by Vision Transformer to optimize the efficiency of the method. These embeddings are then processed by a transformer model trained for object and gesture recognition. Using this approach, we achieve performances at the state-of-the-art with a lower latency than competitors. We also demonstrate that our asynchronous model can operate at any desired sampling rate.",
        "authors": [
          "Carmen Martin Turrero",
          "Maxence Bouvier",
          "Manuel Breitenstein",
          "Pietro Zanuttigh",
          "Vincent Parret"
        ],
        "decision": null,
        "pdate": 1714610333339,
        "cdate": 1706630010719
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "8ZDFn7BDaH",
          "forum": "8ZDFn7BDaH",
          "number": 3289,
          "title": "ALERT-Transformer: Bridging Asynchronous and Synchronous Machine Learning for Real-Time Event-based Spatio-Temporal Data",
          "normalized_title": "alert transformer bridging asynchronous and synchronous machine learning for real time event based spatio temporal data",
          "abstract": "We seek to enable classic processing of continuous ultra-sparse spatiotemporal data generated by event-based sensors with dense machine learning models. We propose a novel hybrid pipeline composed of asynchronous sensing and synchronous processing that combines several ideas: (1) an embedding based on PointNet models -- the ALERT module -- that can continuously integrate new and dismiss old events thanks to a leakage mechanism, (2) a flexible readout of the embedded data that allows to feed any downstream model with always up-to-date features at any sampling rate, (3) exploiting the input sparsity in a patch-based approach inspired by Vision Transformer to optimize the efficiency of the method. These embeddings are then processed by a transformer model trained for object and gesture recognition. Using this approach, we achieve performances at the state-of-the-art with a lower latency than competitors. We also demonstrate that our asynchronous model can operate at any desired sampling rate.",
          "authors": [
            "Carmen Martin Turrero",
            "Maxence Bouvier",
            "Manuel Breitenstein",
            "Pietro Zanuttigh",
            "Vincent Parret"
          ],
          "decision": null,
          "pdate": 1714610333339,
          "cdate": 1706630010719
        }
      ]
    }
  },
  {
    "paper_id": 11299,
    "title": "Controlled Decoding from Language Models",
    "abstract": "KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective and propose a modular solver for it, calledcontrolled decoding (CD). CD exerts control through a separateprefix scorermodule, which is trained to learn a value function for the reward. The prefix scorer is used at inference time to control the generation from a frozen base model, provably sampling from a solution to the RL objective. We empirically demonstrate that CD is effective as a control mechanism on popular benchmarks. We also show that prefix scorers for multiple rewards may be combined at inference time, effectively solving a multi-objective RL problem with no additional training. We show that the benefits of applying CD transfer to an unseen base model with no further tuning as well. Finally, we show that CD can be applied in a blockwise decoding fashion at inference-time, essentially bridging the gap between the popular best-of-K𝐾Kitalic_Kstrategy and tokenwise control through reinforcement learning. This makes CD a promising approach for alignment of language models.",
    "introduction": "Generative language models have reached a level where they can effectively solve a variety of open-domain tasks with little task specific supervision. Hence, it is crucial to ask:how can we align machine generated content to rewards when we have no control over the pre-trained representations in a generative language model?\n\n\nControlling language model responses towards high reward outcomes is an area of active research in the literature. We divide the existing alignment methods into two categories that differ significantly in real-world deployment:generator improvementandinference-time add-onsolutions.\n\n\nGenerator improvement solutions, such as KL-regularized PPO(Christiano et al.,2017; Ouyang et al.,2022), direct preference optimization (DPO)(Rafailov et al.,2023), sequence likelihood calibration (SliC)(Zhao et al.,2022), and identity preference optimization (IPO)(Azar et al.,2023)update the weights of the language model to align it with a reward model. They are efficient for inference but offer little configurability on the reward.\n\n\nA simple and effective inference-time add-on solution is best-of-K𝐾Kitalic_K(Nakano et al.,2021; Stiennon et al.,2020; Touvron et al.,2023), whereK𝐾Kitalic_Ki.i.d. samples are drawn from a base model, ranked based on a reward, and the highest ranking one is selected.\nOther methods, such as\nFUDGE(Yang & Klein,2021)or COLD(Qin et al.,2022), offer a prefix scorer that is used at inference-time to control a frozen base model response towards high-reward outcomes.\nDue to their modularity of design which leaves the base model frozen, these methods offer inference-time configurability. Our goal is to propose a learning framework for such methods.\n\n\nOur contributions are summarized below.•We formalize a modular alignment method,controlled decoding (CD), to solve a KL-regularized RL objective. CD learns a prefix scorer for the reward that is used to steer the generation from a partially decoded path.•We show that two variants of CD, namely CD-FUDGE(Yang & Klein,2021)and CD-Q (ours), provably lead to sampling from a solution to the RL objecive.•We proposeblockwise CDwhere the prefix scorer is used to select the best-of-K𝐾Kitalic_Kpaths for a decoded block ofM𝑀Mitalic_Mtokens. This bridges the gap between the sequence-level best-of-K𝐾Kitalic_Kand tokenwise RL methods.•We empirically show that CD offers significant improvement over existing controlled generation/decoding solutions on popular benchmarks.•We show that CD prefix scorer transfers to an unseen base model with no further training.•We demonstrate the modularity of CD at inference-time to integrate multiple rewards into a single prefix scoring rule, and applying it to an unseen base model.\n\n",
    "related_work": "Controlled decoding/generation.FUDGE(Yang & Klein,2021)noticed that decoding subject to a constraint could be achieved by a prefix scorer given by the Bayes rule, and augmented the discriminative data to train the partial scorer.\nDIRECTOR(Arora et al.,2022)further showed that the partial scorer could be jointly learned with the language model itself, which would lead to a reduced latency at inference time.\nGeDi(Krause et al.,2021)proposed to train separate positive and negative scorer networks that could be combined to obtain a prefix score.Kim et al. (2023)showed that the critic in an actor-critic RL framework may be used for controlled decoding. NADO(Meng et al.,2022)considered control subject to a different divergence constraint that lends itself to a closed-form solution.\nAWR(Peng et al.,2019)extended controlled decoding to an expectation maximization setting where the policy could be subsequently updated based on the value function.\nIn contrast to this line of work, we show that the prefix scorer could be trained as the value function for the language model decoding policy, allowing us to establish an exact connection between controlled decoding and KL-regularized reinforcement learning.\n\n\nTree search.Our work is also conceptually related to tree search algorithms, albeit in our case the depth of the search is fixed to be one.Chaffin et al. (2022); Scialom et al. (2021)demonstrate that Monte Carlo tree search (MCTS) methods could be applied to language model decoding to guide the generation.Lu et al. (2022)use tree-search with a heuristic to determine the quality of a given decoding path to steer decoding towards favorable outcomes.Qin et al. (2022)explore gradient-based sampling using Langevin dynamics which significantly outperforms gradient-free sampling.\nIn contrast to all these works, the depth of search in our work is set to be one, due to the inference costs associated with inference from large LMs, which prohibits a deeper search.\n\n\nReinforcement learning (RL).Another line of very relevant work is reinforcement learning subject to a KL penalty with the language model(Ouyang et al.,2022).Korbak et al. (2022)observed that reinforcement learning with a KL penalty could be viewed in a Bayesian manner with a corresponding reward function.\nHowever, their work fell short of making the full connection in an autoregressive decoding setting, which is our contribution in this work through CD. Another closely related work to ours is that ofSnell et al. (2023)that designs a value-based offline algorithm, albeit with a different learning objective than ours (and that of the KL-regularized PPO).Li et al. (2017)also use a variant of Q-learning to optimize BLEU or ROUGE scores.\nOther related RL work includes generator improvement solutions through on-policy RL.\nSparrow(Glaese et al.,2022)showed that a variant of proximal policy optimization (PPO)(Schulman et al.,2017)with an additional LM regularizer is effective at a variety of safety objectives and alignment with human preference(Ouyang et al.,2022).\nFinally, the configurability of reward is conceptually related to(Ramé et al.,2024), where it is shown that reward soups may be used to a similar effect.\n\n\nSupervised learning from negative examples.Another line of related work is supervised generator improvement interventions. These include unlikelihood training(Welleck et al.,2020; Zhang & Song,2022), contrastive losses(Adolphs et al.,2022), direct preference optimization(Rafailov et al.,2023), and identity preference optimization(Azar et al.,2023). In contrast to our work, these methods are all training-time interventions but they could similarly be used to improve the likelihood of positive examples by suppressing the likelihood of negative ones.\n\n",
    "origin_citations": [
      "FUDGE: Controlled text generation with future discriminators.",
      "Director: Generator-classifiers for supervised language modeling.",
      "GeDi: Generative discriminator guided sequence generation.",
      "Critic-guided decoding for controlled text generation.",
      "Controllable text generation with neurally-decomposed oracle.",
      "Advantage-weighted regression: Simple and scalable off-policy reinforcement learning.",
      "PPL-MCTS: Constrained textual generation through discriminator-guided MCTS decoding.",
      "To beam or not to beam: That is a question of cooperation for language gans.",
      "NeuroLogic a*esque decoding: Constrained text generation with lookahead heuristics.",
      "COLD decoding: Energy-based constrained text generation with langevin dynamics.",
      "Training language models to follow instructions with human feedback.",
      "RL with KL penalties is better viewed as Bayesian inference.",
      "Offline rl for natural language generation with implicit language q learning.",
      "Learning to decode for future success.",
      "Improving alignment of dialogue agents via targeted human judgements.",
      "Proximal policy optimization algorithms.",
      "WARM: On the benefits of weight averaged reward models.",
      "Neural text generation with unlikelihood training.",
      "Discup: Discriminator cooperative unlikelihood prompt-tuning for controllable text generation.",
      "The cringe loss: Learning what language not to model.",
      "Direct preference optimization: Your language model is secretly a reward model.",
      "A general theoretical paradigm to understand learning from human preferences."
    ],
    "citations": [
      8047,
      883076,
      188091,
      139500,
      145876,
      416537,
      364672,
      193882,
      537823,
      173820,
      364,
      147312,
      2499,
      486686,
      2500,
      2219,
      125217,
      2336,
      336684,
      32856,
      5930,
      315102
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.30000000000001,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 82\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 83.74000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "bVIcZb7Qa0",
        "forum": "bVIcZb7Qa0",
        "number": 1952,
        "title": "Controlled Decoding from Language Models",
        "normalized_title": "controlled decoding from language models",
        "abstract": "KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective and propose a modular solver for it, called *controlled decoding (CD)*. CD exerts control through a separate *prefix scorer* module, which is trained to learn a value function for the reward. The prefix scorer is used at inference time to control the generation from a frozen base model, provably sampling from a solution to the RL objective. We empirically demonstrate that CD is effective as a control mechanism on popular benchmarks. We also show that prefix scorers for multiple rewards may be combined at inference time, effectively solving a multi-objective RL problem with no additional training. We show that the benefits of applying CD transfer to an unseen base model with no further tuning as well. Finally, we show that CD can be applied in a blockwise decoding fashion at inference-time, essentially bridging the gap between the popular best-of-$K$ strategy and tokenwise control through reinforcement learning. This makes CD a promising approach for alignment of language models.",
        "authors": [
          "Sidharth Mudgal",
          "Jong Lee",
          "Harish Ganapathy",
          "YaGuang Li",
          "Tao Wang",
          "Yanping Huang",
          "Zhifeng Chen",
          "Heng-Tze Cheng",
          "Michael Collins",
          "Trevor Strohman",
          "Jilin Chen",
          "Alex Beutel",
          "Ahmad Beirami"
        ],
        "decision": null,
        "pdate": 1714610300748,
        "cdate": 1706228696041
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "bVIcZb7Qa0",
          "forum": "bVIcZb7Qa0",
          "number": 1952,
          "title": "Controlled Decoding from Language Models",
          "normalized_title": "controlled decoding from language models",
          "abstract": "KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective and propose a modular solver for it, called *controlled decoding (CD)*. CD exerts control through a separate *prefix scorer* module, which is trained to learn a value function for the reward. The prefix scorer is used at inference time to control the generation from a frozen base model, provably sampling from a solution to the RL objective. We empirically demonstrate that CD is effective as a control mechanism on popular benchmarks. We also show that prefix scorers for multiple rewards may be combined at inference time, effectively solving a multi-objective RL problem with no additional training. We show that the benefits of applying CD transfer to an unseen base model with no further tuning as well. Finally, we show that CD can be applied in a blockwise decoding fashion at inference-time, essentially bridging the gap between the popular best-of-$K$ strategy and tokenwise control through reinforcement learning. This makes CD a promising approach for alignment of language models.",
          "authors": [
            "Sidharth Mudgal",
            "Jong Lee",
            "Harish Ganapathy",
            "YaGuang Li",
            "Tao Wang",
            "Yanping Huang",
            "Zhifeng Chen",
            "Heng-Tze Cheng",
            "Michael Collins",
            "Trevor Strohman",
            "Jilin Chen",
            "Alex Beutel",
            "Ahmad Beirami"
          ],
          "decision": null,
          "pdate": 1714610300748,
          "cdate": 1706228696041
        }
      ]
    }
  },
  {
    "paper_id": 11286,
    "title": "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution",
    "abstract": "Despite their groundbreaking performance for many generative modeling tasks, diffusion models have fallen short on discrete data domains such as natural language. Crucially, standard diffusion models rely on the well-established theory of score matching, but efforts to generalize this to discrete structures have not yielded the same empirical gains. In this work, we bridge this gap by proposing score entropy, a novel loss that naturally extends score matching to discrete spaces, integrates seamlessly to build discrete diffusion models, and significantly boosts performance. Experimentally, we test our Score Entropy Discrete Diffusion models (SEDD) on standard language modeling tasks. For comparable model sizes, SEDD beats existing language diffusion paradigms (reducing perplexity by25252525-75757575%) and is competitive with autoregressive models, in particular outperforming GPT-2. Furthermore, compared to autoregressive mdoels, SEDD generates faithful text without requiring distribution annealing techniques like temperature scaling (around6666-8×8\\times8 ×better generative perplexity than un-annealed GPT-2), can trade compute and quality (similar quality with32×32\\times32 ×fewer network evaluations), and enables controllable infilling (matching nucleus sampling quality while enabling other strategies besides left to right prompting).",
    "introduction": "Many recent advances in deep learning have centered around generative modeling. Here, a model learns how to generate novel samples from unstructured data. With the powerful capabilities of modern neural networks, these “generative AI” systems have developed unparalleled capabilities, such as creating images given only text(Ramesh et al.,2022)and answering complex questions(Brown et al.,2020).\n\n\nThe crucial part for any deep generative model is the probabilistic modeling technique. For discrete data such as natural language, autoregressive modeling(Yule,1971)–arguably the simplest modeling type since it derives from the probabilistic chain rule–has remained the only competitive method for decades. Although modern autoregressive transformers have produced stunning results(Vaswani et al.,2017; Radford et al.,2019), there are limits. For example, the sequential sampling of tokens is slow, hard to control, and often degrades without distribution annealing techniques like nucleus sampling(Holtzman et al.,2019).\n\n\nTo alleviate these issues, researchers have sought alternative approaches to generating text data. In particular, inspired by their success in the image domain, many works have extended diffusion models(Sohl-Dickstein et al.,2015; Ho et al.,2020; Song et al.,2021c)to language domains(Li et al.,2022; Austin et al.,2021). Yet, despite considerable effort, no such approach yet rivals autoregressive modeling, as they are not competitive on likelihoods, are slower to sample from, and do not generate comparable samples without resorting to heavy annealing and empirical alterations.\n\n\nIn our work, we challenge the longstanding dominance of autoregressive models by introducing Score Entropy Discrete Diffusion models (SEDD). SEDD parameterizes a reverse discrete diffusion process using the ratios of the data distribution. These are learned using score entropy, a novel loss that is analogous to score matching for standard diffusion models(Hyvärinen,2005; Song & Ermon,2019)and results in several empirical benefits††We open source our code atgithub.com/louaaron/Score-Entropy-Discrete-Diffusion:1.On core language modeling tasks, SEDD outperforms all existing language diffusion models(Li et al.,2022; Austin et al.,2021; Gulrajani & Hashimoto,2023; He et al.,2022)by large margins and is competitive with autoregressive models of the same size (beating GPT-2 on its zero-shot perplexity tasks(Radford et al.,2019)).2.SEDD generates high quality unconditional samples and enables one to naturally trade off compute for quality. When measuring the generative perplexity (given by large models) of unconditional and un-annealed samples from similarly sized models, SEDD beats GPT-2 by6666-8×8\\times8 ×and can match performance using32×32\\times32 ×fewer function evaluations.3.By directly parameterizing probability ratios, SEDD is highly controllable. In particular, one can prompt SEDD from arbitrary positions without specialized training. For both standard (left to right) and infilling, SEDD outperforms language diffusion models and is comparable with autoregressive models with nucleus sampling (as measured by MAUVE score(Pillutla et al.,2021)).\n\n",
    "related_work": "Continuous Diffusion Models for Text Data.Initially proposed byLi et al. (2022), continuous language diffusion models embed tokens in a latent space, learn a diffusion model there, and take the nearest neighbor to dequantize. While initial versions struggled, these models have achieved significant results by iterating on several empirical components. For example, prior works improve downstream performance with alternative loss functions (moving away from likelihood-based score matching)(Han et al.,2022; Mahabadi et al.,2023)and explicitly encoding conditional information (e.g. inputting an infilling mask)(Gong et al.,2023; Dieleman et al.,2022). Additionally, distribution annealing methods like thresholding(Li et al.,2022)and classifier-free guidance(Ho,2022)can further improve generation quality, although recent work has shown that methods like self-conditioning(Strudel et al.,2022)and designing a less sparse embedding space (e.g. based on bits)(Chen et al.,2022)can obviate the need for such methods. Finally,Gulrajani & Hashimoto (2023)showed that, with many surgical changes to the training paradigm, it is possible for language diffusion models to begin approaching autoregressive performance for likelihoods.\n\n\nDiscrete Diffusion Models.Most discrete diffusion works follow the framework set out by D3PM(Austin et al.,2021)which mimics “mean prediction”(Ho et al.,2020). These discrete diffusion methods are largely applied to fields other than language (e.g. images), likely due to empirical challenges. Despite this, some works have shown strong performance on language, particularly for seq-to-seq tasks and more efficient generation(Zheng et al.,2023; Chen et al.,2023; Ye et al.,2023). Notably, from these works discrete diffusion has tended to be advantageous over continuous diffusion in reducing network evaluations.\n\n\nSEDD vs Prior Work.SEDD is a discrete diffusion model that focuses on score matching, the crucial ingredient for continuous diffusions(Song & Ermon,2019; Ho et al.,2020). Many such works also focus on reversing a discrete diffusion process(Campbell et al.,2022; Benton et al.,2022; Sun et al.,2023), so score entropy is naturally related with prior training objectives. However, SEDD focuses on a principled, scalable, and performant objective (namely denoising score entropy), filling in shortcomings found in previous works. In particular, prior methods train either with the equivalent of implicit score entropy (which is intractable and high variance) or propose alternate losses that suffer from other issues. These critical differences enable large improvements for language tasks, where prior discrete diffusion models have conspicuously struggled on.\n\n\nFurthermore, SEDD achieves better results (for both perplexity and generation) than even continuous diffusion models (without resorting to empirically driven heuristics). This is desirable since discrete data should necessitate a novel approach. Future work could adapt empirical designs from continuous diffusion, further improving performance.\n\n\nFinally, SEDD challenges autoregressive models, achieving competitive perplexities (beating GPT-2) and generation quality (beating nucleus sampling). While there is still a large gap with modern large language models, we believe that future work can bridge this using SEDD as a backbone.\n\n\nSEDD vs Autoregressive Sampling Iterations.SEDD and autoregressive models have significantly different sampling procedures due to the introduction of the KV-cache for standard decoder-only transformer models. In particular, this complicates the inference code (as each network pass changes from being a standard full batch forward) and trades off speed with memory. For example, for our (known) unoptimized codebase and the existing huggingface transformers library(Wolf et al.,2020), we observed that SEDD matches autoregressive inference time when using around 100 steps but can increase the batch size by roughly4−6464-64 - 6times by removing the KV-cache memory. Future work will likely decrease the steps required for optimal generation (similar to existing work in standard diffusion(Song et al.,2021a)) which can improve this tradeoff.\n\n",
    "origin_citations": [
      "Diffusion-lm improves controllable text generation.",
      "Ssd-lm: Semi-autoregressive simplex-based diffusion language model for text generation and modular control.",
      "Tess: Text-to-text self-conditioned simplex diffusion.",
      "Diffuseq: Sequence to sequence text generation with diffusion models.",
      "Continuous diffusion for categorical data.",
      "Classifier-free diffusion guidance.",
      "Self-conditioned embedding diffusion for text generation.",
      "Analog bits: Generating discrete data using diffusion models with self-conditioning.",
      "Likelihood-based diffusion language models.",
      "Structured denoising diffusion models in discrete state-spaces.",
      "Denoising diffusion probabilistic models.",
      "A reparameterized discrete diffusion model for text generation.",
      "Fast sampling via de-randomization for discrete diffusion models.",
      "Dinoiser: Diffused conditional sequence learning by manipulating noises.",
      "Generative modeling by estimating gradients of the data distribution.",
      "A continuous time framework for discrete denoising models.",
      "From denoising diffusions to denoising markov models.",
      "Score-based continuous-time discrete diffusion models.",
      "Transformers: State-of-the-art natural language processing.",
      "Denoising diffusion implicit models."
    ],
    "citations": [
      7959,
      967120,
      704220,
      5924,
      242786,
      5922,
      33473,
      234920,
      256597,
      372220,
      4827,
      489508,
      270323,
      227232,
      5921,
      145461,
      33768,
      1089431,
      356595
    ],
    "extract_rate": 0.95,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 75,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 88,
      "Final_Score": 81.7,
      "raw_text": "Content_Coherence: 75\nCitation_Quality: 82\nSynthesis_Analysis: 88\nFinal_Score: 82"
    },
    "overall_score": 83.62,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "CNicRIVIPA",
        "forum": "CNicRIVIPA",
        "number": 6964,
        "title": "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution",
        "normalized_title": "discrete diffusion modeling by estimating the ratios of the data distribution",
        "abstract": "Despite their groundbreaking performance for many generative modeling tasks, diffusion models have fallen short on discrete data domains such as natural language. Crucially, standard diffusion models rely on the well-established theory of score matching, but efforts to generalize this to discrete structures have not yielded the same empirical gains. In this work, we bridge this gap by proposing score entropy, a novel loss that naturally extends score matching to discrete spaces, integrates seamlessly to build discrete diffusion models, and significantly boosts performance. Experimentally, we test our Score Entropy Discrete Diffusion models (SEDD) on standard language modeling tasks. For comparable model sizes, SEDD beats existing language diffusion paradigms (reducing perplexity by $25$-$75$%) and is competitive with autoregressive models, in particular outperforming GPT-2. Furthermore, compared to autoregressive mdoels, SEDD generates faithful text without requiring distribution annealing techniques like temperature scaling (around $6$-$8\\times$ better generative perplexity than un-annealed GPT-2), can trade compute and quality (similar quality with $32\\times$ fewer network evaluations), and enables controllable infilling (matching nucleus sampling quality while enabling other strategies besides left to right prompting).",
        "authors": [
          "Aaron Lou",
          "Chenlin Meng",
          "Stefano Ermon"
        ],
        "decision": null,
        "pdate": 1714610430040,
        "cdate": 1706832682687
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "CNicRIVIPA",
          "forum": "CNicRIVIPA",
          "number": 6964,
          "title": "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution",
          "normalized_title": "discrete diffusion modeling by estimating the ratios of the data distribution",
          "abstract": "Despite their groundbreaking performance for many generative modeling tasks, diffusion models have fallen short on discrete data domains such as natural language. Crucially, standard diffusion models rely on the well-established theory of score matching, but efforts to generalize this to discrete structures have not yielded the same empirical gains. In this work, we bridge this gap by proposing score entropy, a novel loss that naturally extends score matching to discrete spaces, integrates seamlessly to build discrete diffusion models, and significantly boosts performance. Experimentally, we test our Score Entropy Discrete Diffusion models (SEDD) on standard language modeling tasks. For comparable model sizes, SEDD beats existing language diffusion paradigms (reducing perplexity by $25$-$75$%) and is competitive with autoregressive models, in particular outperforming GPT-2. Furthermore, compared to autoregressive mdoels, SEDD generates faithful text without requiring distribution annealing techniques like temperature scaling (around $6$-$8\\times$ better generative perplexity than un-annealed GPT-2), can trade compute and quality (similar quality with $32\\times$ fewer network evaluations), and enables controllable infilling (matching nucleus sampling quality while enabling other strategies besides left to right prompting).",
          "authors": [
            "Aaron Lou",
            "Chenlin Meng",
            "Stefano Ermon"
          ],
          "decision": null,
          "pdate": 1714610430040,
          "cdate": 1706832682687
        }
      ]
    }
  },
  {
    "paper_id": 15272,
    "title": "SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process",
    "abstract": "In this paper, we explore a principal way to enhance the quality of object masks produced by different segmentation models. We propose a model-agnostic solution called SegRefiner, which offers a novel perspective on this problem by interpreting segmentation refinement as a data generation process. As a result, the refinement process can be smoothly implemented through a series of denoising diffusion steps. Specifically, SegRefiner takes coarse masks as inputs and refines them using a discrete diffusion process. By predicting the label and corresponding states-transition probabilities for each pixel, SegRefiner progressively refines the noisy masks in a conditional denoising manner. To assess the effectiveness of SegRefiner, we conduct comprehensive experiments on various segmentation tasks, including semantic segmentation, instance segmentation, and dichotomous image segmentation. The results demonstrate the superiority of our SegRefiner from multiple aspects. Firstly, it consistently improves both the segmentation metrics and boundary metrics across different types of coarse masks. Secondly, it outperforms previous model-agnostic refinement methods by a significant margin. Lastly, it exhibits a strong capability to capture extremely fine details when refining high-resolution images. The source code and trained models are available atgithub.com/MengyuWang826/SegRefiner.",
    "introduction": "Although segmentation in image[35;23;24;14;11]and video[54;37;38;16;15]has been extensively studied in the past decades, obtaining accurate and detailed segmentation masks is always challenging since high-quality segmentation requires the model to capture both high-level semantic information and fine-grained texture information to make accurate predictions. This challenge is particularly pronounced for images with resolutions of2Kor higher, which requires considerable computational complexity and memory usage. As a result, existing segmentation algorithms often predict masks at a smaller size, inevitably leading to lower accuracy due to the loss of fine-grained information during downsampling.\n\n\nSince directly predicting high-quality masks is challenging, some previous works have shifted their attention to the refinement of coarse masks obtained from a preceding segmentation model. A popular line of direction[29;57;27;28]is to augment the segmentation models (and features) with a new module for masks correction. However, such approaches are usually model-specific, and hence cannot be generalized to refine coarse masks produced by other segmentation models. Another part segmentation refinement works[47;56;12;44], on the other hand, resort to model-agnostic approach by taking only an image and the coarse segmentation as input for refinement. These methods have greater practical utility, as they are applicable to refine different segmentation models. However, the diverse types of errors (e.g., errors along object boundaries, failure in capturing fine-grained details in high-resolution images and errors due to incorrect semantics) presented in the coarse masks pose a great challenge to the refinement model thus causing underperformance (refer toFig.1).\n\n\n(a)Image(b)Coarse Mask(c)CRM[44](d)SegRefiner(ours)Figure 1:Diverse errors presented in the previous segmentation results. The top row and middle row show the significant false positives and false negatives caused by incorrect semantics, the bottom row shows the inaccurate in capturing fine-grained details. Please zoom in for a better view.\n\n\nIn response to this challenge, we draw inspiration from the working principle of denoising diffusion models[45;22;2]. Diffusion models perform denoising at each timestep and gradually approach the image distribution through multiple iterations. This iterative strategy significantly reduces the difficulty of fitting the target distribution all at once, empowering diffusion models to generate high-quality images. Intuitively, by applying this strategy to the segmentation refinement task, refinement model can focus on correcting only some “most obvious errors” at each step and iteratively converge to an accurate result, thus reducing the difficulty of correcting all errors in a single pass and enabling refinement model to handle more challenging instances.\n\n\nUnder this perspective, we present an innovative interpretation of the task by representing segmentation refinement as a data generation process. As a result, refinement can be implemented through a sequence of denoising diffusion steps with coarse segmentation masks being the noisy version of ground truth. To work with binary masks, we further devise a novel discrete diffusion process, where every pixel performsunidirectional randomly states-transition. The proposed process can gradually convert ground truth into a coarse mask during training and employ the coarse mask as sampling origination during inference.\nIn other words, we formulate the mask refinement task as a conditional generation problem, where the input image serves as the condition for iteratively updating/refining the erroneous predictions in the coarse mask.\n\n\nTo our best knowledge, we are the first to introduce diffusion-based refinement for segmentation masks.\nOur method, called SegRefiner, is model-agnostic, and thus applicable across different segmentation models and tasks. We extensively analyze the performance of SegRefiner across various segmentation tasks, demonstrating that our SegRefiner not only outperforms all previous model-agnostic refinement methods (with +3.42 IoU, +2.21 mBA in semantic segmentation and +0.9 Mask AP, +2.2 Boundary AP in instance segmentation), but can also be effortlessly transferred to other segmentation tasks (e.g., the recently proposed dichotomous image segmentation task[40]) without any modification. Additionally, SegRefiner demonstrates a strong capability for capturing extremely fine details when applied to high-resolution images (seeFig.1andFig.4).\n\n",
    "related_work": "Segmentation RefinementThe aim of segmentation refinement is to improve the quality of masks in pre-existing segmentation models. Some works focus on enhancing specific segmentation models. PointRend[29]employs an MLP to predict the labels of pixels with low-confidence scores output from Mask R-CNN[21]. RefineMask[57]incorporates a semantic head to Mask R-CNN as additional guidance. MaskTransfiner[27]employs an independent FCN[35]to detect incoherent regions and refines their labels with a Transformer[49]. These works have demonstrated notable performance enhancements over their preceding models. However, their scope of improvement is model-specific and they lack the capacity to directly refine the coarse mask derived from other models.\nThere are also some refinement methods that adopt model-agnostic approaches, such as[47;56;61;12;44;31]. These strategies emphasize utilizing diverse forms of input, including whole images, boundary patches, edge strips,etc.Even though these techniques can refine coarse masks derived from different models, their applicability remains confined to specific segmentation tasks.\nAs a special case, SegFix[56], which learns a mapping function between edge pixels and inner pixels and subsequently replaces inaccurate edge predictions with corresponding inner pixel predictions, is employed in both semantic segmentation and instance segmentation within the Cityscapes dataset[13]. While the performance of SegFix is significantly constrained by its ability to accurately identify objects within an image, which consequently leads to a decline in performance on datasets with a more extensive range of categories (e.g., COCO[34]).\n\n\nDiffusion Models for Detection and SegmentationRecently, diffusion models have received a lot of attention in research. Initial studies[45;22;2;46;8;3]primarily sought to enhance and expand the diffusion framework. Following these, subsequent works ventured to incorporate diffusion models across a broader array of tasks[43;9;30]and to formulate comprehensive conditional generation frameworks[41;58]. The application of diffusion models to detection and segmentation tasks has also been the focus of an escalating number of studies. Baranchuket al.[4]capture the intermediate activations from diffusion models and employ an MLP to execute per-pixel classification. DiffusionDet[6]and DiffusionInst[19]adapt the diffusion process to perform denoising in object boxes and mask filters.\nSome other works[1;52;53;7;26]consider the image segmentation task as mask generation. These studies predominantly follow the Gaussian diffusion process of DDPM[22]and leverage an additional image encoder to extract image features as condition to generate masks. To the best of our knowledge, our SegRefiner is the first work that applies a diffusion model to the image segmentation refinement task, and it also pioneers the abandonment of the Gaussian assumption in favor of a newly designed discrete diffusion process in diffusion-based segmentation tasks.\n\n",
    "origin_citations": [
      "Pointrend: Image segmentation as rendering.",
      "Mask r-cnn.",
      "Refinemask: Towards high-quality instance segmentation with fine-grained features.",
      "Mask transfiner for high-quality instance segmentation.",
      "Fully convolutional networks for semantic segmentation.",
      "Attention is all you need.",
      "Look closer to segment better: Boundary patch refinement for instance segmentation.",
      "Segfix: Model-agnostic boundary refinement for segmentation.",
      "Deepstrip: High-resolution boundary refinement.",
      "Cascadepsp: Toward class-agnostic and very high-resolution segmentation via global and local refinement.",
      "High quality segmentation for ultra high-resolution images.",
      "Polytransform: Deep polygon transformer for instance segmentation.",
      "The cityscapes dataset.",
      "Microsoft coco: Common objects in context.",
      "Deep unsupervised learning using nonequilibrium thermodynamics.",
      "Denoising diffusion probabilistic models.",
      "Structured denoising diffusion models in discrete state-spaces.",
      "Denoising diffusion implicit models.",
      "Analog bits: Generating discrete data using diffusion models with self-conditioning.",
      "Cold diffusion: Inverting arbitrary image transforms without noise.",
      "Image super-resolution via iterative refinement.",
      "Motion-conditioned diffusion model for controllable video synthesis.",
      "GLIGEN: open-set grounded text-to-image generation.",
      "High-resolution image synthesis with latent diffusion models.",
      "Adding conditional control to text-to-image diffusion models.",
      "Label-efficient semantic segmentation with diffusion models.",
      "Diffusiondet: Diffusion model for object detection.",
      "Diffusioninst: Diffusion model for instance segmentation.",
      "Segdiff: Image segmentation with diffusion probabilistic models.",
      "Medsegdiff: Medical image segmentation with diffusion probabilistic model.",
      "Medsegdiff-v2: Diffusion based medical image segmentation with transformer.",
      "A generalist framework for panoptic segmentation of images and videos.",
      "DDP: diffusion model for dense visual prediction."
    ],
    "citations": [
      738,
      520,
      1016589,
      337729,
      810,
      38,
      28112,
      201142,
      252589,
      418276,
      337381,
      204147,
      486,
      5918,
      4827,
      372220,
      356595,
      234920,
      68496,
      894646,
      51850,
      214623,
      2240,
      489610,
      152182,
      245286,
      222845,
      153340,
      35078,
      214364,
      459055,
      492186
    ],
    "extract_rate": 0.9696969696969697,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 83,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.1,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 83\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 83.62,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2023/Conference",
        "venue": "NeurIPS",
        "year": "2023",
        "openreview_id": "Q9CNA7B7v2",
        "forum": "Q9CNA7B7v2",
        "number": 6329,
        "title": "SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process",
        "normalized_title": "segrefiner towards model agnostic segmentation refinement with discrete diffusion process",
        "abstract": "In this paper, we explore a principal way to enhance the quality of object masks produced by different segmentation models. We propose a model-agnostic solution called SegRefiner, which offers a novel perspective on this problem by interpreting segmentation refinement as a data generation process. As a result, the refinement process can be smoothly implemented through a series of denoising diffusion steps. Specifically, SegRefiner takes coarse masks as inputs and refines them using a discrete diffusion process. By predicting the label and corresponding states-transition probabilities for each pixel, SegRefiner progressively refines the noisy masks in a conditional denoising manner. To assess the effectiveness of SegRefiner, we conduct comprehensive experiments on various segmentation tasks, including semantic segmentation, instance segmentation, and dichotomous image segmentation. The results demonstrate the superiority of our SegRefiner from multiple aspects. Firstly, it consistently improves both the segmentation metrics and boundary metrics across different types of coarse masks. Secondly, it outperforms previous model-agnostic refinement methods by a significant margin. Lastly, it exhibits a strong capability to capture extremely fine details when refining high-resolution images. The source code and trained models are available at [SegRefiner.git](https://github.com/MengyuWang826/SegRefiner)",
        "authors": [
          "Mengyu Wang",
          "Henghui Ding",
          "Jun Hao Liew",
          "Jiajun Liu",
          "Yao Zhao",
          "Yunchao Wei"
        ],
        "decision": null,
        "pdate": 1695325922457,
        "cdate": 1683705857437
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2023/Conference",
          "venue": "NeurIPS",
          "year": "2023",
          "openreview_id": "Q9CNA7B7v2",
          "forum": "Q9CNA7B7v2",
          "number": 6329,
          "title": "SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process",
          "normalized_title": "segrefiner towards model agnostic segmentation refinement with discrete diffusion process",
          "abstract": "In this paper, we explore a principal way to enhance the quality of object masks produced by different segmentation models. We propose a model-agnostic solution called SegRefiner, which offers a novel perspective on this problem by interpreting segmentation refinement as a data generation process. As a result, the refinement process can be smoothly implemented through a series of denoising diffusion steps. Specifically, SegRefiner takes coarse masks as inputs and refines them using a discrete diffusion process. By predicting the label and corresponding states-transition probabilities for each pixel, SegRefiner progressively refines the noisy masks in a conditional denoising manner. To assess the effectiveness of SegRefiner, we conduct comprehensive experiments on various segmentation tasks, including semantic segmentation, instance segmentation, and dichotomous image segmentation. The results demonstrate the superiority of our SegRefiner from multiple aspects. Firstly, it consistently improves both the segmentation metrics and boundary metrics across different types of coarse masks. Secondly, it outperforms previous model-agnostic refinement methods by a significant margin. Lastly, it exhibits a strong capability to capture extremely fine details when refining high-resolution images. The source code and trained models are available at [SegRefiner.git](https://github.com/MengyuWang826/SegRefiner)",
          "authors": [
            "Mengyu Wang",
            "Henghui Ding",
            "Jun Hao Liew",
            "Jiajun Liu",
            "Yao Zhao",
            "Yunchao Wei"
          ],
          "decision": null,
          "pdate": 1695325922457,
          "cdate": 1683705857437
        }
      ]
    }
  },
  {
    "paper_id": 41799,
    "title": "A Simple Approach to Unifying Diffusion-based Conditional Generation",
    "abstract": "Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g.image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g.depth to image), estimation (e.g.image to depth), signal guidance, joint generation (image & depth), and coarse control. Previous attempts at unification often introduce significant complexity through multi-stage training, architectural modification,\nor increased parameter counts. In contrast, our simple formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods.\nWe also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.",
    "introduction": "Text-to-image diffusion models, such as Dall-E 2(OpenAI,2023)and Imagen(Ho et al.,2022a), have revolutionized the field of image generation, leading to contemporary models(Midjourney,2023; Baldridge et al.,2024)that can produce images almost indistinguishable from real ones.\nThis progress in generative modeling, particularly with diffusion models, has spawned new research areas and reshaped existing fields within computer vision.\nWith advancements in image quality, the generative community has expanded its focus to controllability, resulting in many different approaches, each promoting a distinct scheme for guiding the generative process.\nControlNet(Zhang et al.,2023b)highlights the effectiveness of using modalities like depth and edges as conditional input. Meanwhile, other works, such as Loose Control(Bhat et al.,2024)and Readout Guidance(Luo et al.,2024), propose alternative conditioning types (e.g.coarse depth maps) and control mechanisms (e.g.guidance through a prediction head).\nConcurrently, the estimation community has seen diffusion models advance the state-of-the-art for predicting various modalities from RGB images,e.g.Marigold(Ke et al.,2024)repurposes a pretrained image generator to generate depth instead. In addition, other workStan et al. (2023)has explored joint diffusion, generating paired image and depth simultaneously.\n\n\nAlthough typically addressed as separate tasks within distinct communities, these problems share a common underlying structure: conditional generation between correlated images.\nConsider the relationship between an image and its depth map: controllable generation translates depth to image, estimation maps image to depth, guidance uses depth predictions to guide image generation,\nand joint generation produces image-depth pairs.\nThis observation motivates us to unify all these tasks under a global distribution modeling problem.\nWhile a few works(Qi et al.,2024; Zhang et al.,2023a)have also explored unified models capable of handling these diverse tasks, they often introduce significant complexity through multi-stage training, increased parameter counts, or architectural modifications. This additional complexity makes creating and using these models difficult, hindering their adoption.\n\n\nIn this paper, we propose UniCon, a unified diffusion model that learns an image-condition joint distribution with a flexible model architecture and simple but effective training strategy to support diverse inference behaviors.\nWe propose an architecture adaptation to the standard image generator diffusion model that is more flexible than ControlNet (allowing for non-pixel-aligned conditioning signals) and more efficient to train, decreasing both the number of learned parameters and required training samples.\nInspired by Diffusion Forcing(Chen et al.,2024a), we use a training scheme that disentangles the noise sampling of the image and the condition, allowing flexible sampling strategies at inference time to achieve different conditional generation tasks without explicit mask input.\n\n\nAs shown in Fig.1, with the same model but different sampling schedules, UniCon can do: 1) controllable image generation in the form of ControlNet, Readout Guidance, and Loose Control, 2) estimation, and 3) joint generation.\nWe train our models based on a large text-to-image model for several different modalities (depth, edges, human poses, image appearance) and show that the behavior of our single model is similar to or better than specialized methods using standard image quality and alignment metrics. We demonstrate significant improvements over prior unified models in conditional generation, training efficiency, and generation flexibility.\nWe also show that UniCon can combine multiple models for multi-signal conditional generation or switch our model to other base model checkpoints. Our models are trained in about 13 hours on 2-4 Nvidia A800 80G GPUs, adding 15% parameters to the base model.\n\n\nThe main contributions of this work are:•Proposing a framework that unifies controllable generation, estimation, and joint generation, including model adaptation, training strategy, and sampling methods allowing flexible conditional generation at inference.•Demonstrating that our architecture and training can work on a large-scale text-to-image diffusion model with a small number of learned parameters and a relatively small training data scale.•Showing that our unified models can perform similar to specialized methods or better than current unified approaches on different modalities.\n\n",
    "related_work": "Controllable Generation.Fine-tuning text-to-image diffusion models to conditional image generation on signals beyond text has gained significant popularity(Huang et al.,2023; Zhang et al.,2023b; Ye et al.,2023; Mou et al.,2023; Sohn et al.,2023).\nControlNet(Zhang et al.,2023b)trains a control network attached to pre-trained diffusion models to incorporate condition signals, such as edge maps, segmentation masks, and pose estimation.\nBased on ControlNet, LooseControl(Bhat et al.,2024)generalizes depth conditioning to loose depth maps that specify scene boundaries and object positions.\nReadout-Guidance(Luo et al.,2024)proposes a new control scheme by adding prediction heads to internal features and guiding the generation by the predicted condition signal.\nConditional editing tasks, such as DiffEdit(Couairon et al.,2023), have enhanced conditional image manipulation by applying diffusion-based models for inpainting and editing tasks.\nInstead of one specific control behavior, our proposed framework provides diverse controlling abilities through different sampling strategies.\n\n\nEstimation.Extracting signals like depth, surface normals, or segmentation maps from RGB images has been a longstanding challenge in computer vision.\nTypically, each task has been addressed in isolation or limited combinations,e.g.joint depth and segmentation, but distinct from image generation.\nStarting works like DDVM(Saxena et al.,2023a), these tasks have started to be addressed directly with diffusion models including depth prediction(Saxena et al.,2023b;a), optical flow prediction(Saxena et al.,2023a),\ncorrespondence matching(Nam et al.,2024), etc.. Recently, there has been considerable interest in either using generative features inside estimators(Xu et al.,2023; Zhao et al.,2023)or explicitly fine-tuning image generators as estimators, such as Marigold(Ke et al.,2024)which adds a clean RGB conditioning and fine-tunes the entire model to diffuse a depth map. While Marigold shares some similarities with our depth estimation setting for our RGBD model, our approach differs significantly in both goals and techniques. Unlike Marigold, which focuses solely on depth and discards image generation capabilities, our method retains the ability to perform image generation, depth estimation, and other tasks, through lightweight LoRA fine-tuning.\n\n\nJoint and Unified Generation.While less common than controllability or estimation, several works have attempted to unify multiple images and modalities within a single model. LDM3D(Stan et al.,2023)jointly generates image and depth data in an RGBD latent space. Following approaches commonly include an inpainting mask in the input to extend joint generation to bidirectional conditional generation. For example, UniGD(Qi et al.,2024)unifies image synthesis and segmentation through a diffusion model trained with image, segmentation, and inpainting mask as inputs.\nJeDi(Zeng et al.,2024)learns a joint distribution over images that share a common object, facilitating personalized image generation. Among the most relevant works, JointNet(Zhang et al.,2023a)adopts a symmetric ControlNet-like structure for generating both image and depth, utilizing an inpainting scheme to support depth-to-image and image-to-depth generation.\nOur approach presents several improvements over these unified methods. First, our training strategy enables flexible conditional generation without requiring an explicit mask input. Thus we can avoid concatenating multiple inputs in the feature dimension or adding inpainting masks, which allows our model to act like adapters that can be plugged into the base model checkpoints. Furthermore, our structure supports both loosely correlated image pairs (as in JeDi) and densely correlated pairs (as in JointNet), providing more versatile capabilities across different scenarios.\n\n",
    "origin_citations": [
      "Composer: Creative and controllable image synthesis with composable conditions.",
      "Adding conditional control to text-to-image diffusion models.",
      "Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.",
      "T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.",
      "Styledrop: Text-to-image generation in any style.",
      "Loosecontrol: Lifting controlnet for generalized depth conditioning.",
      "Readout guidance: Learning control from diffusion features.",
      "Diffedit: Diffusion-based semantic image editing with mask guidance.",
      "The surprising effectiveness of diffusion models for optical flow and monocular depth estimation.",
      "Zero-shot metric depth with a field-of-view conditioned diffusion model.",
      "Diffusion model for dense matching.",
      "Open-vocabulary panoptic segmentation with text-to-image diffusion models.",
      "Unleashing text-to-image diffusion models for visual perception.",
      "Repurposing diffusion-based image generators for monocular depth estimation.",
      "Ldm3d: Latent diffusion model for 3d.",
      "Unigs: Unified representation for image generation and segmentation.",
      "Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation.",
      "Jointnet: Extending text-to-image diffusion for dense distribution modeling."
    ],
    "citations": [
      227373,
      489610,
      179737,
      228018,
      308044,
      273379,
      273769,
      335814,
      307591,
      348093,
      256652,
      77161,
      78386,
      273773,
      261371,
      273863,
      884928,
      321384
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 85,
      "Final_Score": 81.7,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 82\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 83.62,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "tAGmxz1TUi",
        "forum": "tAGmxz1TUi",
        "number": 4930,
        "title": "A Simple Approach to Unifying Diffusion-based Conditional Generation",
        "normalized_title": "a simple approach to unifying diffusion based conditional generation",
        "abstract": "Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g. image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g. depth to image), estimation (e.g. image to depth), signal guidance, joint generation (image \\& depth), and coarse control. Previous attempts at unification often introduce complexity through multi-stage training, architectural modification, or increased parameter counts. In contrast, our simplified formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods. We also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.",
        "authors": [
          "Xirui Li",
          "Charles Herrmann",
          "Kelvin C.K. Chan",
          "Yinxiao Li",
          "Deqing Sun",
          "Chao Ma",
          "Ming-Hsuan Yang"
        ],
        "decision": null,
        "pdate": 1737562382553,
        "cdate": 1727289828010
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "tAGmxz1TUi",
          "forum": "tAGmxz1TUi",
          "number": 4930,
          "title": "A Simple Approach to Unifying Diffusion-based Conditional Generation",
          "normalized_title": "a simple approach to unifying diffusion based conditional generation",
          "abstract": "Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g. image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g. depth to image), estimation (e.g. image to depth), signal guidance, joint generation (image \\& depth), and coarse control. Previous attempts at unification often introduce complexity through multi-stage training, architectural modification, or increased parameter counts. In contrast, our simplified formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods. We also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.",
          "authors": [
            "Xirui Li",
            "Charles Herrmann",
            "Kelvin C.K. Chan",
            "Yinxiao Li",
            "Deqing Sun",
            "Chao Ma",
            "Ming-Hsuan Yang"
          ],
          "decision": null,
          "pdate": 1737562382553,
          "cdate": 1727289828010
        }
      ]
    }
  },
  {
    "paper_id": 52778,
    "title": "FedLWS: Federated Learning with Adaptive Layer-wise Weight Shrinking",
    "abstract": "In Federated Learning (FL), weighted aggregation of local models is conducted to generate a new global model, and the aggregation weights are typically normalized to 1. A recent study identifies the global weight shrinking effect in FL, indicating an enhancement in the global model’s generalization when the sum of weights (i.e., the shrinking factor) is smaller than 1, where how to learn the shrinking factor becomes crucial. However, principled approaches to this solution have not been carefully studied from the adequate consideration of privacy concerns and layer-wise distinctions. To this end, we propose a novel model aggregation strategy, Federated Learning with Adaptive Layer-wise Weight Shrinking (FedLWS), which adaptively designs the shrinking factor in a layer-wise manner and avoids optimizing the shrinking factors on a proxy dataset. We initially explored the factors affecting the shrinking factor during the training process. Then we calculate the layer-wise shrinking factors by considering the distinctions among each layer of the global model. FedLWS can be easily incorporated with various existing methods due to its flexibility. Extensive experiments under diverse scenarios demonstrate the superiority of our method over several state-of-the-art approaches, providing a promising tool for enhancing the global model in FL.",
    "introduction": "Federated Learning (FL) as an innovative paradigm in machine learning has attracted substantial attention in recent years(Li et al.,2020a; Zhang et al.,2021; Kairouz et al.,2021; Qi et al.,2023). This distributed optimization approach allows model updates to be computed locally and aggregated without exposing raw data, thereby effectively addressing the challenges arising from distributed data sources while simultaneously preserving privacy and security. In FL, weighted aggregation of local models is conducted to generate the global model, where how to design the aggregation scheme is a critical problem(Kairouz et al.,2021; Qi et al.,2023; Ye et al.,2023).\n\n\nMost previous works(Li et al.,2020b; McMahan et al.,2017)conduct model aggregation simply based on the local dataset relative size, which however could be sub-optimal empirically due to the data heterogeneity.\nConsequently, many methods(Ye et al.,2023; Hsu et al.,2019; Wang et al.,2020b)focus on adjusting the model aggregation scheme to enhance the global model’s performance, with the majority of these methods normalizing the aggregation weights (i.e., the sum of aggregation weights is 1, represented asγ=1𝛾1\\gamma=1italic_γ = 1, whereγ𝛾\\gammaitalic_γis the sum of weights).\nRecently,Li et al. (2023a)revisited the weighted aggregation process and gained new insights into the training dynamics of FL, identifying the Global Weight Shrinking effect (GWS, analogous to weight decay) in FL, whenγ𝛾\\gammaitalic_γis smaller than 1, which can further enhance the global model’s generalization.\nThe study demonstrates that during the training process, the shrinking factorγ𝛾\\gammaitalic_γplays a crucial role in maintaining the balance between the regularization term and the optimization term.Li et al. (2023a)proposed FedLAW, which optimizes the value ofγ𝛾\\gammaitalic_γusing gradient descent on a proxy dataset that has the same distribution as the global dataset. Despite its notable performance improvements, two key issues limit its practical applicability and adaptability.\n\n\nFirstly, the use of proxy datasets may raise privacy-related concerns, which are particularly crucial in the context of FL. Additionally, obtaining a dataset with a distribution identical to the global dataset is challenging in practice, and discrepancies between the proxy and global data distributions can negatively impact the method’s effectiveness.\nSecondly, divergence across different layers of deep neural networks has been demonstrated in numerous previous studies(Ma et al.,2022; Rehman et al.,2023; Lee et al.,2023). Hence, simply applying a single shrinking factorγ𝛾\\gammaitalic_γacross the entire model may not effectively harness the benefits of global weight shrinking.\"\n\n\nTo address the aforementioned issues, we initially conducted a series of experiments to explore factors influencing the shrinking factor‘s value during Federated Learning training. Building on these empirical observations from our experiments, we subsequently propose our method, Federated Learning with Adaptive Layer-wise Weight Shrinking (FedLWS).\nBy analyzing the relationship between the regularization and optimization terms, and examining how their ratio correlates with the variance of local gradients, we derive an expression for the shrinking factorγ𝛾\\gammaitalic_γ. This factor is computed directly using the gradients and parameters of the global model, which are readily accessible in Federated Learning.\nConsequently, FedLWS eliminates the need to optimize the shrinking factorγ𝛾\\gammaitalic_γon a proxy dataset, as required by previous work(Li et al.,2023a). This not only addresses privacy-related concerns but also enhances its feasibility and applicability in practical, real-world deployments.\nFurthermore, by leveraging the gradients and parameters of each layer in the global model, it is easy to calculate the layer-wiseγ𝛾\\gammaitalic_γfor each layer. This allows for an improvement in model generalization by considering layer-wise differences of shrinking factor. Moreover, our approach is conducted after the server-side model aggregation. That is to say, it is orthogonal to many existing Federated Learning methods, making it easily integrated with them to further enhance model performance.\nTo validate the effectiveness of our proposed FedLWS, we conduct extensive experiments under diverse scenarios. We observe that FedLWS can significantly improve the performance of existing FL algorithms.\n\n\nThe contributions of this work are summarized as follows:∙∙\\bullet∙We empirically show that the ratio of regularization to optimization terms in FL model aggregation is positively correlated with local gradient variance. This allows us to directly calculate the shrinking factor, eliminating the need for fine-tuning on a proxy dataset.∙∙\\bullet∙We experimentally demonstrate that the model-wise shrinking factor is suboptimal to improve the generalization of the global model. It is essential to consider the discrepancy between different layers.∙∙\\bullet∙We propose FedLWS, a simple method that generates the global model with better generalization through layer-wise weight shrinking. FedLWS requires neither additional data nor transmission of the original data, thus raising no privacy concern.∙∙\\bullet∙We conduct extensive experiments under diverse scenarios to demonstrate that FedLWS brings considerable accuracy gains over the state-of-the-art FL approaches.\n\n",
    "related_work": "Federated learning (FL) is a rapidly advancing research field with many remaining open problems to address. FedAvg(McMahan et al.,2017)has been the standard algorithm of FL. It trains local models separately and conducts model aggregation based on the data size. However, the distribution heterogeneity of local datasets may significantly degrade FL’s performance.\nIn this paper, we aim to attain a more effective and better generalized global model through collaborative training on both the client and server sides.\nMany previous works focus on this can be mainly divided into two directions.\n\n\n2.1Client-side Drift AdjustmentDue to the data heterogeneity, local models trained on the clients may exhibit different degrees of bias, thereby affecting the performance of the global model. Many methods aim to reduce this bias by adjusting the training process of local models. FedProx(Li et al.,2020b)utilizes thel2subscript𝑙2l_{2}italic_l start_POSTSUBSCRIPT 2 end_POSTSUBSCRIPT-distance between the global model and the local model as a regularization term during the training of the local model. FedDyn(Acar et al.,2021)proposes a dynamic regularizer for each client to align the global and local solutions. MOON(Li et al.,2021)aligns the features of global and local models through contrastive learning. FedDC(Gao et al.,2022)and SCAFFOLD(Karimireddy et al.,2020)adjust the drift in the local model by introducing control variates.\nFedETF(Li et al.,2023b)employs a fixed ETF classifier during training to learn unified feature representations.\nThese methods concentrate on adjusting local models, merely assigning aggregation weights based on the size of the local dataset. In contrast, our approach focuses on the server-side aggregation process, which can be easily combined with these methods to enhance model performance further due to its flexibility.\n\n\n2.2Server-side Aggregation SchemeSeveral other previous works adjust the server-side model by improving the model aggregation stage or incorporating a fine-tuning step to obtain a better global model. FedAvgM(Hsu et al.,2019)adopts momentum updates on the server-side to stabilize the training process.\nCCVR(Luo et al.,2021)adjusts the global model’s classifier using virtual representations.\nFedDF(Lin et al.,2020)and FedBE(Chen & Chao,2021)fine-tune the global model on the server. FedDisco(Ye et al.,2023)leverages both dataset size and the discrepancy between local and global category distributions to determine more distinguishing aggregation weights. Several prior studies have investigated layer-wise model aggregation.FedLAMA(Lee et al.,2023)adjusts the aggregation frequency for each layer to reduce communication costs while accounting for inter-layer differences. pFedLA(Ma et al.,2022)focuses on personalized FL, it designs a hyper-network to predict the layer-wise aggregation weights for each client.L-DAWA(Rehman et al.,2023)employs cosine similarity between local models and the global model as aggregation weights for model aggregation, simultaneously taking into account variations across different layers of the model.In comparison to these methods, our method is neither focused on aggregation frequency adjustment nor layer-wise aggregation weight adjustment. Instead, we propose an adaptive layer-wise weight shrinking step after model aggregation to mitigate aggregation bias, which is both computationally efficient and modular, enabling seamless integration with various FL frameworks and baselines.Recently, FedLAW(Li et al.,2023a)identifies the global weight shrinking phenomenon and then learns the optimal shrinking factorγ𝛾\\gammaitalic_γand the aggregation weightsλ𝜆\\lambdaitalic_λat the server with a proxy dataset, which is assumed to have the same distribution as the test dataset. Its success is inseparable from the high degree of consistency between the proxy data and the test data. Considering data privacy is a significant concern in FL, obtaining the proxy dataset with a distribution identical to the test dataset in practice is challenging, limiting its application in real-world. In addition, FedLAW ignores the variations across different layers of model for model aggregation. In contrast to FedLAW, our FedLWS calculate the shrinking factors directly through the easily available gradient and parameters of the global model, which takes into account the layer-wise differences, avoiding demanding proxy dataset and optimization. Moving beyond FedLAW, ours can be easily integrated with most of related model aggregation methods for decoupling shrinking factor and aggregation weights.\n\n",
    "origin_citations": [
      "Communication-efficient learning of deep networks from decentralized data.",
      "Federated optimization in heterogeneous networks.",
      "Federated learning based on dynamic regularization.",
      "Model-contrastive federated learning.",
      "Feddc: Federated learning with non-iid data via local drift decoupling and correction.",
      "Scaffold: Stochastic controlled averaging for federated learning.",
      "No fear of classifier biases: Neural collapse inspired federated learning with synthetic and fixed classifier.",
      "Measuring the effects of non-identical data distribution for federated visual classification.",
      "No fear of heterogeneity: Classifier calibration for federated learning with non-iid data.",
      "Ensemble distillation for robust model fusion in federated learning.",
      "Fedbe: Making bayesian model ensemble applicable to federated learning.",
      "Feddisco: Federated learning with discrepancy-aware collaboration.",
      "Layer-wise adaptive model aggregation for scalable federated learning.",
      "Layer-wised model aggregation for personalized federated learning.",
      "L-dawa: Layer-wise divergence aware weight aggregation in federated self-supervised visual representation learning.",
      "Revisiting weighted aggregation in federated learning with neural networks."
    ],
    "citations": [
      582,
      632,
      633,
      506040,
      501598,
      635,
      495742,
      648,
      194519,
      593,
      189783,
      256578,
      513788,
      378182,
      184725,
      488951
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 82,
      "Final_Score": 84.1,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 82\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 86,
      "Synthesis_Analysis": 85,
      "Final_Score": 83.3,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 86\nSynthesis_Analysis: 85\nFinal_Score: 83"
    },
    "overall_score": 83.62,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "6RjQ54M1rM",
        "forum": "6RjQ54M1rM",
        "number": 9107,
        "title": "FedLWS: Federated Learning with Adaptive Layer-wise Weight Shrinking",
        "normalized_title": "fedlws federated learning with adaptive layer wise weight shrinking",
        "abstract": "In Federated Learning (FL), weighted aggregation of local models is conducted to generate a new global model, and the aggregation weights are typically normalized to 1. A recent study identifies the global weight shrinking effect in FL, indicating an enhancement in the global model’s generalization when the sum of weights (i.e., the shrinking factor) is smaller than 1, where how to learn the shrinking factor becomes crucial. However, principled approaches to this solution have not been carefully studied from the adequate consideration of privacy concerns and layer-wise distinctions. To this end, we propose a novel model aggregation strategy, Federated Learning with Adaptive Layer-wise Weight Shrinking (FedLWS), which adaptively designs the shrinking factor in a layer-wise manner and avoids optimizing the shrinking factors on a proxy dataset. We initially explored the factors affecting the shrinking factor during the training process. Then we calculate the layer-wise shrinking factors by considering the distinctions among each layer of the global model. FedLWS can be easily incorporated with various existing methods due to its flexibility. Extensive experiments under diverse scenarios demonstrate the superiority of our method over several state-of-the-art approaches, providing a promising tool for enhancing the global model in FL.",
        "authors": [
          "Changlong Shi",
          "Jinmeng Li",
          "He Zhao",
          "Dan dan Guo",
          "Yi Chang"
        ],
        "decision": null,
        "pdate": 1737562637035,
        "cdate": 1727414551813
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "6RjQ54M1rM",
          "forum": "6RjQ54M1rM",
          "number": 9107,
          "title": "FedLWS: Federated Learning with Adaptive Layer-wise Weight Shrinking",
          "normalized_title": "fedlws federated learning with adaptive layer wise weight shrinking",
          "abstract": "In Federated Learning (FL), weighted aggregation of local models is conducted to generate a new global model, and the aggregation weights are typically normalized to 1. A recent study identifies the global weight shrinking effect in FL, indicating an enhancement in the global model’s generalization when the sum of weights (i.e., the shrinking factor) is smaller than 1, where how to learn the shrinking factor becomes crucial. However, principled approaches to this solution have not been carefully studied from the adequate consideration of privacy concerns and layer-wise distinctions. To this end, we propose a novel model aggregation strategy, Federated Learning with Adaptive Layer-wise Weight Shrinking (FedLWS), which adaptively designs the shrinking factor in a layer-wise manner and avoids optimizing the shrinking factors on a proxy dataset. We initially explored the factors affecting the shrinking factor during the training process. Then we calculate the layer-wise shrinking factors by considering the distinctions among each layer of the global model. FedLWS can be easily incorporated with various existing methods due to its flexibility. Extensive experiments under diverse scenarios demonstrate the superiority of our method over several state-of-the-art approaches, providing a promising tool for enhancing the global model in FL.",
          "authors": [
            "Changlong Shi",
            "Jinmeng Li",
            "He Zhao",
            "Dan dan Guo",
            "Yi Chang"
          ],
          "decision": null,
          "pdate": 1737562637035,
          "cdate": 1727414551813
        }
      ]
    }
  },
  {
    "paper_id": 16898,
    "title": "CREMA: Generalizable and Efficient Video-Language Reasoning via Multimodal Modular Fusion",
    "abstract": "Despite impressive advancements in recent multimodal reasoning approaches, they are still limited in flexibility and efficiency, as these models typically process only a few fixed modality inputs and require updates to numerous parameters.\nThis paper tackles these critical challenges and proposesCREMA, a generalizable, highly efficient, and modular modality-fusion framework that can incorporate many new modalities to enhance video reasoning.\nWe first augment multiple informative modalities (such asoptical flow,3D point cloud,audio,thermal heatmap, andtouch map) from given videos without extra human annotation by leveraging sensors or existing pre-trained models.\nNext, we introduce a query transformer with multiple parameter-efficient modules associated with each accessible modality.\nIt projects diverse modality features to the LLM token embedding space, allowing the model to integrate different data types for response generation.\nFurthermore, we propose a novel progressive multimodal fusion design supported by a lightweight fusion module and modality-sequential training strategy.\nIt helps compress information across various assisting modalities, maintaining computational efficiency in the LLM while improving performance.\nWe validate our method on 7 video-language reasoning tasks assisted by diverse modalities, including conventional VideoQA and Video-Audio/3D/Touch/Thermal QA, and achieve better/equivalent performance against strong multimodal LLMs, including OneLLM, BLIP-2, and SeViLA while reducing over 90% trainable parameters.\nWe provide extensive analyses ofCREMA, including the impact of each modality on reasoning domains, the design of the fusion module, and example visualizations.111Project Page:https://CREMA-VideoLLM.github.io/.",
    "introduction": "We humans understand the world through various senses, such as sight, sound, touch, and heat, allowing us to understand our environment and act accordingly.\nThis concept has inspired the field of multimodal learning that connects various perceptions, including vision-language(Alayrac et al.,2022, Li et al.,2023b, Zang et al.,2023, Radford et al.,2021), audio-video(Han et al.,2020, Tang et al.,2022), and 2D-3D joint vision(Li et al.,2020, Hou et al.,2021;2023, Lei et al.,2024).\nIn particular, recent Multimodal Large Language Models (MLLMs)(Yu et al.,2023b, Li et al.,2023b, Liu et al.,2023a, Tang et al.,2023a)have shown promising versatility in handling multiple forms of input data, such as vision, audio, and text.\nThese models are crucial in real-world applications that require a comprehensive understanding of multiple modalities to make decisions in various contexts.\nFor example, autonomous vehicles rely on visual road signs, sirens, and LIDAR for navigation and safe driving. Embodied AI takes visual, heat, and touch information to complete household tasks. Similarly, educational AI enhances the learning experience by integrating various information, such as videos, speech, and textbooks.\n\n\nFigure 1:Overview of theCREMAarchitecture & training.Left:Multimodal encoders, Q-former, and LLM are kept frozen in the process.\nFor each modality input, we extract tokens using a corresponding modality-specific adaptation module. Then, we employ the fusion module to blend and compress the obtained multimodal tokens. In the end, the LLM uses modality-fusion tokens to generate responses.Right:We present a modality-sequential training and modality-adaptive early exist strategy, further boosting the training efficiency while allowing faster modality adaptation.\n\n\nDespite their recent advancements, deploying a generic MLLM that handles multiple diverse modalities is still very challenging in terms ofcostandflexibility. For different types of inputs, MLLMs have required extremely large computational budgets to update the LLM with individual encoders for modalities.\nAlternatively, recent efficient MLLMs using separate projection modules(Zhang et al.,2023, Sun et al.,2023a, Li et al.,2023b, Han et al.,2023)provide a more efficient and flexible way for multimodal reasoning. However, as each modality module contains hundreds of millions of parameters for training, this approach is still computationally intensive, and balancing as well as fusing various types of inputs becomes even more complex and costly when more modalities are introduced.\nSuch challenges also exist in very recent pioneering works(Liu et al.,2023c, Panagopoulou et al.,2023, Lu et al.,2023); these models aim to integrate more diverse sensory data for compositional understanding via partial updates to the models,\nyet still require notable training resources to adapt to different modalities (7B for Unified-IO 2(Lu et al.,2023)).\nMoreover, they focus primarily on fixed modality pairs (like 3D-text and visual-text), limiting their adaptability to new data forms and broader applications.\nThis raises the crucial research question:Can we enable MLLMs to efficiently leverage multiple modalities for Video-Language Reasoning at lower costs, similar to how humans do?.\n\n\nIn this paper, we presentCREMA: a generalizable and highly efficient MLLM framework designed for video-language reasoning with many modalities.\nThe proposed framework enables us to incorporate any new set of modalities,\nincludingvideo,depth map,optical flow,surface normals,audio,thermal heatmap,touch map,3D point cloud, notably with very few trainable parameters (4∼similar-to\\sim∼10M for each new modality) as compared to BLIP-2(Li et al.,2023b)(∼similar-to\\sim∼108M) and SeViLA(Yu et al.,2023a)(∼similar-to\\sim∼216M) but higher performance due to assistance of diverse modalities.\nAs illustrated inFigure1Left, given a frozen pre-trained vision-language backbone, our approach introduces modality-adaptive modules on top of the Q-Former(Li et al.,2023b)architecture, including linear projectors, low-rank adapters(Hu et al.,2022), and learnable queries.\nOur parameter-efficient modular design ensures that the pre-trained backbone remains unchanged and enables updates with new modalities and more advanced LLMs in the future without complex architecture changes.\nTo enrich the input modalities, we utilize pre-trained models to extract features from raw videos, e.g., depth maps and optical flow.\n\n\nFurthermore, despite the effectiveness of our versatile video reasoning framework for multimodal data, handling multiple modalities is not always beneficial; some modalities may be redundant or irrelevant to the reasoning tasks, and optimizing with all modalities simultaneously can lead to a certain deficiency, such as suboptimal convergence of the modality inputs.\nBesides, the LLM receives longer input contexts, which include token embeddings from all modality queries, resulting in increased computations to produce responses.\nTo address these remaining concerns, we propose a new modality fusion technique that effectively integrates various modality tokens through a novel self-gated attention module and performs alternative modality training within a minibatch, preventing interference in the learning of modality-specific representation. Furthermore, we present a modality-adaptive early exit strategy to bypass the training of a specific modality if it is considered as converged. This further increases the efficiency ofCREMAwhile allowing faster adaptation and maintaining comparable performance. (Please seeFigure1Right). In the end, we allow the model to maintain GFLOPs while achieving enhanced reasoning ability, even when the LLM processes many modalities.\n\n\nWe validateCREMAon various video-language reasoning benchmarks assisted by diverse modalities, including VideoQA (NExT-QA(Xiao et al.,2021), PerceptionTest(Pătrăucean et al.,2023)), 3D-VideoQA (SQA3D(Ma et al.,2023)), Audio-VideoQA (MUSIC-AVQA(Li et al.,2022), VGGSound(Chen et al.,2020)), Touch-VideoQA (Touch&Go(Yang et al.,2022)), Thermal-VideoQA (Thermal-IM(Tang et al.,2023b)).CREMAsurpasses other strong multimodal video reasoning baselines, improving fine-tuning performance by+2.4%on MUSIC-AVQA,+2.6%on SQA3D, and+0.6%on NExT-QA while reducing by90% of the trainable parameters.CREMAalso outperforms general-purpose baselines in the zero-shot setting.\nWe further provide comprehensive analyses of applyingCREMAwith different LLM, varying sets of modalities, different modality fusion strategies, benefits of adding more modality, and qualitative analysis with input/response visualizations to highlight the efficiency/effectiveness ofCREMAin various video-language reasoning across diverse modalities. Our contributions are summarized as follows:\n\n\n•We propose ahighly efficient and generalizable modality-extensible learning framework, coinedCREMA, which learns multiple modality-adaptive modules to understand given data through augmented senses.•We present anovel modality fusion and training designthat efficiently weighs modalities, integrating useful modality features into response generation.•CREMA’s design allows easy embracing ofany new modalitiesby adding additional modality-adaptive modules without any need to modify the existing framework.•We show the efficacy ofCREMAon seven video reasoning datasets by achieving better/equivalent performance whilereducing over 90% of trainable parametersthan strong baseline models.•WithCREMAframework, we conduct extensive studies to provide acomprehensive analysisof how multimodal information helps video-centric language reasoning tasks.\n\n",
    "related_work": "Learning with Multiple Modalities.Beyond conventional unimodal learning, leveraging additional modalities, such as visual and audio, in learning models is increasingly popular and has demonstrated remarkable success in solving diverse tasks(Zhu et al.,2024, Liu et al.,2023b, Lu et al.,2023, Moon et al.,2023, Wang et al.,2024a).\nVision Language Models(Huang et al.,2023, Li et al.,2023a, Gong et al.,2023, Chen et al.,2023b)are the most prevalent branch ofmultimodal learningthat combine vision and language by training on massive data to understand and generate outputs involving visual and text-based information.\nAudio-Language Models(Chuang et al.,2020, Castellon et al.,2021, Wang et al.,2023)have been proposed for various audio-associated language tasks,e.g., spoken question answering and speech synthesis.\nAlso, 2D-3D Joint Vision (and Language) Models(Li et al.,2020, Hou et al.,2021;2023, Lei et al.,2024)aim to combine features of both two-dimensional (2D) and three-dimensional (3D) data to interpret and analyze both modalities, allowing for a more comprehensive understanding of visual information.\nHowever, these approaches are not scalable to other tasks involving different modality inputs since they focus on handling fine-grained problems with pre-defined modalities.\n\n\nMultimodal Large Language Model.Very recently, several works propose integrated pipelines using more than two different data sources for general-purpose reasoning(Zellers et al.,2022, Han et al.,2023, Li et al.,2023b, Girdhar et al.,2023, Liu et al.,2023c).\nMERLOT-REVERSE(Zellers et al.,2022)introduces a new training objective that learns\nfrom audio, subtitles, and video frames. Given modality-specific features extracted from the encoder/tokenizer for these inputs, the joint transformer learns to predict the masked text and audio.\nTherefore, incorporating new types of inputs can be challenging since it requires new pre-training steps for all other modalities.\nX-InstructBLIP (X-BLIP)(Panagopoulou et al.,2023)integrates various modalities into a framework using frozen LLMs, employing modality-specific Q-Formers as adapters to connect different encoders. However, this method needs to train the individual Q-Former for each modality to enable modality-aligned instruction tuning, which is still resource-intensive.MultiPLY(Hong et al.,2024)is a multisensory embodied LLM for interaction within 3D environments with fixed modality set.\nX-VILA(Ye et al.,2024)is an omni-modality model aimed at cross-modality alignment, understanding, and generation. It concentrates on large-scale cross-modality alignment.OneLLM(Han et al.,2023)presents a universal encoder and projection module to align various modalities with language, its flexibility is limited in adapting new modalities, as the pre-trained projection may be impaired with unseen input format.\nOurCREMAmethod adopts an efficient and modular approach, using parameter-efficient adapters for each modality, and enhances flexibility in combining any new modalities.\n\n",
    "origin_citations": [
      "Minigpt-4: Enhancing vision-language understanding with advanced large language models.",
      "Visual instruction tuning.",
      "Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action.",
      "Anymal: An efficient and scalable any-modality augmented language model.",
      "Sok-bench: A situated video reasoning benchmark with aligned open-world knowledge.",
      "Large language models can self-improve.",
      "Otter: A multi-modal model with in-context instruction tuning.",
      "Multimodal-gpt: A vision and language model for dialogue with humans.",
      "Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset.",
      "Speechbert: An audio-and-text jointly learned language model for end-to-end spoken question answering.",
      "Codified audio language modeling learns useful representations for music information retrieval.",
      "Neural codec language models are zero-shot text to speech synthesizers.",
      "Detailed 2d-3d joint representation for human-object interaction.",
      "Pri3d: Can 3d priors help 2d representation learning?",
      "Mask3d: Pre-training 2d vision transformers by learning masked 3d priors.",
      "Vit-lens: Towards omni-modal representations.",
      "Merlot reserve: Neural script knowledge through vision and language and sound.",
      "Onellm: One framework to align all modalities with language.",
      "Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.",
      "Imagebind: One embedding space to bind them all.",
      "Prismer: A vision-language model with an ensemble of experts.",
      "X-instructblip: A framework for aligning x-modal instruction-aware representations to llms and emergent cross-modal reasoning.",
      "Multiply: A multisensory object-centric embodied large language model in 3d world.",
      "X-vila: Cross-modality alignment for large language model."
    ],
    "citations": [
      2243,
      2238,
      346280,
      477994,
      781726,
      8064,
      2237,
      2239,
      257055,
      70267,
      371097,
      3004,
      192913,
      164374,
      705135,
      22948,
      255470,
      273028,
      212324,
      2236,
      1090052,
      781507,
      126865,
      739404
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 87,
      "Publication_Potential": 91,
      "Final_Score": 88.2,
      "raw_text": "Content_Quality: 87\nPublication_Potential: 91\nFinal_Score: 88"
    },
    "related_work_scores": {
      "Content_Coherence": 73,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 82,
      "Final_Score": 80.5,
      "raw_text": "Content_Coherence: 73\nCitation_Quality: 85\nSynthesis_Analysis: 82\nFinal_Score: 81"
    },
    "overall_score": 83.58,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "3UaOlzDEt2",
        "forum": "3UaOlzDEt2",
        "number": 5036,
        "title": "CREMA: Generalizable and Efficient Video-Language Reasoning via Multimodal Modular Fusion",
        "normalized_title": "crema generalizable and efficient video language reasoning via multimodal modular fusion",
        "abstract": "Despite impressive advancements in recent multimodal reasoning approaches, they are still limited in flexibility and efficiency, as these models typically process only a few fixed modality inputs and require updates to numerous parameters. This paper tackles these critical challenges and proposes CREMA, a generalizable, highly efficient, and modular modality-fusion framework that can incorporate many new modalities to enhance video reasoning. We first augment multiple informative modalities (such as optical flow, 3D point cloud, audio, thermal heatmap, and touch map) from given videos without extra human annotation by leveraging sensors or existing pre-trained models. \nNext, we introduce a query transformer with multiple parameter-efficient modules associated with each accessible modality. \nIt projects diverse modality features to the LLM token embedding space, allowing the model to integrate different data types for response generation. Furthermore, we propose a novel progressive multimodal fusion design supported by a lightweight fusion module and modality-sequential training strategy. It helps compress information across various assisting modalities, maintaining computational efficiency in the LLM while improving performance. We validate our method on seven video-language reasoning tasks assisted by diverse modalities, including conventional VideoQA and Video-Audio/3D/Touch/Thermal QA, and achieve better/equivalent performance against strong multimodal LLMs, including OneLLM, BLIP-2, and SeViLA while reducing over 90% trainable parameters. We provide extensive analyses of CREMA, including the impact of each modality on reasoning domains, the design of the fusion module, and example visualizations.",
        "authors": [
          "Shoubin Yu",
          "Jaehong Yoon",
          "Mohit Bansal"
        ],
        "decision": null,
        "pdate": 1737562387961,
        "cdate": 1727293286018
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "3UaOlzDEt2",
          "forum": "3UaOlzDEt2",
          "number": 5036,
          "title": "CREMA: Generalizable and Efficient Video-Language Reasoning via Multimodal Modular Fusion",
          "normalized_title": "crema generalizable and efficient video language reasoning via multimodal modular fusion",
          "abstract": "Despite impressive advancements in recent multimodal reasoning approaches, they are still limited in flexibility and efficiency, as these models typically process only a few fixed modality inputs and require updates to numerous parameters. This paper tackles these critical challenges and proposes CREMA, a generalizable, highly efficient, and modular modality-fusion framework that can incorporate many new modalities to enhance video reasoning. We first augment multiple informative modalities (such as optical flow, 3D point cloud, audio, thermal heatmap, and touch map) from given videos without extra human annotation by leveraging sensors or existing pre-trained models. \nNext, we introduce a query transformer with multiple parameter-efficient modules associated with each accessible modality. \nIt projects diverse modality features to the LLM token embedding space, allowing the model to integrate different data types for response generation. Furthermore, we propose a novel progressive multimodal fusion design supported by a lightweight fusion module and modality-sequential training strategy. It helps compress information across various assisting modalities, maintaining computational efficiency in the LLM while improving performance. We validate our method on seven video-language reasoning tasks assisted by diverse modalities, including conventional VideoQA and Video-Audio/3D/Touch/Thermal QA, and achieve better/equivalent performance against strong multimodal LLMs, including OneLLM, BLIP-2, and SeViLA while reducing over 90% trainable parameters. We provide extensive analyses of CREMA, including the impact of each modality on reasoning domains, the design of the fusion module, and example visualizations.",
          "authors": [
            "Shoubin Yu",
            "Jaehong Yoon",
            "Mohit Bansal"
          ],
          "decision": null,
          "pdate": 1737562387961,
          "cdate": 1727293286018
        }
      ]
    }
  },
  {
    "paper_id": 47627,
    "title": "iFormer: Integrating ConvNet and Transformer for Mobile Application",
    "abstract": "We present a new family of mobile hybrid vision networks, called iFormer,\nwith a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network,i.e., ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks.\nNotably, iFormer achieves an impressive Top-1 accuracy of 80.4% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13,\nsurpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios. Code and models are available at:https://github.com/ChuanyangZheng/iFormer.",
    "introduction": "Building lightweight neural networks facilitates real-time analysis of images and videos captured by mobile applications such as smartphones.\nThis not only enhances privacy protection and security by processing data locally on the device but also improves overall user experience.\n\n\nFigure 1:Comparison of latency and accuracy between our iFormer and other existing methods on ImageNet-1k.The latency is measured on an iPhone 13. Our iFormer is Pareto-optimal.\n\n\nThrough the decades, convolutional neural networks (CNNs)(Krizhevsky et al.,2012; Szegedy et al.,2015; He et al.,2016)have emerged as the primary choice for balancing latency and performance on resource-constrained mobile devices.\nHowever, a significant limitation of CNNs is their reliance on a local sliding window mechanism, which imposes crucial inductive biases that may hinder modeling flexibility.\nRecently, the soaring development of vision transformers (ViTs)(Dosovitskiy et al.,2020)has begun to dominate various computer vision tasks, including image classification(Zhai et al.,2022), object detection(Liu et al.,2021a), and semantic segmentation(Xie et al.,2021). The core mechanism underlying ViTs is self-attention, which dynamically learns interactions between all image patches. This enables the model to focus on important regions adaptively and capture more global features. Nevertheless, deploying ViTs on mobile devices with limited resources poses significant challenges. On the one hand, the quadratic computational complexity of attention renders them unsuitable for large feature maps, which are common in the early stages of vision networks. On the other hand, the multi-head mechanism requires reshaping operations, leading to increased memory usage.\n\n\nMany research efforts are devoted to combining the advantages of both CNNs and ViTs in designing lightweight networks while mitigating inefficient operations in mobile applications. Some studies(Zhang et al.,2023; Wang et al.,2024; Ma et al.,2024)revisit the architectural designs of lightweight CNNs from a ViT perspective and incorporate key components that contribute to the performance of ViTs into CNNs. Although these pure lightweight CNNs show improved performance compared to previous mobile networks(Howard et al.,2017; Zhang et al.,2018; Sandler et al.,2018), they still lag behind the powerful self-attention in ViTs.\nAnother line of works(Mehta & Rastegari,2021; Chen et al.,2022b; Li et al.,2023; Cai et al.,2023; Shaker et al.,2023; Vasu et al.,2023a; Qin et al.,2024)proposes innovative attention mechanisms to address the limitation of standard attention(Vaswani,2017)and blend convolutions to achieve a better balance between latency and performance. These attention mechanisms either reduce the number of queries and keys(Shaker et al.,2023; Qin et al.,2024), limit the attention span(Wan et al.,2023), or adopt linear attention(Cai et al.,2023), which may compromise performance to some extent.\n\n\nIn this work, we present the iFormer, a herd of lightweight models that integrates the strengths of both CNNs and ViTs, achieving a state-of-the-art balance between latency and accuracy. Specifically, we employ a hierarchical architecture consisting of four stages. In the earlier, high-resolution stages, we utilize fast convolution to extract local representations. To construct the convolutional block, we start with a “modern” ConvNeXt(Liu et al.,2022), which incorporates a series of design decisions inspired by ViTs. Then we progressively “lighten” the ConvNeXt to create a streamlined lightweight network, optimizing it for real-time mobile latency on an iPhone 13, in contrast to the FLOPs and parameters used in prior works(Mehta & Rastegari,2021; Chen et al.,2022b). This results in a fast convolutional architecture with strong performance. To further enhance the dynamic properties and its ability to model long-range contexts, we incorporate self-attention in the later low-resolution stages. However, direct implementation of standard multi-head self-attention (MHA) brings notable memory overheads and slows down inference speed on mobile devices. We identify that the increased latency stems primarily from the reshaping operations in MHA. More analyses reveal that multiple attention heads behave similarly. Therefore, we propose a simple yet effective single-head modulation self-attention (SHMA), which significantly minimizes memory costs while preserving strong performance. Fig.4provides an illustration of SHMA. In detail, SHMA learns spatial context interactions through optimized self-attention. Concurrently, a parallel feature extraction branch is employed to capture informative features. Finally, we fuse the outputs of these two branches to facilitate a more flexible and dynamic exchange of information, compensating for the slight performance degradation of the single-head attention when compared to MHA.\n\n\nBenefiting from the fast local representation capacity of convolution and the efficient global modeling proficiency of the proposed SHMA, iFormer outperforms existing pure lightweight CNNs and hybrid networks across multiple visual recognition tasks, including image classification, object detection, instance segmentation, and semantic segmentation. For instance, in the context of image classification as shown in Fig.1, iFormer-M achieves a Top-1 accuracy of 80.4% with only 1.10 ms on an iPhone 13 without advanced training strategies such as knowledge distillation(Touvron et al.,2021a)or reparameterization(Ding et al.,2021). Notably, our model obtains a 0.5% improvement in Top-1 accuracy compared to the recent MNV4-Conv-M(Qin et al.,2024), while being 1.4×\\times×faster than FastViT-SA12(Vasu et al.,2023a)with similar accuracy. These results demonstrate the effectiveness of the proposed network in capturing both local and global feature representations.\n\n",
    "related_work": "2.1Efficient Convolutional NetworksIn the past 2010s, computer vision was dominated by CNNs, and so were efficient networks. The first remarkable breakthrough in mobile CNNs is MobileNets(Howard et al.,2017), which hatches the concept of decomposing standard convolution into depthwise and pointwise counterparts. Subsequently, MobileNetV2(Sandler et al.,2018)introduces an inverted residual bottleneck block to push the state-of-the-art for mobile models. Numerous studies have aimed to accelerate CNNs via various approaches, such as channel shuffle in ShuffleNet(Zhang et al.,2018; Ma et al.,2018)and cheap linear transformations in GhostNet(Han et al.,2020). Meanwhile, Neural architecture search (NAS) has emerged as a method for automating the design of neural networks, optimizing for performance under resource constraints. EfficientNet(Tan & Le,2019), MobileNetV3(Howard et al.,2019), and FBNet(Wu et al.,2019)all achieve rather good performance. Besides, MobileOne(Vasu et al.,2023b)proposes to train a model using reparameterizable branches, which are merged during inference. Recently, following the revolution of ViTs, several methods reexamine the design spaces and training strategies(Liu et al.,2024)for mobile CNNs. For instance, RepViT(Wang et al.,2024)integrates efficient architectural designs from ViTs into MobileNetV3, outperforming existing lightweight CNNs.\nOther approaches, such as FocalNet(Yang et al.,2022a), Conv2Former(Hou et al.,2024), and EfficientMod(Ma et al.,2024), fuse features from context modeling and feature projection branches, also known as modulation mechanism, to enhance the model with dynamic properties analogous to attention.\nHowever, pure CNNs remain inherently spatially localized and their reliance on stationary weights restricts their flexibility. Although modulation can partially mitigate this limitation by enhancing dynamic capacity, they still exhibit deficiencies in building global interactions.\n\n\n2.2Efficient Vision TransformersThe success of Vision Transformer(Dosovitskiy et al.,2020)offers a compelling demonstration of the potential to apply transformer to computer vision tasks. Following this, ViT and its numerous variants(Liu et al.,2021a; Dong et al.,2022; Li et al.,2022a)sweep across various scenarios. However, the quadratic complexity of self-attention behind ViTs poses significant challenges for efficiency. The following researches seek to boost ViT efficiency through efficient attention mechanisms(Wang et al.,2021; Zhu et al.,2023; Hatamizadeh et al.,2023), model compression(Liu et al.,2021b; Zheng et al.,2022), knowledge distillation(Hao et al.,2021), and token reduction(Rao et al.,2021; Bolya et al.,2022). Recent studies further introduce ViTs into mobile applications. One mainstream of work combines efficient convolution and ViT to create lightweight hybrid networks(Mehta & Rastegari,2022; Vasu et al.,2023a). MobileViT(Mehta & Rastegari,2021)directly integrates MobileNetv2 blocks and ViT blocks, while Mobile-Former(Chen et al.,2022b)features a parallel design of MobileNet and ViT with a two-way bridge connecting the two.\nTo further accelerate inference, some approaches replace the standard attention(Vaswani,2017)with efficient variants within the hybrid networks. These include reducing the number of delegate tokens for computing attention(Pan et al.,2022), employing channel attention(Maaz et al.,2022), substituting projection in attention with efficient ghost modules(Ma et al.,2022), and utilizing linear attention mechanisms(Zhao et al.,2022).\nBesides manual designs, EfficientFormer(Li et al.,2022b;2023)and MobileNetV4(Qin et al.,2024)search for efficient architectures in a unified space encompassing both convolution operators and transformer operators. Another stream of work focuses on efficient attention mechanisms and directly employs them throughout the entire network(Shaker et al.,2023; Cai et al.,2023). For example, CMT(Guo et al.,2022)takes advantage of depth-wise convolution to downsample key and value to reduce computation.\nGhostNetV2(Tang et al.,2022)applies two fully connected layers along the horizontal and vertical directions to compute attention, a decoupled version of MLP-Mixer(Tolstikhin et al.,2021). Recently, SHViT observes computational redundancy in the multi-head attention module and proposes to apply sing-head attention. In contrast to these existing approaches, we introduce a novel efficient attention module without sacrificing informative interactions, thereby maintaining strong representational capacity. Regarding attention design, ours is a bit similar to SHViT but is considerably superior as shown in Table18in the supplementary material. The key difference lies in the novel modulation attention. In addition, we explore efficient attention mechanisms in an on-device environment while SHViT focuses on general-purpose GPUs, fundamentally different hardware.\n\n",
    "origin_citations": [
      "Mobilenets: Efficient convolutional neural networks for mobile vision applications.",
      "Mobilenetv2: Inverted residuals and linear bottlenecks.",
      "Shufflenet: An extremely efficient convolutional neural network for mobile devices.",
      "Shufflenet v2: Practical guidelines for efficient cnn architecture design.",
      "Ghostnet: More features from cheap operations.",
      "Efficientnet: Rethinking model scaling for convolutional neural networks.",
      "Searching for mobilenetv3.",
      "Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search.",
      "Mobileone: An improved one millisecond mobile backbone.",
      "Ghostnetv3: Exploring the training strategies for compact models.",
      "Repvit: Revisiting mobile cnn from vit perspective.",
      "Focal modulation networks.",
      "Conv2former: A simple transformer-style convnet for visual recognition.",
      "Efficient modulation for vision networks.",
      "An image is worth 16x16 words: Transformers for image recognition at scale.",
      "Swin transformer: Hierarchical vision transformer using shifted windows.",
      "Cswin transformer: A general vision transformer backbone with cross-shaped windows.",
      "Mvitv2: Improved multiscale vision transformers for classification and detection.",
      "Pyramid vision transformer: A versatile backbone for dense prediction without convolutions.",
      "Biformer: Vision transformer with bi-level routing attention.",
      "Fastervit: Fast vision transformers with hierarchical attention.",
      "Post-training quantization for vision transformer.",
      "Savit: Structure-aware vision transformer pruning via collaborative optimization.",
      "Learning efficient vision transformers via fine-grained manifold distillation.",
      "Dynamicvit: Efficient vision transformers with dynamic token sparsification.",
      "Token merging: Your vit but faster.",
      "Separable self-attention for mobile vision transformers.",
      "Fastvit: A fast hybrid vision transformer using structural reparameterization.",
      "Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer.",
      "Mobile-former: Bridging mobilenet and transformer.",
      "Attention is all you need.",
      "Edgevits: Competing light-weight cnns on mobile devices with vision transformers.",
      "Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications.",
      "Mocovit: Mobile convolutional vision transformer.",
      "Lightweight vision transformer with cross feature attention.",
      "Efficientformer: Vision transformers at mobilenet speed.",
      "Rethinking vision transformers for mobilenet size and speed.",
      "Mobilenetv4-universal models for the mobile ecosystem.",
      "Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications.",
      "Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction.",
      "Cmt: Convolutional neural networks meet vision transformers.",
      "Ghostnetv2: Enhance cheap operation with long-range attention.",
      "Mlp-mixer: An all-mlp architecture for vision."
    ],
    "citations": [
      504,
      842,
      1791,
      112006,
      4847,
      7167,
      850,
      873,
      412439,
      136296,
      470333,
      501524,
      244285,
      422787,
      732,
      1501,
      4780,
      152905,
      2577,
      704886,
      305429,
      4841,
      372973,
      787,
      336733,
      413047,
      493791,
      343099,
      94226,
      38,
      378483,
      81144,
      146623,
      906364,
      413773,
      1017440,
      136625,
      493111,
      4773,
      243823,
      7660
    ],
    "extract_rate": 0.9534883720930233,
    "article_scores": {
      "Content_Quality": 88,
      "Publication_Potential": 84,
      "Final_Score": 86.8,
      "raw_text": "Content_Quality: 88\nPublication_Potential: 84\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 78,
      "Final_Score": 81.4,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 85\nSynthesis_Analysis: 78\nFinal_Score: 81"
    },
    "overall_score": 83.56,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "4ytHislqDS",
        "forum": "4ytHislqDS",
        "number": 938,
        "title": "IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION",
        "normalized_title": "iformer integrating convnet and transformer for mobile application",
        "abstract": "We present a new family of mobile hybrid vision networks, called iFormer, with a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network, \\textit{i.e.}, ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks. Notably, iFormer achieves an impressive Top-1 accuracy of 80.4% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13, surpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios. Code and models are available at: https://github.com/ChuanyangZheng/iFormer.",
        "authors": [
          "Chuanyang Zheng"
        ],
        "decision": null,
        "pdate": 1737562212781,
        "cdate": 1726416349380
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "4ytHislqDS",
          "forum": "4ytHislqDS",
          "number": 938,
          "title": "IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION",
          "normalized_title": "iformer integrating convnet and transformer for mobile application",
          "abstract": "We present a new family of mobile hybrid vision networks, called iFormer, with a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network, \\textit{i.e.}, ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks. Notably, iFormer achieves an impressive Top-1 accuracy of 80.4% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13, surpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios. Code and models are available at: https://github.com/ChuanyangZheng/iFormer.",
          "authors": [
            "Chuanyang Zheng"
          ],
          "decision": null,
          "pdate": 1737562212781,
          "cdate": 1726416349380
        }
      ]
    }
  },
  {
    "paper_id": 4268,
    "title": "Multi-View Representation is What You Need for Point-Cloud Pre-Training",
    "abstract": "promising direction for pre-training 3D point clouds is to leverage the massive amount of data in 2D, whereas the domain gap between 2D and 3D creates a fundamental challenge. This paper proposes a novel approach to point-cloud pre-training that learns 3D representations by leveraging pre-trained 2D networks. Different from the popular practice of predicting 2D features first and then obtaining 3D features through dimensionality lifting, our approach directly uses a 3D network for feature extraction. We train the 3D feature extraction network with the help of the novel 2D knowledge transfer loss, which enforces the 2D projections of the 3D feature to be consistent with the output of pre-trained 2D networks. To prevent the feature from discarding 3D signals, we introduce the multi-view consistency loss that additionally encourages the projected 2D feature representations to capture pixel-wise correspondences across different views. Such correspondences induce 3D geometry and effectively retain 3D features in the projected 2D features. Experimental results demonstrate that our pre-trained model can be successfully transferred to various downstream tasks, including 3D shape classification, part segmentation, 3D object detection, and semantic segmentation, achieving state-of-the-art performance.",
    "introduction": "The rapid development of commercial data acquisition devices(Daneshmand et al.,2018)and point-based deep learning networks(Qi et al.,2017a;b; Choy et al.,2019)has led to a growing research interest in models that can directly process 3D point clouds without voxelization. Remarkable success has been achieved in various applications, including but not limited to object detection(Misra et al.,2021; Liu et al.,2021b; Wang et al.,2022a), segmentation(Qian et al.,2022; Tang et al.,2022; Zhao et al.,2021; Yan et al.,2021), and tracking(Qi et al.,2020; Zheng et al.,2021; Shan et al.,2021).\n\n\nDespite the significant advances in 3D point cloud processing, acquiring task-specific 3D annotations is a highly expensive and severely limited process due to the geometric complexity.\nThe shortage of data annotations highlights the need for adapting pre-training paradigms. Instead of training the deep network from randomly initialized weights, prior work suggests that pre-training the network on a relevant but different pre-task and later fine-tuning the weights using task-specific labels often leads to superior performance. In natural language processing(Devlin et al.,2018)and 2D vision(He et al.,2022; Radford et al.,2015; Doersch et al.,2015; He et al.,2020; Zhuang et al.,2021;2019), pre-trained models are the backbones of many exciting applications, such as real-time chatbots(Touvron et al.,2023; OpenAI,2023)and graphic designers(Meng et al.,2021; Wang et al.,2022b).\nHowever, pre-training on point clouds has yet to demonstrate a universal performance improvement. From-scratch training remains a common practice in 3D vision.\n\n\nInitial attempts towards 3D point-cloud pre-training primarily leverage contrastive learning(Chopra et al.,2005), especially when the point clouds are collected from indoor scenes(Xie et al.,2020; Rao et al.,2021; Liu et al.,2021a; Zhang et al.,2021; Chen et al.,2022). However, the broad application of contrastive learning-based pre-training techniques is impeded by the requirement of large batch sizes and the necessity to carefully define positive and negative pairs. In contrast to natural language processing and 2D vision, pre-training on 3D point clouds presents two unique challenges.\n\n\nFigure 1:Our model (blue) achieves state-of-the-art performance across a broad range of tasks at both the scene and shape levels. The distance to the origin indicates the task result.\n\n\nFirst, the data is extremely scarce, even without annotations. Public 3D datasets are orders of magnitude smaller than 2D image datasets.\nSecond, the lack of data annotations necessitates 3D pre-training methods to adhere to the self-supervised learning paradigm.\nWithout strong supervision, pre-task design becomes particularly crucial in effective knowledge acquisition.\n\n\nFigure2illustrates our novel approach to 3D pre-training, which is designed to address the aforementioned challenges. Our key idea is to leverage pre-trained 2D networks in the image domain to enhance the performance of 3D representation learning while bridging the 2D-3D domain gap. We begin by applying a 3D-based network to an input point cloud withM𝑀Mitalic_Mpoints, which generates a 3D feature volume represented by anM×C𝑀𝐶M\\times Citalic_M × italic_C-dimensional embedding, whereC𝐶Citalic_Cis the length of the per-point feature vectors. To compensate for data scarcity, we align the 3D feature volume with features predicted by a 2D encoder trained on hundreds of millions of images. To this end, we project the 3D feature volume into pixel embeddings and obtainH×W×C𝐻𝑊𝐶H\\times W\\times Citalic_H × italic_W × italic_Cimage-like feature maps in 2D. We then use pre-trained image encoders to extract hierarchical features from corresponding RGB images and train a similarly structured encoder on projected 2D maps to enforce feature alignment. To ensure geometrical meaningfulness, we project the same 3D feature volume into multiple camera views and perform hierarchical feature alignment in each individual view. Our pipeline is generic enough to support all major 2D pre-trained models (e.g., CLIP(Radford et al.,2021), SAM(Kirillov et al.,2023), DINOv2(Oquab et al.,2023)). In our experiments, DINOv2 exhibits the best performance.\n\n\nOne issue with the approach described above is that the 3D feature learning may overfit to the pre-trained 2D networks. Consequently, it could potentially discard 3D features that are critical for 3D recognition but not well-captured during image-based 2D pre-training. To address this issue, we introduce an auxiliary pre-task to predict 2D multi-view pixel-wise correspondences from pairs of projected 2D feature maps. As such correspondences induce 3D depth information, the learned feature representations are forced to capture 3D signals.\n\n\nOur approach is fundamentally different from existing work based on multi-view representations(Su et al.,2015; Kalogerakis et al.,2017; Kanezaki et al.,2018; Wei et al.,2020; Kundu et al.,2020; Hamdi et al.,2021;2023), which either aggregate image-based analysis results or lift 2D feature maps into 3D for understanding. Our approach employs a 3D network for feature extraction trained with the additional help ofpre-trained2D models. While multi-view based 3D understanding approaches(Su et al.,2015; Kalogerakis et al.,2017)tremendously benefit from the rich experience in 2D vision and have long dominated the design of shape classification networks when 3D data is limited, recent 3D-based understanding networks(Riegler et al.,2017; Qi et al.,2017a;b; Wang et al.,2017)are shown to outperform their multi-view counterparts. The superior performance is the result of enhanced architecture design and the clear advantage of directly learning 3D patterns for geometry understanding. In our setting, the 3D feature extraction network provides implicit regularization on the type of features extracted from pre-trained 2D models. We seek to combine the advantage of multi-view representations and the geometric capacity of 3D networks, while leveraging the rich knowledge embedded in pre-trained 2D networks.\n\n\nIn summary, we make the following contributions:\n\n\n•We formulate point-cloud pre-training as learning a multi-view consistent 3D feature volume.•To compensate for data scarcity, we leverage pre-trained 2D image-based models to supervise 3D pre-training through perspective projection and hierarchical feature-based 2D knowledge transfer.•To prevent overfitting to pre-trained 2D networks, we develop an auxiliary pre-task where the goal is to predict the multi-view pixel-wise correspondences from the 2D pixel embeddings.•We conduct extensive experiments to demonstrate the effectiveness of our approach across a wide range of downstream tasks, including shape classification, part segmentation, 3D detection, and semantic segmentation, achieving consistent improvement over baselines (Figure1).\n\n",
    "related_work": "Point-cloud pre-trainingThe success of point-based deep neural networks demonstrates the potential for machine learning models to directly perceive point clouds. Point-based pre-training refers to the practice ofpre-trainingthe point-based prediction network on one or morepre-tasksbefore fine-tuning the weights ondownstream tasks. The expectation is that the knowledge about pre-tasks will be transferred to downstream tasks, and the network can achieve better performance than random parameter initialization.\nWe refer readers to(Xiao et al.,2023)for a comprehensive survey of the field, in which we highlight one important observation, stating that point-based pre-training “still lags far behind as compared with its\ncounterparts”, and training from-scratch “is still the prevalent approach”. The challenges and opportunities call for innovations in point-based pre-training methods.Shape-level pre-training.Self-reconstruction is a popular pre-task at the shape level, where the network encodes the given point cloud as representation vectors capable of being decoded back to the original input data(Yang et al.,2018; Sauder & Sievers,2019). To increase the pre-task difficulty, it is common for self-reconstruction-based methods to randomly remove a certain percentage of the points from the input(Wang et al.,2021; Pang et al.,2022; Zhang et al.,2022a; Yu et al.,2022; Yan et al.,2023b;a). In addition to self-reconstruction,(Rao et al.,2020)develop a multi-task approach that unifies contrastive learning, normal estimation, and self-reconstruction into the same pipeline. There are also attempts to connect 3D pre-training to 2D(Afham et al.,2022; Xu et al.,2021; Dong et al.,2022; Zhang et al.,2022b), or to explore the inverse relationship(Hou et al.,2023). For example,(Dong et al.,2022)utilize pre-trained 2D transformers as cross-modal teachers.(Zhang et al.,2022b)leverage self-supervised pre-training and masked autoencoding to obtain high-quality 3D features from 2D pre-trained models.\nHowever, these methods focus on synthetic objects, leading to significant performance degradation in downstream tasks using real data from indoor scenes.Scene-level pre-training.Existing scene-level pre-training methods focus on exploiting the contrastive learning paradigm. As the first pre-training method with demonstrated success at the scene level, PointContrast(Xie et al.,2020)defines the contrastive loss using pairs of 3D points across multiple views. RandomRoom(Rao et al.,2021)exploits synthetic datasets and defines the contrastive loss using pairs of CAD objects instead. Experimentally, RandomRoom additionally confirms that synthetic shape-level pre-training is beneficial to real-world scene-level downstream tasks. DepthContrast(Zhang et al.,2021)simplifies PointContrast by proposing an effective pre-training method requiring only single-view depth scans. 4DContrast(Chen et al.,2022)further includes temporal dynamics in the contrastive formulation, where the pre-training data comprises sequences of synthetic scenes with objects in different locations. In contrast to these approaches, we focus on utilizing pre-trained 2D networks to boost the performance of 3D pre-training without leveraging contrastive learning.\n\n",
    "origin_citations": [
      "Unsupervised point cloud representation learning with deep neural networks: A survey.",
      "Foldingnet: Point cloud auto-encoder via deep grid deformation.",
      "Self-supervised deep learning on point clouds by reconstructing space.",
      "Unsupervised point cloud pre-training via occlusion completion.",
      "Masked autoencoders for point cloud self-supervised learning.",
      "Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training.",
      "Point-bert: Pre-training 3d point cloud transformers with masked point modeling.",
      "Implicit autoencoder for point-cloud self-supervised representation learning.",
      "3d feature prediction for masked-autoencoder-based point cloud pretraining.",
      "Global-local bidirectional reasoning for unsupervised representation learning of 3d point clouds.",
      "Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding.",
      "Image2point: 3d point-cloud understanding with 2d image pretrained models.",
      "Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning?",
      "Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders.",
      "Mask3d: Pre-training 2d vision transformers by learning masked 3d priors.",
      "Pointcontrast: Unsupervised pre-training for 3d point cloud understanding.",
      "Randomrooms: unsupervised pre-training from synthetic shapes and randomized layouts for 3d object detection.",
      "Self-supervised pretraining of 3d features on any point-cloud.",
      "4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding."
    ],
    "citations": [
      708745,
      286541,
      1532,
      357278,
      516828,
      145768,
      337259,
      256176,
      17961,
      252119,
      519427,
      194894,
      140642,
      221158,
      705135,
      521402,
      454618,
      39396,
      152223
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 71,
      "Final_Score": 81.5,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 92\nSynthesis_Analysis: 71\nFinal_Score: 81"
    },
    "overall_score": 83.5,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2024/Conference",
        "venue": "ICLR",
        "year": "2024",
        "openreview_id": "imZcqOrbig",
        "forum": "imZcqOrbig",
        "number": 716,
        "title": "Multi-View Representation is What You Need for Point-Cloud Pre-Training",
        "normalized_title": "multi view representation is what you need for point cloud pre training",
        "abstract": "A promising direction for pre-training 3D point clouds is to leverage the massive amount of data in 2D, whereas the domain gap between 2D and 3D creates a fundamental challenge. This paper proposes a novel approach to point-cloud pre-training that learns 3D representations by leveraging pre-trained 2D networks. Different from the popular practice of predicting 2D features first and then obtaining 3D features through dimensionality lifting, our approach directly uses a 3D network for feature extraction. We train the 3D feature extraction network with the help of the novel 2D knowledge transfer loss, which enforces the 2D projections of the 3D feature to be consistent with the output of pre-trained 2D networks. To prevent the feature from discarding 3D signals, we introduce the multi-view consistency loss that additionally encourages the projected 2D feature representations to capture pixel-wise correspondences across different views. Such correspondences induce 3D geometry and effectively retain 3D features in the projected 2D features. Experimental results demonstrate that our pre-trained model can be successfully transferred to various downstream tasks, including 3D shape classification, part segmentation, 3D object detection, and semantic segmentation, achieving state-of-the-art performance.",
        "authors": [
          "Siming Yan",
          "Chen Song",
          "Youkang Kong",
          "Qixing Huang"
        ],
        "decision": null,
        "pdate": 1705410801916,
        "cdate": 1694884629883
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2024/Conference",
          "venue": "ICLR",
          "year": "2024",
          "openreview_id": "imZcqOrbig",
          "forum": "imZcqOrbig",
          "number": 716,
          "title": "Multi-View Representation is What You Need for Point-Cloud Pre-Training",
          "normalized_title": "multi view representation is what you need for point cloud pre training",
          "abstract": "A promising direction for pre-training 3D point clouds is to leverage the massive amount of data in 2D, whereas the domain gap between 2D and 3D creates a fundamental challenge. This paper proposes a novel approach to point-cloud pre-training that learns 3D representations by leveraging pre-trained 2D networks. Different from the popular practice of predicting 2D features first and then obtaining 3D features through dimensionality lifting, our approach directly uses a 3D network for feature extraction. We train the 3D feature extraction network with the help of the novel 2D knowledge transfer loss, which enforces the 2D projections of the 3D feature to be consistent with the output of pre-trained 2D networks. To prevent the feature from discarding 3D signals, we introduce the multi-view consistency loss that additionally encourages the projected 2D feature representations to capture pixel-wise correspondences across different views. Such correspondences induce 3D geometry and effectively retain 3D features in the projected 2D features. Experimental results demonstrate that our pre-trained model can be successfully transferred to various downstream tasks, including 3D shape classification, part segmentation, 3D object detection, and semantic segmentation, achieving state-of-the-art performance.",
          "authors": [
            "Siming Yan",
            "Chen Song",
            "Youkang Kong",
            "Qixing Huang"
          ],
          "decision": null,
          "pdate": 1705410801916,
          "cdate": 1694884629883
        }
      ]
    }
  },
  {
    "paper_id": 2103,
    "title": "Graph Generation with Diffusion Mixture",
    "abstract": "Generation of graphs is a major challenge for real-world tasks that require understanding the complex nature of their non-Euclidean structures. Although diffusion models have achieved notable success in graph generation recently, they are ill-suited for modeling the topological properties of graphs since learning to denoise the noisy samples does not explicitly learn the graph structures to be generated. To tackle this limitation, we propose a generative framework that models the topology of graphs by explicitly learning the final graph structures of the diffusion process.\nSpecifically, we design the generative process as a mixture of endpoint-conditioned diffusion processes which is driven toward the predicted graph that results in rapid convergence.\nWe further introduce a simple parameterization of the mixture process and develop an objective for learning the final graph structure, which enables maximum likelihood training.\nThrough extensive experimental validation on general graph and 2D/3D molecule generation tasks, we show that our method outperforms previous generative models, generating graphs with correct topology with both continuous (e.g. 3D coordinates) and discrete (e.g. atom types) features. Our code is available at https://github.com/harryjo97/GruM.",
    "introduction": "Generation of graph-structured data has emerged as a crucial task for real-world problems such as drug discovery(Simonovsky & Komodakis,2018), protein design(Ingraham et al.,2019), and program synthesis(Brockschmidt et al.,2019). To tackle the challenge of learning the underlying distribution of graphs, deep generative models have been proposed, including models based on generative adversarial networks (GANs)(De Cao & Kipf,2018; Martinkus et al.,2022), recurrent neural networks (RNNs)(You et al.,2018), and variational autoencoders (VAEs)(Jin et al.,2018).\n\n\nRecently, diffusion models have achieved state-of-the-art performance on the generation of graph-structured data(Niu et al.,2020; Jo et al.,2022; Hoogeboom et al.,2022). These models learn the generation process as the time reversal of the forward process, which corrupts the graphs by gradually adding noise that destroys its topological properties. Since the generative process is derived from the unknown score function(Song et al.,2021)or noise(Ho et al.,2020), existing graph diffusion models aim to estimate them in order to denoise the data from noise, which are commonly referred to as thedenoising diffusion models(Figure1(a)).\n\n\nDespite their success, learning the score or noise is fundamentally ill-suited for the generation of graphs. In contrast to other types of data such as images, the key to generating valid graphs is accurately modeling the discrete structures that determine the topological properties such as connectivity or clusteredness. However, learning the score or noise does not explicitly model these features, as it aims to gradually denoise the corrupted structures. Thereby it is challenging for the diffusion models to recover the topological properties, which leads to failure cases even for small graphs. A way to more accurately generate graphs with correct topology would be directly learning the final graph and its structure, instead of learning how to denoise a noisy version of the original graphs.\n\n\nHowever, predicting the final graph structure of the diffusion process is difficult since the prediction would be highly inaccurate in the early steps of the diffusion process, and such an inaccurate prediction may lead the process in the wrong direction resulting in invalid results. Few existing works(Hoogeboom et al.,2021; Austin et al.,2021; Vignac et al.,2023)based on denoising diffusion models aim to predict the probability of the final states by parameterizing the denoising process, but it is only applicable to categorical data with a finite number of states and thus cannot generate graphs with continuous features, which is not suitable for tasks such as 3D molecule generation.\n\n\nTo address these limitations of existing graph diffusion models, we propose a novel framework that explicitly models the graph topology, by learning the prediction of the resulting graph of the generative process which is represented as a weighted mean of graph data (Figure1(b)).\nSpecifically, we construct a diffusion process via the mixture of endpoint-conditioned Ornstein-Uhlenbeck processes for which the drift drives the process in the direction of the predicted graph, differing from the denoising diffusion process used in previous works (Section3.1).\nIn order to model the mixture of the diffusion process, we develop a simple parameterization of the graph generative model with respect to the prediction of the final graph.\nWe further derive an efficient training objective for learning the graph prediction, which guarantees to maximize the likelihood of our generative model (Section3.2).\nThanks to its ability to capture accurate graph structures, our framework achieves fast convergence to the correct graph topology in an early sampling step (Figure1(c) and Figure3(Right)).\n\n\nWe experimentally validate our method on diverse real-world graph generation tasks. We first validate it on general graph generation benchmarks with synthetic and real-world graphs, on which it outperforms previous deep graph generative models including graph diffusion models, by being able to generate valid graphs with correct topologies. We further validate our method on 2D and 3D molecule generation tasks to demonstrate its ability to generate graphs with both the continuous and discrete features, on which ours generates a significantly larger number of valid and stable molecules compared to the state-of-the-art baselines.\nOur main contributions can be summarized as follows:•We observe that previous diffusion models cannot accurately model the graph structures as they learn to denoise at each step without considering the topology of the graphs to be generated.•To fix such a myopic behavior of previous diffusion models, we propose a new graph generation framework that captures the graph structures by directly predicting the final graph of the diffusion process modeled by a mixture of endpoint-conditioned diffusion processes.•We develop a simple parameterization of the graph generative model for modeling the mixture process and present a simulation-free training objective for graph prediction.•Our method significantly outperforms previous graph diffusion models on the generation of diverse real and synthetic graphs, as well as on 2D/3D molecule generation tasks, by being able to generate graphs with accurate topologies, and both the discrete and continuous features.\n\n\nFigure 1:Illustration of the graph generative process.(a: Denoising diffusion model, b: GruM (ours), c: Graph mixture) For GruM, we design the generative process as a mixture of endpoint-conditioned diffusion processes (Eq. (3)), namely the OU bridge mixture (Eq. (6)), which is driven toward the graph mixture (green) by its drift (Eq. (8)). Our GruM in (b) successfully generates graphs with valid topology by predicting the final result via learning the graph mixture as a weighted mean of data (Eq. (1)). The predicted graph of GruM converges in an early stage to the correct topology as visualized in (c).\nIn contrast, previous denoising diffusion models in (a) often fail to capture the correct topology as they learn the score or noise for denoising (red), without explicit knowledge of final graph structure.\n\n",
    "related_work": "Diffusion ModelsDiffusion models have been shown to successfully generate high-quality samples from diverse data domains such as images(Dhariwal & Nichol,2021; Saharia et al.,2022)and videos(Ho et al.,2022). Despite their success, existing diffusion models for graphs(Niu et al.,2020; Jo et al.,2022)often fail to generate graphs with correct structures since they learn to estimate the score or noise for the denoising process which does not explicitly capture the final graph and its structure. To address these limitations, we propose a graph diffusion framework that models the generative process as a mixture of diffusion processes, which learns to predict the final graph with valid topology instead of predicting the denoising function at each step. This promotes our generative process to be driven toward the prediction of the final graph, resulting in generation of valid graphs with correct topology.\n\n\nDiffusion Bridge ProcessA line of recent works has improved the generative framework of diffusion models by leveraging the diffusion bridge processes, i.e., processes conditioned to the endpoints. Schrödinger Bridge(Bortoli et al.,2021b; Chen et al.,2022)aims to find both the forward and the backward process that transforms two distributions back and forth using iterative proportional fittings that require heavy computations. More recent works(Peluchetti,2021; Wu et al.,2022; Liu et al.,2023; Jo & Hwang,2023)consider learning the generation process as a mixture of diffusion processes instead of reversing the noising process as in denoising diffusion models, which we describe in detail in AppendixA.11.\nHowever, previous works aim to approximate the drifts of the diffusion processes which cannot accurately capture the discrete structure of graphs as it does not explicitly learn the graph structures to be generated. Moreover, learning the drift could be problematic as the drift of the diffusion process diverges near the terminal time. Instead, we present a new approach to parameterizing the mixture process with the prediction of the final graph which allows it to model valid graph topology.\n\n\nGraph Generative ModelsDeep generative models for graphs either generate nodes and edges in an autoregressive manner or all the nodes and edges at once using VAE(Jin et al.,2018), RNN(You et al.,2018), normalizing flow(Zang & Wang,2020; Shi et al.,2020; Luo et al.,2021), and GAN(De Cao & Kipf,2018; Martinkus et al.,2022). However, these models show poor performance due to restrictive model architectures for modeling the likelihood or their inability to model the permutation equivariant nature of graphs. Recently, diffusion models for graphs(Jo et al.,2022; Hoogeboom et al.,2022; Vignac et al.,2023)have made large progress, but either fail to capture the graph topology or are not applicable to general tasks due to the architectural restriction of the framework.\nIn our work, we introduce a diffusion framework that predicts the final graph structure instead of denoising noisy graphs. Our method largely outperforms existing models(Jo et al.,2022; Vignac et al.,2023; Hoogeboom et al.,2022)on generation tasks including general graphs as well as 2D and 3D molecules.\n\n",
    "origin_citations": [
      "Diffusion models beat gans on image synthesis.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Video diffusion models.",
      "Permutation invariant graph generation via score-based generative modeling.",
      "Score-based generative modeling of graphs via the system of stochastic differential equations.",
      "Diffusion schrödinger bridge with applications to score-based generative modeling.",
      "Likelihood training of schrödinger bridge using forward-backward sdes theory.",
      "Non-denoising forward-time diffusions.",
      "Diffusion-based molecule generation with informative prior bridges.",
      "Learning diffusion bridges on constrained domains.",
      "Generative modeling on manifolds through mixture of riemannian diffusion processes.",
      "Junction tree variational autoencoder for molecular graph generation.",
      "Graphrnn: Generating realistic graphs with deep auto-regressive models.",
      "Moflow: an invertible flow model for generating molecular graphs.",
      "Graphaf: a flow-based autoregressive model for molecular graph generation.",
      "Graphdf: A discrete flow model for molecular graph generation.",
      "Molgan: An implicit generative model for small molecular graphs.",
      "SPECTRE: spectral conditioning helps to overcome the expressivity limits of one-shot graph generators.",
      "Equivariant diffusion for molecule generation in 3d.",
      "Digress: Discrete denoising diffusion for graph generation."
    ],
    "citations": [
      165,
      5926,
      5927,
      384900,
      90229,
      196673,
      513356,
      347525,
      230716,
      320990,
      1652,
      3996,
      169536,
      3983,
      120530,
      7214,
      65617,
      9005,
      176219
    ],
    "extract_rate": 0.95,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 88,
      "Final_Score": 85.89999999999999,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 88\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 83,
      "Synthesis_Analysis": 84,
      "Final_Score": 81.8,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 83\nSynthesis_Analysis: 84\nFinal_Score: 82"
    },
    "overall_score": 83.44,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "cZTFxktg23",
        "forum": "cZTFxktg23",
        "number": 1170,
        "title": "Graph Generation with Diffusion Mixture",
        "normalized_title": "graph generation with diffusion mixture",
        "abstract": "Generation of graphs is a major challenge for real-world tasks that require understanding the complex nature of their non-Euclidean structures. Although diffusion models have achieved notable success in graph generation recently, they are ill-suited for modeling the topological properties of graphs since learning to denoise the noisy samples does not explicitly learn the graph structures to be generated. To tackle this limitation, we propose a generative framework that models the topology of graphs by explicitly learning the final graph structures of the diffusion process. Specifically, we design the generative process as a mixture of endpoint-conditioned diffusion processes which is driven toward the predicted graph that results in rapid convergence. We further introduce a simple parameterization of the mixture process and develop an objective for learning the final graph structure, which enables maximum likelihood training. Through extensive experimental validation on general graph and 2D/3D molecule generation tasks, we show that our method outperforms previous generative models, generating graphs with correct topology with both continuous (e.g. 3D coordinates) and discrete (e.g. atom types) features. Our code is available at https://github.com/harryjo97/GruM.",
        "authors": [
          "Jaehyeong Jo",
          "Dongki Kim",
          "Sung Ju Hwang"
        ],
        "decision": null,
        "pdate": 1714610282048,
        "cdate": 1705895572923
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "cZTFxktg23",
          "forum": "cZTFxktg23",
          "number": 1170,
          "title": "Graph Generation with Diffusion Mixture",
          "normalized_title": "graph generation with diffusion mixture",
          "abstract": "Generation of graphs is a major challenge for real-world tasks that require understanding the complex nature of their non-Euclidean structures. Although diffusion models have achieved notable success in graph generation recently, they are ill-suited for modeling the topological properties of graphs since learning to denoise the noisy samples does not explicitly learn the graph structures to be generated. To tackle this limitation, we propose a generative framework that models the topology of graphs by explicitly learning the final graph structures of the diffusion process. Specifically, we design the generative process as a mixture of endpoint-conditioned diffusion processes which is driven toward the predicted graph that results in rapid convergence. We further introduce a simple parameterization of the mixture process and develop an objective for learning the final graph structure, which enables maximum likelihood training. Through extensive experimental validation on general graph and 2D/3D molecule generation tasks, we show that our method outperforms previous generative models, generating graphs with correct topology with both continuous (e.g. 3D coordinates) and discrete (e.g. atom types) features. Our code is available at https://github.com/harryjo97/GruM.",
          "authors": [
            "Jaehyeong Jo",
            "Dongki Kim",
            "Sung Ju Hwang"
          ],
          "decision": null,
          "pdate": 1714610282048,
          "cdate": 1705895572923
        }
      ]
    }
  },
  {
    "paper_id": 16704,
    "title": "BiLLM: Pushing the Limit of Post-Training Quantization for LLMs",
    "abstract": "Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effectivebinary residual approximationstrategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose anoptimal splitting searchto group and binarize them accurately. BiLLM, for the first time, achieves high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only1.08-bitweights across various LLM families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of a 7-billion LLM within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency.\nOur code is available athttps://github.com/Aaronhuang-778/BiLLM.",
    "introduction": "Recently, large language models (LLMs) based on transformers(Vaswani et al.,2017)have garnered significant attention in natural language processing. Pre-trained LLMs such as OPT(Zhang et al.,2022)and LLaMA(Touvron et al.,2023a), have demonstrated excellent performance across a range of evaluation benchmarks. However, LLMs pose substantial challenges in deployment on memory-constrained devices due to their immense parameter size and computation requirements. For instance, the widely-used LLaMA2-70B(Touvron et al.,2023b)model, with its 70 billion parameters, requires 150 GB of storage in half-precision (FP16) format. This necessitates at least two A100 GPUs, each with 80 GB of storage space, for inference.\n\n\nFigure 1:Perplexity of LLaMA-13B on WikiText2 under different bit-widths. Round-to-nearest (RTN), GPTQ, and PB-LLM (10% weight of INT8) suffer accuracy loss at ultra-low bits, facing the sharply increasing perplexity (↓↓\\downarrow↓).BiLLMdemonstrates exceptional performance under binarization.\n\n\nModel quantization has emerged as a highly effective technology for compressing neural networks, thereby reducing the model size of LLMs and substantially saving GPU memory consumption(Dettmers et al.,2022). Current quantization techniques primarily fall into Quantization-Aware Training (QAT) and Post-Training Quantization (PTQ). QAT involves fine-tuning and retraining during the quantization process, while PTQ significantly streamlines the computation by eliminating back-propagation, enabling a faster quantization process and promoting the practicality of quantization(Frantar et al.,2022;Shang et al.,2023;Lin et al.,2023). Given the deep structures and numerous parameters of LLMs, PTQ stands out for its ability to rapidly perform the quantization process, especially on time and resource-constrained scenarios(Zhu et al.,2023).\n\n\nDespite the success of previous PTQ methods in 8-bit and 4-bit quantization(Dettmers et al.,2022,2023b;Frantar et al.,2022;Xiao et al.,2023;Frantar & Alistarh,2022), the expanding size of LLMs demands more aggressive quantization approaches(Shang et al.,2023). Neural network binarization, which reduces the weight bit-width to only 1 bit, is a promising approach(Helwegen et al.,2019;Qin et al.,2020,2023). However, as depicted in Figure1, current advanced PTQ methods for LLMs exhibit a performance collapse under ultra-low bit (⩽\\leqslant⩽3 bits) quantization. This phenomenon can be attributed to the significant difference between quantized and original weights. Even the recent binary PTQ method for LLMs, PB-LLM(Shang et al.,2023), only maintains a perplexity metric of around 800 with an average weight of 1.7 bits. This observation underscores the challenges existing PTQ methods face in promoting the weight binarization of LLMs.\n\n\nIn pursuit of this goal, we conducted an empirical study to analyze the distribution of pre-trained weights in LLMs. The findings derived from this study are presented in AppendixG, revealing two key observations:•The second-order Hessian matrix of weights demonstrates anexceptionally long-tail distributionand is often used to measure the importance of weight elements in neural networks(LeCun et al.,1989;Dong et al.,2019). As depicted in Figure2, a small fraction of weights elements possesses significantly high Hessian values, substantially influencing the layer output. In contrast, most Hessian values cluster around 0.•The density distribution of weight magnitudes in LLMs follows abell-shaped pattern. This bell-shaped distribution exhibits a significant resemblance to both the Gaussian or Laplace distribution in terms of its characteristics(Blundell et al.,2015). Figure2illustrates that most weight values cluster around zero with a non-uniform bell-shaped distribution.\n\n\nFigure 2:The Hessian metrics (sensitivity) and magnitude (value) of weights in LLMs.\nThe weights of different layers in LLMs are characterized by bell-shaped distribution, accompanied by a few salient values.\n\n\nThe above implies:\na) A minority of weights play an important role in LLMs, whereas the majority of weights exhibit characteristics of redundancy(Shang et al.,2023;Dettmers et al.,2023b);\nb) With the most aggressive bit-width, binarization incurs most severe error among quantization under bell-shaped distributions in LLMs(Jacob et al.,2018).\n\n\nMotivated by the above observation, we propose a novel 1-bit PTQ framework for LLMs, namelyBiLLM, incorporating two core designs to achieve highly accurate weight binarization.\nFirst, guided by the Hessian-based metric, we select the salient weights structurally (Figure3upper-right) to achieve a trade-off between accuracy and storage savings and develop a residual approximation to maximize the restoration of salient weights with highly dynamic range.\nSecond, for the remaining non-salient weights (Figure3lower-right), we design an optimal splitting binarization strategy, where a meticulous search process is applied to determine an optimal break-point for weight distribution and binarization of the segments is then processed separately to minimize binarization errors.\nMoreover,BiLLMincorporates error compensation on a block-wise basis by default following existing common practices(Frantar et al.,2022;Shang et al.,2023), which further reduces quantization error.\n\n\nExtensive experiments demonstrate thatBiLLMachieve the state-of-the-art (SOTA) performance for LLMs across multiple LLM families on various evaluation metrics, and first achieves extremely compact 1.07∼similar-to\\sim∼1.11 bit-width in average for the PTQ binarization. For example, on the Wikitext2(Merity et al.,2016)metric,BiLLMachieved perplexities of 8.49 and 8.41 with only 1.08-bit weights on LLaMA-65B(Touvron et al.,2023a)and LLaMA2-70B(Touvron et al.,2023b), respectively, even surpassing the 9.34 performance of the FP16 OPT-66B(Zhang et al.,2022).\n\n",
    "related_work": "Figure 3:Schematic of the PTQ binarization framework for LLMs. The left side shows the structure of the Transformer block after binarization. The right side shows the binarization process ofBiLLM, which consists of two parts,Residual Approximationfor salient weights andBell-shaped Splittingfor non-salient weights.\n\n\n2.1Large Language Model QuantizationQuantization maps high-precision parameters to a discrete range. This method, which compresses parameters without altering the model structure, effectively reduces the storage and computational overhead of deep neural networks. Recent work has successfully applied QAT and PTQ to LLMs. QAT, through a quantization-aware retraining strategy, better preserves the performance of quantized models. LLM-QAT(Liu et al.,2023)addressed data barrier issues in QAT training through data-free distillation. However, for LLMs with extremely large parameter sizes, the cost of retraining is prohibitively high and inefficient. Therefore, techniques such as QLoRA(Dettmers et al.,2023a)focus on parameter-efficient fine-tuning (PEFT) methods for quantizing LLMs, enhancing the efficiency of QAT. Nevertheless, even these efficient fine-tuning quantization strategies require over 24 hours of GPU time.Therefore, the PTQ strategy has become a significant option for quantizing LLMs efficiently. Works like BRECQ(Li et al.,2021), ZerqQuant(Yao et al.,)and LLM.int8()(Dettmers et al.,2022)enhance quantization accuracy by adding additional grouping labels for custom quantization blocks. Other studies adopt a feature segmentation strategy, such as PB-LLM(Shang et al.,2023)and SpQR(Dettmers et al.,2023b). They preserve the bit-width of outlier features or those with higher quantization errors to FP16 or INT8, mitigating the precision loss due to quantization. GPTQ(Frantar et al.,2022)employs a more precise quantization framework, reducing the block quantization errors of LLMs through Hessian-based second-order error compensation(Frantar & Alistarh,2022), achieving commendable performance in low-bits (4 bits) quantization. Smoothquant(Xiao et al.,2023)introduces a strategy of scaling weight and activation outliers to simplify quantization. Subsequently, AWQ(Lin et al.,2023)and OWQ(Lee et al.,2023)also proposed scale transformations of more crucial weight channels for activation features, preserving their information representation capacity.\n\n\n2.2Network BinarizationBinarized compression can quantize parameters to only 1 bit, expressed as±plus-or-minus\\pm±1. In forward propagation, the sign function is used to binarize the original parameter tensor:𝐖b=α⋅sign⁡(𝐖f),subscript𝐖𝑏⋅𝛼signsubscript𝐖𝑓\\mathbf{W}_{b}=\\alpha\\cdot\\operatorname{sign}(\\mathbf{W}_{f}),bold_W start_POSTSUBSCRIPT italic_b end_POSTSUBSCRIPT = italic_α ⋅ roman_sign ( bold_W start_POSTSUBSCRIPT italic_f end_POSTSUBSCRIPT ) ,(1)sign⁡(x)={1ifx≥0,−1others.sign𝑥cases1ifx≥01others\\operatorname{sign}(x)=\\begin{cases}1&\\text{if $x\\geq 0$},\\\\\n-1&\\text{others}.\\end{cases}roman_sign ( italic_x ) = { start_ROW start_CELL 1 end_CELL start_CELL if italic_x ≥ 0 , end_CELL end_ROW start_ROW start_CELL - 1 end_CELL start_CELL others . end_CELL end_ROW(2)where𝐖f∈ℝn×msubscript𝐖𝑓superscriptℝ𝑛𝑚\\mathbf{W}_{f}\\in\\mathbb{R}^{n\\times m}bold_W start_POSTSUBSCRIPT italic_f end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_n × italic_m end_POSTSUPERSCRIPTis the full precision weight and𝐖b∈ℝn×msubscript𝐖𝑏superscriptℝ𝑛𝑚\\mathbf{W}_{b}\\in\\mathbb{R}^{n\\times m}bold_W start_POSTSUBSCRIPT italic_b end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_n × italic_m end_POSTSUPERSCRIPTis the binarized output.n𝑛nitalic_nandm𝑚mitalic_mrepresent the size of the weight matrix.α𝛼\\alphaitalic_αdenotes the scaling factor(Courbariaux et al.,2016). Binarization usually uses the channel-wise scale(Rastegari et al.,2016;Qin et al.,2023), soα∈ℝn𝛼superscriptℝ𝑛\\alpha\\in\\mathbb{R}^{n}italic_α ∈ blackboard_R start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPT.Most previous binarization works adopt a framework based on QAT for quantization(Qin et al.,2023). Straight through estimator (STE)(Bengio et al.,2013)is deployed to address the issue of gradient vanishing caused by thesign⁡(⋅)sign⋅\\operatorname{sign}(\\cdot)roman_sign ( ⋅ )function. Binary Weight Network (BWN)(Rastegari et al.,2016)was initially proposed for executing neural network computations by binarizing weights and using full-precision activations, while XNOR-Net(Rastegari et al.,2016)extends this approach by binarizing both weights and activations. Both methods minimize quantization errors through dynamic searching ofα𝛼\\alphaitalic_α. DoReFa-Net(Zhou et al.,2016)further expands upon XNOR-Net, employing quantized gradients to accelerate network training. Group segmentation is also applied in binarization tasks, with Syq(Faraone et al.,2018)utilizing network weight to the small size of groups for minimizing binarization errors.Based on the successful application of binarization in Transformers(Wang et al.,2023)and Bert(Qin et al.,2022), we believe that the binarization of LLMs is filled with potential. PB-LLM(Shang et al.,2023)investigates the impact of binarized QAT and PTQ strategies on LLMs, but it is necessary to retain a significant proportion (over 30%) of the weights at 8 bits to enable LLMs to produce reasonable answers. Due to the presence of a large amount of INT8, LLMs still have a relatively high average bit-width. To address this issue, we proposedBiLLM, which aims to push the limit of PTQ binarization for LLMs.\n\n",
    "origin_citations": [
      "Llm-qat: Data-free quantization aware training for large language models.",
      "Qlora: Efficient finetuning of quantized llms.",
      "Brecq: Pushing the limit of post-training quantization by block reconstruction.",
      "Efficient and affordable post-training quantization for large-scale transformers, 2022.",
      "Llm. int8 (): 8-bit matrix multiplication for transformers at scale.",
      "Pb-llm: Partially binarized large language models.",
      "Spqr: A sparse-quantized representation for near-lossless llm weight compression.",
      "Gptq: Accurate post-training quantization for generative pre-trained transformers.",
      "Optimal brain compression: A framework for accurate post-training quantization and pruning.",
      "Smoothquant: Accurate and efficient post-training quantization for large language models.",
      "Awq: Activation-aware weight quantization for llm compression and acceleration.",
      "Owq: Lessons learned from activation outliers for weight quantization in large language models.",
      "Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1.",
      "Xnor-net: Imagenet classification using binary convolutional neural networks.",
      "Bibench: Benchmarking and analyzing network binarization.",
      "Estimating or propagating gradients through stochastic neurons for conditional computation.",
      "Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients.",
      "Syq: Learning symmetric quantization for efficient deep neural networks.",
      "Bitnet: Scaling 1-bit transformers for large language models.",
      "Bibert: Accurate fully binarized bert."
    ],
    "citations": [
      257203,
      2253,
      118706,
      888003,
      477369,
      306939,
      945903,
      67497,
      245034,
      308051,
      1088952,
      481944,
      851,
      212999,
      8375,
      8241,
      44580,
      315383,
      516929
    ],
    "extract_rate": 0.95,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 76,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 78,
      "Final_Score": 81.4,
      "raw_text": "Content_Coherence: 76\nCitation_Quality: 88\nSynthesis_Analysis: 78\nFinal_Score: 81"
    },
    "overall_score": 83.44,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "qOl2WWOqFg",
        "forum": "qOl2WWOqFg",
        "number": 2137,
        "title": "BiLLM: Pushing the Limit of Post-Training Quantization for LLMs",
        "normalized_title": "billm pushing the limit of post training quantization for llms",
        "abstract": "Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effective binary residual approximation strategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose an optimal splitting search to group and binarize them accurately. BiLLM, for the first time, achieves high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only 1.08-bit weights across various LLM families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of a 7-billion LLM within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency. Our code is available at https://github.com/Aaronhuang-778/BiLLM .",
        "authors": [
          "Wei Huang",
          "Yangdong Liu",
          "Haotong Qin",
          "Ying Li",
          "Shiming Zhang",
          "Xianglong Liu",
          "Michele Magno",
          "XIAOJUAN QI"
        ],
        "decision": null,
        "pdate": 1714610305268,
        "cdate": 1706288456538
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "qOl2WWOqFg",
          "forum": "qOl2WWOqFg",
          "number": 2137,
          "title": "BiLLM: Pushing the Limit of Post-Training Quantization for LLMs",
          "normalized_title": "billm pushing the limit of post training quantization for llms",
          "abstract": "Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effective binary residual approximation strategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose an optimal splitting search to group and binarize them accurately. BiLLM, for the first time, achieves high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only 1.08-bit weights across various LLM families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of a 7-billion LLM within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency. Our code is available at https://github.com/Aaronhuang-778/BiLLM .",
          "authors": [
            "Wei Huang",
            "Yangdong Liu",
            "Haotong Qin",
            "Ying Li",
            "Shiming Zhang",
            "Xianglong Liu",
            "Michele Magno",
            "XIAOJUAN QI"
          ],
          "decision": null,
          "pdate": 1714610305268,
          "cdate": 1706288456538
        }
      ]
    }
  },
  {
    "paper_id": 27002,
    "title": "OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary Understanding",
    "abstract": "This paper introduces OpenGaussian, a method based on 3D Gaussian Splatting (3DGS) capable of 3D point-level open vocabulary understanding. Our primary motivation stems from observing that existing 3DGS-based open vocabulary methods mainly focus on 2D pixel-level parsing. These methods struggle with 3D point-level tasks due to weak feature expressiveness and inaccurate 2D-3D feature associations.\nTo ensure robust feature presentation and 3D point-level understanding, we first employ SAM masks without cross-frame associations to train instance features with 3D consistency. These features exhibit both intra-object consistency and inter-object distinction. Then, we propose a two-stage codebook to discretize these features from coarse to fine levels. At the coarse level, we consider the positional information of 3D points to achieve location-based clustering, which is then refined at the fine level.\nFinally, we introduce an instance-level 3D-2D feature association method that links 3D points to 2D masks, which are further associated with 2D CLIP features. Extensive experiments, including open vocabulary-based 3D object selection, 3D point cloud understanding, click-based 3D object selection, and ablation studies, demonstrate the effectiveness of our proposed method. The source code is available at our project pagehttps://3d-aigc.github.io/OpenGaussian.",
    "introduction": "The recently proposed neural rendering method 3D Gaussian Splatting (3DGS)[19]has rapidly gained popularity and is being widely applied in various areas such as 3D reconstruction[18,10,38], 4D reconstruction[23,28,42,47], generation[41,25,49], and understanding[44,51]. This is primarily due to its fast training, real-time rendering capabilities, and explicit point-based representation. Within the realm of 3D vision learning, 3D scene understanding has embraced the 3DGS framework to achieve an integrated process encompassing explicit reconstruction, novel view synthesis, and semantic understanding. Incorporating open vocabulary for 3D understanding is considered a more promising, practical, and natural approach for intelligent agent understanding, interaction, and decision-making. In this paper, we specifically concentrate on point-level open-vocabulary 3D scene understanding based on the 3DGS framework.\n\n\nFigure 1:Illustration of two text query strategies. (a)\ndemonstrates the process of rendering a feature map and computing its similarity with text features to obtain a 2D mask, which is then used to generate a corresponding rendered image. (b) demonstrates the direct similarity computation of 3D Gaussian language features with text features, selecting Gaussian points with high similarity, and rendering to obtain a rendered image corresponding to the text.\n\n\nDespite several efforts[33,37,52,48,54,15,24]to incorporate learnable language attributes into 3DGS, aiming to enhance them with language-grounded capabilities, the primary objective of these approaches is to render language attributes onto images for 2D pixel-level understanding, lifting 2d feature into view-consistent understanding/segmentation. As shown in Fig.1(a), we demonstrate this using an example involving the rendering of language-embedded 3D Gaussians into 2D feature maps using LangSplat[33], followed by the utilization of open vocabulary for matching and identifying target areas. While these methods exhibit impressive performance in view-consistent lifting, we find a few drawbacks such as:1)inability to recognize occluded objects or parts, compromising the inherent 3D capabilities of 3DGS;2)incompatibility with robotics and embodied intelligence applications that necessitate 3D point-level understanding, localization, and interaction. Therefore, we aim to empower 3DGS with the capability of3D point-level open-vocabulary understanding.\n\n\nWe first investigate the 3D point-level understanding of existing works. As depicted in Fig1(b), we measure the similarity between textual features and 3D Gaussian language features, selecting points with high relevance, and subsequently render these points through the rasterization module of 3DGS to generate images.\nThe results highlight challenges in effectively matching target objects, indicating that although these approaches exhibit strong performance on 2D images, their 3D understanding capabilities are limited. As demonstrated in Fig.6, our visualization of their 3D point features reveals a lack of discriminability between different objects and low consistency within objects. We attribute these limitations to two main factors:1) Weak feature expressiveness: Due to the memory and speed constraints associated with 3D point-wise training and rendering in 3DGS, training high-dimensional language features for millions of Gaussian points in a scene becomes challenging. As a result, existing methods rely on dimension reduction techniques such as distillation[52,8]or quantization[37]to reduce dimensions. However, this inevitably compromises the expressiveness and distinguishability of the features.2) Inaccurate 2D-3D correspondence: The alpha-blending rendering technique accumulates the values of 3D points based on opacity weights to render 2D pixels, which prevents the establishment of a one-to-one correspondence between 2D and 3D. Consequently, a performance mismatch occurs between 2D and 3D interpretations.\n\n\nTo address these challenges, we propose OpenGaussian, an approach that learns distinctive and consistent features at the 3D point-level, both across objects and within objects. Our method associates high-dimensional lossless CLIP[34]features with 3D Gaussian points, enabling open-vocabulary 3D scene understanding. Specifically, the technical contributions of this paper are summarized as:1)Training of 3D point-level instance features that are both distinctive and 3D consistent using the proposed intra-mask smoothing loss and inter-mask contrastive loss, leveraging boolean masks from SAM[21]without cross-frame associations;2)Introducing a two-level coarse-to-fine codebook to discretize the instance features, resulting in discrete 3D instance clusters.3)Proposing an instance-level 2D-3D association method based on IoU and feature distance to associate CLIP features from multiple views for each 3D instance.\nThrough comprehensive experiments covering open-vocabulary object selection at the 3D point level, open-vocabulary 3D point cloud understanding, click-based 3D object selection, and module ablation, we demonstrate the simplicity and efficiency of our method. OpenGaussian eliminates the need for an additional network for feature dimensionality compression or quantization while inheriting the open-vocabulary capabilities of the original CLIP features.\n\n",
    "related_work": "2.1Neural RenderingNeural 3D scene representation, for example, the recently proposed\nNeRF[29]has demonstrated remarkable advancements in novel view synthesis quality using learning-based optimization techniques. While many methods focus on improving NeRF’s rendering quality[2,4,3,17], they often suffer from slow training and rendering speeds.\nAlternatively, methods based on explicit representation, such as voxels[39,14,35], hash grids[30]and point clouds[50,1,36], have emerged. These methods employ techniques to reduce the computational cost of large neural networks. The recent development of 3DGS[19]sets a new benchmark in terms of both rendering quality and rendering speed by employing fast differentiable rasterization of 3D Gaussians instead of volume rendering. As neural rendering proves to be an effective connection between 2D images and 3D scenes, our work builds upon this paradigm with a particular focus on 3D point-level open-vocabulary understanding.\n\n\n2.23D Open Vocabulary UnderstandingRecent advancements in open vocabulary scene understanding have seen the integration of 2D Vision-Language Models (VLMs) with 3D point cloud processing, resulting in significant progress in the field[16,45,53]. These approaches primarily concentrate on aligning features and projecting 3D data into 2D, thereby enhancing zero-shot learning capabilities. Furthermore, significant progress has been made in 3D object detection and segmentation[11,27,40], demonstrating the efficacy of merging point cloud data with visual features extracted from images for scene analysis.The significant advancements in 2D scene understanding, pioneered by SAM[21]and its variants, have motivated the exploration of integrating semantic features into NeRF. Methods have been developed to incorporate semantic features from models such as CLIP[34]and DINO[7]into NeRF, enabling more effective handling of 3D segmentation, understanding, and editing tasks. LERF[20]distills features from readily available VLMs like CLIP into a 3D scene represented by NeRF.[26]also introduces a 3D open-vocabulary segmentation pipeline using NeRF. Recent efforts have been made to combine 2D scene understanding techniques with 3D Gaussians to create a real-time and editable 3D scene representation, addressing the computational challenges of NeRF-based methods.LEGaussians[37]introduces uncertainty and semantic feature attributes to each Gaussian, to render a semantic map with corresponding uncertainties. This rendered map is compared with the quantized CLIP and DINO dense features extracted from the ground truth image. LangSplat[33]uses a scene-wise language autoencoder to learn language features on the scene-specific latent space, demonstrating to discern clear boundaries between objects in rendered feature images. Feature3DGS[52]proposes a parallelN𝑁Nitalic_N-dimensional Gaussian rasterizer to distill high-dimensional features for view-based tasks such as editing and segmentation. To achieve the 2D mask\nconsistency across views, Gaussian Grouping[48]performs simultaneous reconstruction and segmentation of open-world 3D objects, guided by 2D mask predictions obtained from SAM and 3D spatial consistency constraints. Similar to these works, we leverage the real-time rendering and explicit representation capabilities of 3DGS. However, while those methods primarily focus on pixel-level open-vocabulary understanding (i.e., lifting 2D feature into view-consistent segmentation), our approach diverges as we aim to enhance 3DGS with the ability for 3D point-level open-vocabulary understanding.\n\n",
    "origin_citations": [
      "Nerf: Representing scenes as neural radiance fields for view synthesis.",
      "Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields.",
      "Zip-nerf: Anti-aliased grid-based neural radiance fields.",
      "Mip-nerf 360: Unbounded anti-aliased neural radiance fields.",
      "Alignerf: High-fidelity neural radiance fields via alignment-aware training.",
      "Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction.",
      "Plenoxels: Radiance fields without neural networks.",
      "Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes.",
      "Instant neural graphics primitives with a multiresolution hash encoding.",
      "Differentiable surface splatting for point-based geometry processing.",
      "Neural point-based graphics.",
      "Adop: Approximate differentiable one-pixel point rendering.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "Clip2point: Transfer clip to point cloud classification with image-depth pre-training.",
      "Ulip-2: Towards scalable multimodal pre-training for 3d understanding.",
      "Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning.",
      "Pla: Language-driven open-vocabulary 3d scene understanding.",
      "Open-vocabulary point-cloud object detection without 3d annotation.",
      "Openmask3d: Open-vocabulary 3d instance segmentation.",
      "Segment anything.",
      "Learning transferable visual models from natural language supervision.",
      "Emerging properties in self-supervised vision transformers.",
      "Lerf: Language embedded radiance fields.",
      "Weakly supervised 3d open-vocabulary segmentation.",
      "Language embedded 3d gaussians for open-vocabulary scene understanding.",
      "Langsplat: 3d language gaussian splatting.",
      "Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields.",
      "Gaussian grouping: Segment and edit anything in 3d scenes."
    ],
    "citations": [
      253579,
      507332,
      18058,
      338391,
      245321,
      338715,
      151392,
      226298,
      58406,
      268778,
      439639,
      341006,
      180969,
      461615,
      114176,
      244395,
      242290,
      20840,
      48698,
      2512,
      1639,
      2514,
      496002,
      259505,
      807754,
      346753,
      273318,
      274520
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 86,
      "Final_Score": 83.2,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 86\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 89,
      "Synthesis_Analysis": 78,
      "Final_Score": 83.6,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 89\nSynthesis_Analysis: 78\nFinal_Score: 84"
    },
    "overall_score": 83.44,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "3NAEowLh7Q",
        "forum": "3NAEowLh7Q",
        "number": 1591,
        "title": "OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary Understanding",
        "normalized_title": "opengaussian towards point level 3d gaussian based open vocabulary understanding",
        "abstract": "This paper introduces OpenGaussian, a method based on 3D Gaussian Splatting (3DGS) that possesses the capability for 3D point-level open vocabulary understanding. Our primary motivation stems from observing that existing 3DGS-based open vocabulary methods mainly focus on 2D pixel-level parsing. These methods struggle with 3D point-level tasks due to weak feature expressiveness and inaccurate 2D-3D feature associations. To ensure robust feature presentation and 3D point-level understanding, we first employ SAM masks without cross-frame associations to train instance features with 3D consistency. These features exhibit both intra-object consistency and inter-object distinction. Then, we propose a two-stage codebook to discretize these features from coarse to fine levels. At the coarse level, we consider the positional information of 3D points to achieve location-based clustering, which is then refined at the fine level.\nFinally, we introduce an instance-level 3D-2D feature association method that links 3D points to 2D masks, which are further associated with 2D CLIP features. Extensive experiments, including open vocabulary-based 3D object selection, 3D point cloud understanding, click-based 3D object selection, and ablation studies, demonstrate the effectiveness of our proposed method. The source code is available at our project page https://3d-aigc.github.io/OpenGaussian.",
        "authors": [
          "Yanmin Wu",
          "Jiarui Meng",
          "Haijie LI",
          "Chenming Wu",
          "Yahao Shi",
          "Xinhua Cheng",
          "Chen Zhao",
          "Haocheng Feng",
          "Errui Ding",
          "Jingdong Wang",
          "Jian Zhang"
        ],
        "decision": null,
        "pdate": 1727287665475,
        "cdate": 1714707704171
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "3NAEowLh7Q",
          "forum": "3NAEowLh7Q",
          "number": 1591,
          "title": "OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary Understanding",
          "normalized_title": "opengaussian towards point level 3d gaussian based open vocabulary understanding",
          "abstract": "This paper introduces OpenGaussian, a method based on 3D Gaussian Splatting (3DGS) that possesses the capability for 3D point-level open vocabulary understanding. Our primary motivation stems from observing that existing 3DGS-based open vocabulary methods mainly focus on 2D pixel-level parsing. These methods struggle with 3D point-level tasks due to weak feature expressiveness and inaccurate 2D-3D feature associations. To ensure robust feature presentation and 3D point-level understanding, we first employ SAM masks without cross-frame associations to train instance features with 3D consistency. These features exhibit both intra-object consistency and inter-object distinction. Then, we propose a two-stage codebook to discretize these features from coarse to fine levels. At the coarse level, we consider the positional information of 3D points to achieve location-based clustering, which is then refined at the fine level.\nFinally, we introduce an instance-level 3D-2D feature association method that links 3D points to 2D masks, which are further associated with 2D CLIP features. Extensive experiments, including open vocabulary-based 3D object selection, 3D point cloud understanding, click-based 3D object selection, and ablation studies, demonstrate the effectiveness of our proposed method. The source code is available at our project page https://3d-aigc.github.io/OpenGaussian.",
          "authors": [
            "Yanmin Wu",
            "Jiarui Meng",
            "Haijie LI",
            "Chenming Wu",
            "Yahao Shi",
            "Xinhua Cheng",
            "Chen Zhao",
            "Haocheng Feng",
            "Errui Ding",
            "Jingdong Wang",
            "Jian Zhang"
          ],
          "decision": null,
          "pdate": 1727287665475,
          "cdate": 1714707704171
        }
      ]
    }
  },
  {
    "paper_id": 42338,
    "title": "From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs",
    "abstract": "Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduceMemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree’s depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model’s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.",
    "introduction": "Despite recent advances in large language models (LLMs) where the context window has expanded to millions of tokens(Ding et al.,2024; Bulatov et al.,2023; Beltagy et al.,2020; Chen et al.,2023; Tworkowski et al.,2024), these models continue to struggle with reasoning over long-term memory(Sun et al.,2021; Liu et al.,2024; Kuratov et al.,2024). This challenge arises because LLMs rely primarily on a key-value (KV) cache of past interactions, processed through a fixed number of transformer layers, which lack the capacity to effectively aggregate extensive historical data. Unlike LLMs, the human brain employs dynamic memory structures known as schemas, which facilitate the efficient organization, retrieval, and integration of information as new experiences occur(Anderson,2005; Ghosh & Gilboa,2014; Gilboa & Marlatte,2017). This dynamic restructuring of memory is a cornerstone of human cognition, allowing for the flexible application of accumulated knowledge across varied contexts.\n\n\nA prevalent method to address the limitations of long-term memory in LLMs involves the use of external memory.Weston et al. (2014)introduced the concept of utilizing external memory for the storage and retrieval of relevant information. More recent approaches in LLM research have explored various techniques to manage historical observations in databases, retrieving pertinent data for given queries through vector similarity searches in the embedding space(Park et al.,2023; Packer et al.,2023; Zhong et al.,2024). However, these methods primarily utilize a lookup table for memory representation, which fails to capture the inherent structure in data. Consequently, each past experience is stored as an isolated instance, lacking the interconnectedness and integrative capabilities of the human brain’s schemas. This limitation becomes increasingly problematic as the size of the memory grows or when relevant information is distributed across multiple instances.\n\n\nFigure 2:Illustration of MemTree. MemTree represents knowledge schema via a dynamic tree. Both parent and leaf nodes archive textual content, summarizing information relevant to their respective levels. Upon receiving new information, the system begins traversal from the root node. If the new information is semantically akin to an existing leaf node under the current node, it is routed to that node. Conversely, if it diverges from all existing leaf nodes under the current node, a new leaf node is created under the current node, concluding the traversal. During this process, all ancestor nodes will integrate the new information into the higher-level summaries they maintain.\n\n\nIn this work, we introduceMemTree, an algorithm designed to emulate the schema-like structures of the human brain by maintaining a dynamically structured memory representation during interactions. Within MemTree, each memory unit is represented as a node within a tree, containing node-level information and links to child nodes.\n\n\nUpon encountering new information, MemTree updates its memory structure starting from the root node. It evaluates at each node whether to instantiate a new child node or integrate the information into an existing child node. This decision process is governed by a traversal algorithm that efficiently adds new information with an insertion complexity of O(log N), where N denotes the number of conversational interactions. This structure facilitates the aggregation of knowledge at parent nodes, which evolve to capture high-level semantics as the tree expands. For knowledge retrieval, MemTree computes the cosine similarity between node embeddings and the query embedding. This method maintains the retrieval time complexity comparable to existing approaches, ensuring efficiency.\n\n\nWe evaluated MemTree across four benchmarks, covering both conversational and document question-answering tasks, and compared it against online and offline knowledge representation methods. MemTree enables seamless dynamic updates as new data becomes available, a capability characteristic of online methods. In contrast, offline methods require complete dataset access and are costly to update, as they involve periodic rebuilding to incorporate new information.•In extended conversations, MemTree consistently outperforms other online methods, including MemoryStream(Park et al.,2023)and MemGPT(Packer et al.,2023), maintaining superior accuracy as discussions progress.•For document question-answering, MemTree excels in two key areas. On single-document tasks, it outperforms other online methods and reduces the performance gap with offline models, particularly surpassing RAPTOR(Sarthi et al.,2024)on challenging questions that require deeper reasoning. In multi-document tasks, MemTree not only surpasses online methods but also approaches the performance of offline models, particularly outperforming GraphRAG(Edge et al.,2024). Moreover, on complex temporal queries requiring the analysis of event sequences across multiple documents, MemTree exceeds all offline methods.These results demonstrate that MemTree is a robust and scalable solution, delivering high accuracy across diverse and challenging tasks while retaining the efficiency of online systems.\n\n",
    "related_work": "Memory-Augmented LLMsRecent advancements in memory-augmented LLMs have introduced various strategies for enhancing memory capabilities.Park et al. (2023)developed LLM-based agents that log experiences as timestamped descriptions, retrieving memories based on recency, importance, and relevance. Similarly,Cheng et al. (2023)developed Selfmem with a dedicated memory selector. MemGPT(Packer et al.,2023)proposed automatic memory management through LLM function-calling for conversational agents and document analysis, providing a pre-prompt with detailed instructions on memory hierarchy and utilities, along with a schema for accessing and modifying the memory database.Zhong et al. (2024)introduced MemoryBank, a long-term memory framework that stores timestamped dialogues and uses exponential decay to forget outdated information. Additionally,Mitchell et al. (2022)proposed storing model edits in an explicit memory and learning to reason over them to adjust the base model’s predictions. These methods represent common solutions for adding memory to LLMs, focusing on tabular memory storage and vector similarity retrieval(Zhang et al.,2024). However, as memory scales or information becomes dispersed across multiple entries, their unstructured representations reveal significant limitations.Another line of work explores triplet-based memory. For example,Modarressi et al. (2023)proposed encoding relationships in triplets, andAnokhin et al. (2024)extended this approach to graph-based triplet memory for text-based games. While effective at encoding individual relations or scene graphs at the object level, these methods struggle with scalability and generalization to more complex data that do not fit neatly into a strict triplet format.\n\n\nStructured Retrieval-Augmented Generation ApproachesTo address the limitations of unstructured memory representations, recent advances have integrated structured knowledge into RAG models, enhancing navigation and summarization in complex QA tasks.Trajanoska et al. (2023)leveraged LLMs to extract entities and relationships from unstructured text to construct knowledge graphs. Similarly,Yao et al. (2023)proposed techniques to fill in missing links and nodes in existing knowledge graphs by utilizing LLMs to infer unseen relationships.Ban et al. (2023)identified causal relationships within textual data and represented them in graph form to enhance understanding of causal structures. In the context of retrieval-augmented generation,Gao et al. (2023)provided a comprehensive review of existing RAG methods, andBaek et al. (2023)utilized knowledge graphs as indexes within RAG frameworks for efficient retrieval of structured information.More relevant to our work, RAPTOR(Sarthi et al.,2024)organizes text into a recursive tree, clustering and summarizing chunks at multiple layers to enable efficient retrieval of both high-level themes and detailed information. GraphRAG(Edge et al.,2024)constructs a knowledge graph from LLM-extracted entities and relations, partitioning it into modular communities that are independently summarized and combined via a map-reduce framework. While these methods effectively structure large textual data to improve retrieval and generation capabilities, they are limited to static corpora, requiring full reconstruction to integrate new information, and do not support online memory updates.In this work, we propose a novel structured memory representation that overcomes these limitations by enabling dynamic updates and efficient retrieval in large-scale memory systems without necessitating full reconstruction.\n\n",
    "origin_citations": [
      "Generative agents: Interactive simulacra of human behavior.",
      "Lift yourself up: Retrieval-augmented text generation with self memory.",
      "Memgpt: Towards llms as operating systems.",
      "Memorybank: Enhancing large language models with long-term memory.",
      "Memory-based model editing at scale.",
      "A survey on the memory mechanism of large language model based agents.",
      "Ret-llm: Towards a general read-write memory for large language models.",
      "Arigraph: Learning knowledge graph world models with episodic memory for llm agents.",
      "Enhancing knowledge graph construction using large language models.",
      "Exploring large language models for knowledge graph completion.",
      "From query tools to causal architects: Harnessing large language models for advanced causal discovery from data.",
      "Retrieval-augmented generation for large language models: A survey.",
      "Knowledge-augmented language model prompting for zero-shot knowledge graph question answering.",
      "Raptor: Recursive abstractive processing for tree-organized retrieval.",
      "From local to global: A graph rag approach to query-focused summarization."
    ],
    "citations": [
      2945,
      373,
      471,
      2994,
      2982,
      135148,
      7652,
      740153,
      115891,
      404416,
      47250,
      349125,
      306420,
      8351,
      133921
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 78,
      "Final_Score": 81.4,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 85\nSynthesis_Analysis: 78\nFinal_Score: 81"
    },
    "overall_score": 83.44,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "moXtEmCleY",
        "forum": "moXtEmCleY",
        "number": 3306,
        "title": "From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs",
        "normalized_title": "from isolated conversations to hierarchical schemas dynamic tree memory representation for llms",
        "abstract": "Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduce MemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree's depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model’s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.",
        "authors": [
          "Alireza Rezazadeh",
          "Zichao Li",
          "Wei Wei",
          "Yujia Bao"
        ],
        "decision": null,
        "pdate": 1737562311536,
        "cdate": 1727136844136
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "moXtEmCleY",
          "forum": "moXtEmCleY",
          "number": 3306,
          "title": "From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs",
          "normalized_title": "from isolated conversations to hierarchical schemas dynamic tree memory representation for llms",
          "abstract": "Recent advancements in large language models have significantly improved their context windows, yet challenges in effective long-term memory management remain. We introduce MemTree, an algorithm that leverages a dynamic, tree-structured memory representation to optimize the organization, retrieval, and integration of information, akin to human cognitive schemas. MemTree organizes memory hierarchically, with each node encapsulating aggregated textual content, corresponding semantic embeddings, and varying abstraction levels across the tree's depths. Our algorithm dynamically adapts this memory structure by computing and comparing semantic embeddings of new and existing information to enrich the model’s context-awareness. This approach allows MemTree to handle complex reasoning and extended interactions more effectively than traditional memory augmentation methods, which often rely on flat lookup tables. Evaluations on benchmarks for multi-turn dialogue understanding and document question answering show that MemTree significantly enhances performance in scenarios that demand structured memory management.",
          "authors": [
            "Alireza Rezazadeh",
            "Zichao Li",
            "Wei Wei",
            "Yujia Bao"
          ],
          "decision": null,
          "pdate": 1737562311536,
          "cdate": 1727136844136
        }
      ]
    }
  },
  {
    "paper_id": 48834,
    "title": "VideoGrain: Modulating Space-Time Attention for Multi-grained Video Editing",
    "abstract": "Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge.\nThe major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model.\nTo address these difficulties,\nwe present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content.\nWe enhance text-to-region control by amplifying each local prompt’s attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention.\nAdditionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention.\nExtensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios.\nOur code, data, and demos are available on theproject page.",
    "introduction": "Figure 2:Definition of multi-grained video editing and comparison on instance editing\n\n\nRecent advances in Text-to-Image (T2I) and Text-to-Video (T2V) diffusion models(Rombach et al.,2022; Wang et al.,2023a; Brooks et al.,2024)have enabled video manipulation through natural language prompts.\nIn practical applications, enabling users to edit regions at various levels of granularity based on textual prompts offers greater flexibility.\nTo investigate this, we introduce a new task called multi-grained video editing, which encompasses class-level, instance-level, and part-level editing, as shown in Fig.2left.\nClass-level editing refers to modifying objects within the same class.\nInstance-level editing means editing different instances into distinct objects.\nPart-level going further, requires adding new objects or modifying existing attributes at part-level.\n\n\nWhile existing methods employ various visual consistency techniques, such as optical flow(Cong et al.,2023; Yang et al.,2023), control signals(Zhang et al.,2023b), or feature correspondence(Geyer et al.,2023). These methods remain instance-agnostic, often mixing features of different instances during editing (see Fig.2right).\nGround-A-Video(Jeong & Ye,2023), which inherits text-to-bounding box generation priors(Li et al.,2023), should be instance-level editing but still suffer from artifacts.\nSimilarly, recent T2V-based methods like DMT(Yatim et al.,2024)and Pika(pik,), although equipped with video generation priors, struggle with multi-grained edits.\nWe find that the core issue is that diffusion models tend to treat different instances as the same class segments, leading to strong feature coupling across instances, as illustrated in Figure3.\n\n\nTo address this problem, our primary insight is to 1) enable text-to-region control and 2) keep feature separation between regions.\nIn the typical diffusion models, the cross-attention layer serves as a key component to update textual features control over each spatial region, while the self-attention layer generates globally coherent structures by connecting each frame token across time.\nTherefore, we propose Spatial-Temporal Layout-Guided Attention (ST-Layout Attn), which modulates both space-time cross- and self-attention in a unified manner to achieve the above goals.\n\n\nIn thecross-attention layer, the uniform application of global text prompts across all frame tokens leads to severe semantic misalignment, which reduces the precision of multi-grained text-to-region control.\nTo address this, we modulate cross-attention to amplify each local prompt’s focus on its corresponding spatial-disentangled region while suppressing attention to irrelevant areas.\nIn theself-attention layer, pixels from one region may attend to outside or similar regions within the same class, leading to feature coupling and texture mixing, which is an inherent limitation of diffusion models that complicates multi-grained video editing.\nTo mitigate this, we modulate self-attention to enhance feature separation by increasing intra-region focus and reducing inter-region interactions, ensuring each query attends only to its target region.\n\n\nOur key contributions can be summarized as follows:•To the best of our knowledge, this is the first attempt at multi-grained video editing. Our method enables both class-level, instance-level and part-level editing.•We propose a novel framework, dubbedVideoGrain, which modulates spatial-temporal cross- and self-attention for text-to-region control and feature separation between regions.•Without tuning any parameters, we achieve state-of-the-art results on existing benchmarks and real-world videos both qualitatively and quantitatively.\n\n",
    "related_work": "2.1Text-to-Image Editing/GenerationIn the realm of single attribute text-to-image editing, various approaches have been explored, from manipulating attention maps in Pix2Pix-Zero(Parmar et al.,2023)and Prompt2Prompt(Hertz et al.,2022)to employing masks in DiffEdit(Couairon et al.,2023)and Latent Blend(Avrahami et al.,2022;2023)for foreground modifications while preserving the background.For multi-grained editing, efforts like Attention and Excite(Chefer et al.,2023)and DPL(Wang et al.,2023b)focus on maximizing attention scores for each subject token and reducing attention leakage. In image generation,(Kim et al.,2023)modulates attention based on layout masks and dense captions, while(Phung et al.,2023)proposed an attention refocus loss for regularization. However, using single-frame layout masks and dense captioning alone is insufficient for video editing, as it fails to maintain the original video’s integrity and temporal consistency.\n\n\n2.2Text-to-Video EditingVideo Editing based on Image Diffusion Models.Tune-A-Video (TAV)(Wu et al.,2022)is the first work to extend latent diffusion models to the spatial-temporal domain and encode the source motion implicitly by one-shot tuning but still fails to preserve local details. Fatezero(Qi et al.,2023)and Pix2Video(Ceylan et al.,2023)fuse self- or cross-attention maps in the inversion process for temporal consistency.\nHowever,(Qi et al.,2023)requires extensive RAM usage and suffers from layout preservation even when equipping TAV for local object editing.(Chai et al.,2023)and(Ouyang et al.,2023), following the Neural Atlas(Kasten et al.,2021)or dynamic Nerf’s deformation field(Pumarola et al.,2021), struggle with non-grid human motion.\nSubsequent methods like Rerender-A-Video(Yang et al.,2023), FLATTEN(Cong et al.,2023)ControlVideo(Zhang et al.,2023b)achieve strict temporal consistency via optical-flow, depth/edge maps, but failed in multi-grained editing while preserving original layouts. Tokenflow(Geyer et al.,2023)enforces a linear mix of nearest key-frame features to ensure consistency but results in detail loss. Ground-A-Video(Jeong & Ye,2023)leverages groundings for multi-grained editing, but it suffers from feature mixing when bounding boxes overlap.Video Editing based on Video Diffusion Models.Previous video editing work primarily utilized text-to-image SD model(Rombach et al.,2022).\nRecent advancements in video foundation models(Yu et al.,2023; Guo et al.,2023; Wang et al.,2023a; Yang et al.,2024e)have led efforts like VideoSwap(Gu et al.,2023)to employ temporal priors for customized motion transfer or motion editing(Mou et al.,2025).\nYet, current video foundation models are limited to fixed views and struggle with non-grid human motions. Additionally, these editing methods require tuning parameters, which poses a challenge for real-time video editing applications. In contrast, our VideoGrain method requires no parameter tuning, enabling zero-shot, multi-grained video editing.\n\n",
    "origin_citations": [
      "Zero-shot image-to-image translation.",
      "Prompt-to-prompt image editing with cross attention control.",
      "Diffedit: Diffusion-based semantic image editing with mask guidance.",
      "Blended diffusion for text-driven editing of natural images.",
      "Blended latent diffusion.",
      "Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models.",
      "Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing.",
      "Dense text-to-image generation with attention modulation.",
      "Grounded text-to-image synthesis with attention refocusing.",
      "Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation.",
      "Fatezero: Fusing attentions for zero-shot text-based video editing.",
      "Pix2video: Video editing using image diffusion.",
      "Stablevideo: Text-driven consistency-aware diffusion video editing.",
      "Codef: Content deformation fields for temporally consistent video processing.",
      "Layered neural atlases for consistent video editing.",
      "D-nerf: Neural radiance fields for dynamic scenes.",
      "Rerender a video: Zero-shot text-guided video-to-video translation.",
      "Flatten: optical flow-guided attention for consistent text-to-video editing.",
      "Controlvideo: Training-free controllable text-to-video generation.",
      "Tokenflow: Consistent diffusion features for consistent video editing.",
      "Ground-a-video: Zero-shot grounded video editing using text-to-image diffusion models.",
      "High-resolution image synthesis with latent diffusion models.",
      "Magvit: Masked generative video transformer.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Modelscope text-to-video technical report.",
      "Cogvideox: Text-to-video diffusion models with an expert transformer.",
      "Videoswap: Customized video subject swapping with interactive semantic point correspondence.",
      "Revideo: Remake a video with motion and content control."
    ],
    "citations": [
      490740,
      236009,
      335814,
      337260,
      413004,
      478139,
      404850,
      305758,
      139303,
      496009,
      494475,
      406516,
      907813,
      365602,
      431037,
      304517,
      321553,
      260101,
      469944,
      324072,
      2240,
      221810,
      185878,
      179813,
      977802,
      273814,
      1042029
    ],
    "extract_rate": 0.9642857142857143,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86"
    },
    "related_work_scores": {
      "Content_Coherence": 74,
      "Citation_Quality": 87,
      "Synthesis_Analysis": 81,
      "Final_Score": 81.3,
      "raw_text": "Content_Coherence: 74\nCitation_Quality: 87\nSynthesis_Analysis: 81\nFinal_Score: 81"
    },
    "overall_score": 83.38,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "SSslAtcPB6",
        "forum": "SSslAtcPB6",
        "number": 308,
        "title": "VideoGrain: Modulating Space-Time Attention for Multi-Grained Video Editing",
        "normalized_title": "videograin modulating space time attention for multi grained video editing",
        "abstract": "Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available on the [project page](https://knightyxp.github.io/VideoGrain_project_page/).",
        "authors": [
          "Xiangpeng Yang",
          "Linchao Zhu",
          "Hehe Fan",
          "Yi Yang"
        ],
        "decision": null,
        "pdate": 1737562185646,
        "cdate": 1726214034262
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "SSslAtcPB6",
          "forum": "SSslAtcPB6",
          "number": 308,
          "title": "VideoGrain: Modulating Space-Time Attention for Multi-Grained Video Editing",
          "normalized_title": "videograin modulating space time attention for multi grained video editing",
          "abstract": "Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available on the [project page](https://knightyxp.github.io/VideoGrain_project_page/).",
          "authors": [
            "Xiangpeng Yang",
            "Linchao Zhu",
            "Hehe Fan",
            "Yi Yang"
          ],
          "decision": null,
          "pdate": 1737562185646,
          "cdate": 1726214034262
        }
      ]
    }
  },
  {
    "paper_id": 24757,
    "title": "Learning Latent Space Hierarchical EBM Diffusion Models",
    "abstract": "This work studies the learning problem of the energy-based prior model and the multi-layer generator model. The multi-layer generator model, which contains multiple layers of latent variables organized in a top-down hierarchical structure, typically assumes the Gaussian prior model. Such a prior model can be limited in modelling expressivity, which results in a gap between the generator posterior and the prior model, known as the prior hole problem. Recent works have explored learning the energy-based (EBM) prior model as a second-stage, complementary model to bridge the gap. However, the EBM defined on a multi-layer latent space can be highly multi-modal, which makes sampling from such marginal EBM prior challenging in practice, resulting in ineffectively learned EBM. To tackle the challenge, we propose to leverage the diffusion probabilistic scheme to mitigate the burden of EBM sampling and thus facilitate EBM learning. Our extensive experiments demonstrate a superior performance of our diffusion-learned EBM prior on various challenging tasks.",
    "introduction": "The hierarchical generative model with multiple layers of latent variables (a.k.a.,multi-layer generator model) has made significant progress in learning complex data distribution(Sø nderby et al.,2016; Vahdat & Kautz,2020)and has garnered particular interest for its top-down hierarchical structure, where multiple layers of latent variables that are organized from the top to the bottom layers tend to capture levels of (hierarchical) data representations, with high-level semantic representations captured by the latent variables at the top layers and low-level detail representations by those at the bottom layers(Maaløe et al.,2019; Child,2020). Learning such hierarchical representation can be essential and crucial to various downstream applications(Havtorn et al.,2021; Nijkamp et al.,2020b). However, such multi-layer generator models typically assume the Gaussian prior model, which can be limited in statistical expressivity by primarily focusing on the inter-layer relation between layers of latent variables while largely ignoring the intra-layer relation between latent units within each layer(Cui et al.,2023a,b). This may result in theprior hole problem(Rosca et al.,2018; Hoffman & Johnson,2016; Takahashi et al.,2019; Bauer & Mnih,2019)where the non-expressive Gaussian prior fails to match the aggregated generator posterior.\n\n\nRecent studies(Aneja et al.,2021; Cui et al.,2023a)have investigated the utilization of the energy-based (EBM) prior model as a complementary model to address this limitation. The EBM prior is typically trained with a fixed generator model (referred to as theTwo-Stagelearning scheme) to tilt the non-expressive Gaussian prior to match the generator posterior. However, learning asingle(marginal) EBM is challenging because the generator posterior is often multi-modal, and more importantly, the Markov Chain Monte Carlo (MCMC) sampling required to maximize the marginal EBM likelihood can be difficult, as multiple layers of latent variables are interwoven and require exploration at different latent scales. In addition, MCMC sampling, such as Langevin dynamics, usually starts from a noise-initialized point, which is hard to explore the energy landscape and mix between different local modes. Therefore, for multi-layer latent variables, EBM prior sampling may serve as the bottleneck for effective EBM learning, which still poses a challenge.\n\n\nInspired by recent diffusion probabilistic frameworks(Ho et al.,2020; Gao et al.,2020; Zhu et al.,2023; Du et al.,2023), we propose learning the EBM prior of multi-layer latent variables in a diffusion learning scheme. We construct a series of conditional EBMs prior to gradually matching the highly multi-modal generator posterior, with each conditional EBM prior only matching perturbed samples at each step. Compared to marginal EBM prior, such a conditional EBM prior can be less multi-modal, leading to more tractable conditional likelihood learning. For EBM sampling, the proposed conditional EBM prior can render a smoother energy landscape, which mitigates the burden of MCMC sampling and thus further facilitates effective EBM learning. However, for multi-layer latent variables, MCMC sampling needs to account for their different latent scales at different layers (i.e., the scales of latent variables at the top and bottom layers can be very different); moreover, directly perturbating the latent samples may destroy their inter-layer relation (i.e., conditional dependency formulated in the hierarchical structure). Therefore, we further employ a uni-scale𝐮~~𝐮{\\mathbf{\\tilde{u}}}over~ start_ARG bold_u end_ARG-space (see definition in Sec.3.1) converted from the multi-scale latent space, which allows us to preserve the hierarchical dependency along the forward process while at the same time, further reducing the burden of MCMC sampling by traversing a uni-scale latent space. Our experiments demonstrate the effectiveness of the proposed method in various challenging tasks and show that our model is capable of generating high-quality samples and capturing hierarchical representations at different layers.\n\n\nContribution:1) We develop a learning framework that incorporates the diffusion probabilistic scheme for learning the joint EBM prior for the multi-layer generator model; 2) To preserve hierarchical structures and enable more effective EBM sampling, we adopt a uni-scale space to further mitigate the burden of MCMC sampling; 3) We conduct various experiments to examine our model in generating high-quality samples and learning effective hierarchical representations.\n\n",
    "related_work": "Energy-based model.The EBM is expressive in representing the data uncertainty. Most existing works learn EBM on data space by maximizing EBM likelihood, which involves challenging MCMC sampling for the EBM. To tackle the challenge,(Du & Mordatch,2019; Du et al.,2020)propose the use of a replay buffer, while(Cui & Han,2023; Han et al.,2020b; Xie et al.,2021,2018)consider learning a complementary generator model to jump-start MCMC sampling.(Xiao & Han,2022; Pang et al.,2020a,b)focus on learning EBM prior on low-dimensional latent space to mitigate the burden of EBM sampling, but these works only deal with a single-layer latent space. We consider learning EBM prior for multi-layer latent variables, which enables hierarchical representation learning and better sample generation.\n\n\nHierarchical generative model.For hierarchical generative models(Vahdat & Kautz,2020; Maaløe et al.,2019; Sø nderby et al.,2016), they typically assume a Gaussian prior model, which can be less informative, resulting in the prior hole problem and poor generation quality. Our work aims to learn effective EBM prior for hierarchical generative models, which is related to joint EBM prior (see Sec.2.2). These works intend to learn a single (marginal) EBM on𝐳~~𝐳{\\mathbf{\\tilde{z}}}over~ start_ARG bold_z end_ARG-space, while we study learning a sequence of conditional EBMs on𝐮~~𝐮{\\mathbf{\\tilde{u}}}over~ start_ARG bold_u end_ARG-space. NCP-VAE(Aneja et al.,2021)discuss aautoregressive-style model, which can beintractablefor MLE learning, as the normalization constant includes the top layer𝐳i+1subscript𝐳𝑖1{\\mathbf{z}}_{i+1}bold_z start_POSTSUBSCRIPT italic_i + 1 end_POSTSUBSCRIPTand needs an additional inner loop for sampling. Our model leverages diffusion probabilistic models such that we can directly draw𝐮~t+1subscript~𝐮𝑡1{\\mathbf{\\tilde{u}}}_{t+1}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPTthrough the forward trajectory. Compared to(Cui et al.,2023a)(Eqn.4), our EBM prior only models𝐮~tsubscript~𝐮𝑡{\\mathbf{\\tilde{u}}}_{t}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPTconditioned on the perturbed sample𝐮~t+1subscript~𝐮𝑡1{\\mathbf{\\tilde{u}}}_{t+1}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPTgenerated by𝐮~t+1=αt+1⁢𝐮~t+ϵt+1⁢σt+1subscript~𝐮𝑡1subscript𝛼𝑡1subscript~𝐮𝑡subscriptitalic-ϵ𝑡1subscript𝜎𝑡1{\\mathbf{\\tilde{u}}}_{t+1}=\\alpha_{t+1}{\\mathbf{\\tilde{u}}}_{t}+\\epsilon_{t+1}%\n\\sigma_{t+1}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPT = italic_α start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPT over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT + italic_ϵ start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPT italic_σ start_POSTSUBSCRIPT italic_t + 1 end_POSTSUBSCRIPT, which can be much easier than directly recovering from𝐮~Tsubscript~𝐮𝑇{\\mathbf{\\tilde{u}}}_{T}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPTto𝐮~0subscript~𝐮0{\\mathbf{\\tilde{u}}}_{0}over~ start_ARG bold_u end_ARG start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT(if consider equivalent form Eqn.8).\n\n\nEBM diffusion model.Other diffusion EBMs(Gao et al.,2020; Yu et al.,2022,2023)motivate EBM learning with a diffusion scheme. They build energy-basedrecoverymodel on data space and single-layer latent space, respectively. In this work, we focus on multi-layer generator model with a top-down hierarchical structure, which is shown to be capable of learning meaningful hierarchical representations(Child,2020; Zhao et al.,2017; Havtorn et al.,2021). It is typically challenging for the diffusion scheme to maintain such hierarchical structures as the goal of the forward process is to destroy the data pattern, which includes the inter-layer relation (conditional dependency) among multi-layer latent variables. To preserve the hierarchical structure of latent variables, we conduct the forward and reverse processes on𝐮~~𝐮{\\mathbf{\\tilde{u}}}over~ start_ARG bold_u end_ARG-space, which also leads to more effective learning and sampling than multi-scale𝐱𝐱{\\mathbf{x}}bold_x-space(Gao et al.,2020)and𝐳𝐳{\\mathbf{z}}bold_z-space(Yu et al.,2022).\n\n\nFigure 1:Image synthesis on CelebA-HQ-256 (left), LSUN-Church-64 (center) and CIFAR-10 (right).\n\n\nTable 1:IS(↑)↑(\\uparrow)( ↑ )and FID(↓)↓(\\downarrow)( ↓ )on CIFAR-10. Model∗indicates our backbone model.MethodISFIDOurs(T=3𝑇3T=3italic_T = 3)9.038.93Joint-EBM(Cui et al.,2023a)8.9911.34DRL-EBM (T=6𝑇6T=6italic_T = 6)(Gao et al.,2020)8.409.58NCP-VAE(Aneja et al.,2021)-24.08Hierarchical Generative Models w Gaussian PriorNVAE∗(Vahdat & Kautz,2020)5.3037.73NVAE∗-Recon-0.68HVAE(Sø nderby et al.,2016)-81.44BIVA(Maaløe et al.,2019)-66.37Energy-based ModelsArchitectural-EBM(Cui et al.,2023b)-63.42Dual-MCMC(Cui & Han,2023)8.559.26Adaptive-CE(Xiao & Han,2022)-65.01VAEBM(Xiao et al.,2020)8.4312.19Hat EBM(Hill et al.,2022a)-19.15ImprovedCD(Du et al.,2020)7.8525.1Divergence Triangle(Han et al.,2020a)-30.10Adv-EBM(Yin et al.,2020)9.1013.21GANs+Score+Diffusion ModelsStyleGANv2 w/o ADA(Karras et al.,2020)8.999.9Diffusion-Amortized(Yu et al.,2023)-57.72NCSN(Song & Ermon,2019)8.8725.32LSGM(Vahdat et al.,2021)-2.10DDPM (T=1000𝑇1000T=1000italic_T = 1000)(Ho et al.,2020)9.463.17Table 2:FID on CelebA-HQ-256 and LSUN-Church-64.MethodCelebA-HQ-256LSUN-Church-64Ours(T=3𝑇3T=3italic_T = 3)8.787.34Joint-EBM(Cui et al.,2023a)9.898.38DRL-EBM (T=6𝑇6T=6italic_T = 6)(Gao et al.,2020)-7.04NCP-VAE(Aneja et al.,2021)24.79-NVAE∗(Vahdat & Kautz,2020)30.2538.13NVAE∗-Recon1.642.45Adv-EBM(Yin et al.,2020)17.3110.84GLOW(Kingma & Dhariwal,2018)68.9359.35PGGAN(Karras et al.,2017)8.036.42\n\n",
    "origin_citations": [
      "Implicit generation and generalization in energy-based models.",
      "Improved contrastive divergence training of energy based models.",
      "Learning energy-based model via dual-mcmc teaching.",
      "Joint training of variational auto-encoder and latent energy-based model.",
      "Learning energy-based model with variational auto-encoder as amortized sampler.",
      "Cooperative training of descriptor and generator networks.",
      "Adaptive multi-stage density ratio estimation for learning latent space energy-based model.",
      "Learning latent space energy-based prior model.",
      "Semi-supervised learning by latent space energy-based model of symbol-vector coupling.",
      "Nvae: A deep hierarchical variational autoencoder.",
      "Biva: A very deep hierarchy of latent variables for generative modeling.",
      "Ladder variational autoencoders.",
      "A contrastive learning approach for training variational autoencoder priors.",
      "Learning joint latent space ebm prior model for multi-layer generator.",
      "Learning energy-based models by diffusion recovery likelihood.",
      "Latent diffusion energy-based model for interpretable text modeling.",
      "Learning energy-based prior model with diffusion-amortized mcmc.",
      "Very deep vaes generalize autoregressive models and can outperform them on images.",
      "Learning hierarchical features from generative models.",
      "Hierarchical vaes know what they don’t know.",
      "Learning hierarchical features with joint latent space energy-based prior.",
      "Vaebm: A symbiosis between variational autoencoders and energy-based models.",
      "Learning probabilistic models from generator latent spaces with hat ebm.",
      "Analyzing and improving generative adversarial training for generative modeling and out-of-distribution detection.",
      "Training generative adversarial networks with limited data.",
      "Generative modeling by estimating gradients of the data distribution.",
      "Score-based generative modeling in latent space.",
      "Denoising diffusion probabilistic models.",
      "Glow: Generative flow with invertible 1x1 convolutions.",
      "Progressive growing of gans for improved quality, stability, and variation."
    ],
    "citations": [
      401602,
      111024,
      273658,
      350470,
      104690,
      522731,
      178816,
      1090572,
      437456,
      201284,
      515779,
      482054,
      5896,
      305360,
      895354,
      411609,
      322988,
      432405,
      144273,
      474787,
      319810,
      357489,
      333296,
      108,
      5921,
      194114,
      4827,
      3284,
      62
    ],
    "extract_rate": 0.9666666666666667,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 87,
      "Final_Score": 83.5,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 87\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 75,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 85,
      "Final_Score": 83.2,
      "raw_text": "Content_Coherence: 75\nCitation_Quality: 88\nSynthesis_Analysis: 85\nFinal_Score: 83"
    },
    "overall_score": 83.32,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "o9uOuIwhZK",
        "forum": "o9uOuIwhZK",
        "number": 3460,
        "title": "Learning Latent Space Hierarchical EBM Diffusion Models",
        "normalized_title": "learning latent space hierarchical ebm diffusion models",
        "abstract": "This work studies the learning problem of the energy-based prior model and the multi-layer generator model. The multi-layer generator model, which contains multiple layers of latent variables organized in a top-down hierarchical structure, typically assumes the Gaussian prior model. Such a prior model can be limited in modelling expressivity, which results in a gap between the generator posterior and the prior model, known as the prior hole problem. Recent works have explored learning the energy-based (EBM) prior model as a second-stage, complementary model to bridge the gap. However, the EBM defined on a multi-layer latent space can be highly multi-modal, which makes sampling from such marginal EBM prior challenging in practice, resulting in ineffectively learned EBM. To tackle the challenge, we propose to leverage the diffusion probabilistic scheme to mitigate the burden of EBM sampling and thus facilitate EBM learning. Our extensive experiments demonstrate a superior performance of our diffusion-learned EBM prior on various challenging tasks.",
        "authors": [
          "Jiali Cui",
          "Tian Han"
        ],
        "decision": null,
        "pdate": 1714610338003,
        "cdate": 1706652435939
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "o9uOuIwhZK",
          "forum": "o9uOuIwhZK",
          "number": 3460,
          "title": "Learning Latent Space Hierarchical EBM Diffusion Models",
          "normalized_title": "learning latent space hierarchical ebm diffusion models",
          "abstract": "This work studies the learning problem of the energy-based prior model and the multi-layer generator model. The multi-layer generator model, which contains multiple layers of latent variables organized in a top-down hierarchical structure, typically assumes the Gaussian prior model. Such a prior model can be limited in modelling expressivity, which results in a gap between the generator posterior and the prior model, known as the prior hole problem. Recent works have explored learning the energy-based (EBM) prior model as a second-stage, complementary model to bridge the gap. However, the EBM defined on a multi-layer latent space can be highly multi-modal, which makes sampling from such marginal EBM prior challenging in practice, resulting in ineffectively learned EBM. To tackle the challenge, we propose to leverage the diffusion probabilistic scheme to mitigate the burden of EBM sampling and thus facilitate EBM learning. Our extensive experiments demonstrate a superior performance of our diffusion-learned EBM prior on various challenging tasks.",
          "authors": [
            "Jiali Cui",
            "Tian Han"
          ],
          "decision": null,
          "pdate": 1714610338003,
          "cdate": 1706652435939
        }
      ]
    }
  },
  {
    "paper_id": 44821,
    "title": "Neuro-Symbolic Data Generation for Math Reasoning",
    "abstract": "critical question about Large Language Models (LLMs) is whether their apparent deficiency in mathematical reasoning is inherent, or merely a result of insufficient exposure to high-quality mathematical data. To explore this, we developed an automated method for generating high-quality, supervised mathematical datasets. The method carefully mutates existing math problems, ensuring both diversity and validity of the newly generated problems. This is achieved by a neuro-symbolic data generation framework combining the intuitive informalization strengths of LLMs, and the precise symbolic reasoning of math solvers along with projected Markov chain Monte Carlo sampling in the highly-irregular symbolic space.\nEmpirical experiments demonstrate the high quality of data generated by the proposed method, and that the LLMs, specifically LLaMA-2 and Mistral, when realigned with the generated data, surpass their state-of-the-art counterparts.",
    "introduction": "Despite recent progress[1,2,3,4,5,6], both proprietary and open-source LLMs are still far from satisfactory in mathematical reasoning[7,8,9].\nIt is an open question whether LLM’s subpar reasoning capability is inherent or due to the the extreme scarcity of high-quality mathematical datasets[10,11,12,13].\nAs an initial step towards answer this question, a data generation framework that could create high-quality math datasets is required.\nTo this end,\ncurrent two lines of research struggle in thediversity-validitydilemma:\n(1) to produce diverse math data, the prompt-based method effectively rephrases math problems using LLMs, but may induce errors thus ruining thevalidity, especially considering the rigor of maths;\n(2) to ensure the validity, template-based methods are often used by rewriting math problems with certain rules, sacrificing thediversityand thus confining data scale.\n\n\nTo address this dilemma, we propose a novelneuro-symbolicframework that automatically generates high-quality, supervised mathematical data.\nThe merit of this paradigm lies in leveraging both neural and symbolic strengths:\n(1) the math problem is generated in the symbolic space, achieving diversity through systematic sampling, while maintaining validity through symbolic solvers;\n(2) the translation from the symbolic space back to the natural language space can be effectively supported by LLMs, ensuring the consistency between newly generated formal problems and their corresponding natural language versions.\n\n\nFigure 1:The overview of our neuro-symbolic data generation framework.\nThe framework comprises three steps: (1) Formalize the seed problem into its symbolic version. (2) Mutate the symbolic problem to create new variants. (3) Translate the variants in symbolic form back to the natural language version. Additionally, we prompt GPT-4 to generate reasoning paths, which are verified by symbolic solvers, as part of the supervision.\n\n\nOur framework, as illustrated inFigure 1, initiates with the formalization of the original problem expressed via the math symbolic tools.\nNext, itmutatesthe formal problem into an evolved version, and then derives a new natural language problem byinformalization.\nSpecifically, we design a mutation mechanism, including various simplification and complication strategies, such that the new problems can be generated with a controllable complexity.\nAs shown inFigure 2, our mutation mechanism can properly adjust the complexity of generated problems, and the exposure to more complex math problems can improve the LLM’s reasoning capability.\nMoreover, to ensure the data validity and achieve higher generation diversity, we combine the symbolic solving with the random sampling through the projected Markov chain Monte Carlo technique[14,15].\n\n\nFigure 2:The performance of our proposed mutation mechanism. The first figure illustrates that the generated problems with higher difficulty levels lead to more reasoning steps of GPT-4. The second figure shows that the gradual incorporation of more difficult problems consistently improves the LLM’s reasoning capability.\n\n\nEmpirical evaluation on GSM8K[10]and MATH[11]demonstrates the effectiveness of the proposed method.\nParticularly,\nwe use the proposed framework to generate a mathematical dataset of 620K examples for supervised fine-tuning.\nThe experimental results show that, the fine-tuned models on LLaMA-2[16]and Mistral-7B[17]significantly outperform the existing open-source counterparts on both GSM8K and MATH datasets, as well as two out-of-domain datasets SVAMP[12]and ASDiv[18].\nOn the GSM8K dataset, the model fine-tuned on Mistral-7B even outperforms GPT-3.5-Turbo (by 2.4%), a proprietary model with an order of magnitude larger parameters.\nAdditionally, we evaluate the scalability of our method, and observe consistent performance improvements, as the size of training data increases. This upward trajectory suggests a promising avenue for further enhancing LLMs’ mathematical capabilities.\n\n",
    "related_work": "Recent surveys[37,38,39]have comprehensively discussed the current advances in the mathematical reasoning of LLMs.\nHere, we review three main lines of existing work on enhancing the mathematical reasoning for LLMs related to our study: prompt-based methods, rephrasing-based methods, and tool-based methods.\n\n\nPrompt-based Method.\nPrompt-based methods aim to harness the inherent capabilities of LLMs by carefully designing appropriate input prompts without tuning the model parameters.\nThis line of work starts from the observation that LLMs can effectively tackle more math problems when provided with a simple Chain-of-Thought (CoT) prompt, i.e., “Let’s think step by step”[40].\nBuilding upon the CoT prompt,Wang et al. [41]further propose to consolidate multiple reasoning paths based on the self-consistency of correct answers.\nLater, several researchers propose to prompt LLMs to decompose complex problems.\nFor example,Zhou et al. [42]introduce the least-to-most strategy that prompts LLMs to break down the original problem into a series of sub-problems.Khot et al. [43]further boost this strategy, by assigning each sub-problem to the corresponding LLM that is specifically optimized for it.\nFinally, few-shot prompting, e.g., Few-shot CoT[44]and Complex CoT[45], has also been studied to enhance the reasoning performance of LLMs. To further improve the few-shot prompting, the prompt retrieval is proposed to automatically select high-quality examples[46,47], while the prompt compression is explored to include more examples in restricted context by pruning each example[48].\n\n\nRephrasing-based Method. The second line of existing work aims to generate additional math data, based on which the mathematical reasoning capability of LLMs can be established via supervised fine-tuning. To address the data scarcity issue, current research mainly focuses on rephrasing the problem or the answer.\nFor the answer rephrasing,Magister et al. [49]adopt the PaLM and GPT-3 to generate CoT math data, resulting in improved performance of the T5 model on math reasoning tasks.\nTo mitigate the inclusion of incorrect answers during the supervised fine-tuning, RFT[50]introduces a rejection sampling strategy, whereas AFT[51]trains an LLM to categorize them.\nRegarding the problem rephrasing,\nWizardMath[30]proposes a reinforced evol-instruct method.\nIt instructs ChatGPT and trains a new LLM to rephrase the problem, equipped by a reward model for evaluating the quality of generated problems.\nCombining the rephrasing of problems and answers together,\nMuggleMATH[31]builds the AugGSM8K dataset based on prompting GPT-3.5 and GPT-4.\nMetaMath[33]develops a question bootstrapping method based on LLMs, unifying rephrasing, self-verification[52], FOBAR[53], and answer augmentation strategies, obtaining the MetaMathQA.\nXwin-Math[54]is a peer study that significantly enhances the reasoning capacity of LLMs using problems generated by GPT-4 Turbo. In contrast, our work focuses on generating verifiable problems through controllable mutations, rather than relying entirely on the GPT model.\n\n\nOur proposed method also falls into this category. In contrast to existing methods directly prompting LLMs to rephrase the problem, we mutate the problem in the formal symbolic space, resulting in a more controllable mutation mechanism that ensures both the validity and diversity of the generated problems.\nMoreover, the quality of reasoning paths is also guaranteed by the symbolic solvers.\n\n\nTool-based Method.Tool-based methods aim to enhance the math solving performance of LLMs by instructing them to use external math tools.\nFor instance,\nPoT (Program of Thought)[55]and PAL[56]propose to prompt the LLMs to delegate the computation to a program interpreter, which can be executed to obtain the final answer.\nTo further improve the tool-using ability of LLMs,\nMathCoder[57]constructs a math dataset containing problems and their code-based solutions for the supervised fine-tuning;\nMAmmoTH[32]builds a dataset that combines CoT and PoT reasoning, enabling LLMs to perform hybrid inference.\nSince the interaction with math tools can further boost the performance of LLMs,\nTVA[58]includes the Isabelle theorem prover to check each reasoning step and guide the reflection of LLMs;\nTora[59]generates interactive tool-use trajectories on mathematical datasets and then performs imitation learning on the annotations.\n\n\nOur proposed method shares some similarities with tool-based approaches as both involve symbolic solvers.\nHowever, rather than using external tools to solve mathematical problems, our approach aims to explore the inherent reasoning capability of LLMs. Therefore, symbolic solvers are only used to ensure the validity of the generated data as well as the correctness of the generated reasoning paths.\n\n",
    "origin_citations": [
      "Large language models for mathematical reasoning: Progresses and challenges.",
      "A survey of deep learning for mathematical reasoning.",
      "Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.",
      "Large language models are zero-shot reasoners.",
      "Self-consistency improves chain of thought reasoning in language models.",
      "Least-to-most prompting enables complex reasoning in large language models.",
      "Decomposed prompting: A modular approach for solving complex tasks.",
      "Chain-of-thought prompting elicits reasoning in large language models.",
      "Complexity-based prompting for multi-step reasoning.",
      "Automatic chain of thought prompting in large language models.",
      "Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning.",
      "Boosting llm reasoning: Push the limits of few-shot learning with reinforced in-context pruning.",
      "Teaching small language models to reason.",
      "Scaling relationship on learning mathematical reasoning with large language models, 2023.",
      "Making large language models better reasoners with alignment.",
      "Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct.",
      "Query and response augmentation cannot help out-of-domain math reasoning generalization.",
      "Metamath: Bootstrap your own mathematical questions for large language models.",
      "Large language models are better reasoners with self-verification.",
      "Forward-backward reasoning in large language models for mathematical verification.",
      "Common 7b language models already possess strong math capabilities.",
      "Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks.",
      "PAL: program-aided language models.",
      "Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning.",
      "Mammoth: Building math generalist models through hybrid instruction tuning.",
      "Don’t trust: Verify – grounding LLM quantitative reasoning with autoformalization.",
      "Tora: A tool-integrated reasoning agent for mathematical problem solving."
    ],
    "citations": [
      781004,
      139691,
      323463,
      7094,
      2191,
      424,
      5565,
      1578,
      5953,
      2467,
      176263,
      5959,
      181997,
      129866,
      406519,
      321775,
      3513,
      140092,
      781408,
      290150,
      243983,
      7957,
      322719,
      128244,
      423721,
      477284
    ],
    "extract_rate": 0.9629629629629629,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 78,
      "Final_Score": 80.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 78\nFinal_Score: 81"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 88,
      "Final_Score": 85.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 88\nFinal_Score: 85"
    },
    "overall_score": 83.32,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "CIcMZGLyZW",
        "forum": "CIcMZGLyZW",
        "number": 9836,
        "title": "Neuro-Symbolic Data Generation for Math Reasoning",
        "normalized_title": "neuro symbolic data generation for math reasoning",
        "abstract": "A critical question about Large Language Models (LLMs) is whether their apparent deficiency in mathematical reasoning is inherent, or merely a result of insufficient exposure to high-quality mathematical data. To explore this, we developed an automated method for generating high-quality, supervised mathematical datasets. The method carefully mutates existing math problems, ensuring both diversity and validity of the newly generated problems. This is achieved by a neuro-symbolic data generation framework combining the intuitive informalization strengths of LLMs, and the precise symbolic reasoning of math solvers along with projected Markov chain Monte Carlo sampling in the highly-irregular symbolic space.\nEmpirical experiments demonstrate the high quality of data generated by the proposed method, and that the LLMs, specifically LLaMA-2 and Mistral, when realigned with the generated data, surpass their state-of-the-art counterparts.",
        "authors": [
          "Zenan Li",
          "Zhi Zhou",
          "Yuan Yao",
          "Xian Zhang",
          "Yu-Feng Li",
          "Chun Cao",
          "Fan Yang",
          "Xiaoxing Ma"
        ],
        "decision": null,
        "pdate": 1727287922314,
        "cdate": 1715686478922
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "CIcMZGLyZW",
          "forum": "CIcMZGLyZW",
          "number": 9836,
          "title": "Neuro-Symbolic Data Generation for Math Reasoning",
          "normalized_title": "neuro symbolic data generation for math reasoning",
          "abstract": "A critical question about Large Language Models (LLMs) is whether their apparent deficiency in mathematical reasoning is inherent, or merely a result of insufficient exposure to high-quality mathematical data. To explore this, we developed an automated method for generating high-quality, supervised mathematical datasets. The method carefully mutates existing math problems, ensuring both diversity and validity of the newly generated problems. This is achieved by a neuro-symbolic data generation framework combining the intuitive informalization strengths of LLMs, and the precise symbolic reasoning of math solvers along with projected Markov chain Monte Carlo sampling in the highly-irregular symbolic space.\nEmpirical experiments demonstrate the high quality of data generated by the proposed method, and that the LLMs, specifically LLaMA-2 and Mistral, when realigned with the generated data, surpass their state-of-the-art counterparts.",
          "authors": [
            "Zenan Li",
            "Zhi Zhou",
            "Yuan Yao",
            "Xian Zhang",
            "Yu-Feng Li",
            "Chun Cao",
            "Fan Yang",
            "Xiaoxing Ma"
          ],
          "decision": null,
          "pdate": 1727287922314,
          "cdate": 1715686478922
        }
      ]
    }
  },
  {
    "paper_id": 42231,
    "title": "DN-4DGS: Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering",
    "abstract": "Dynamic scenes rendering is an intriguing yet challenging problem.\nAlthough current methods based on NeRF have achieved satisfactory performance, they still can not reach real-time levels.\nRecently, 3D Gaussian Splatting (3DGS) has garnered researchers’ attention due to their outstanding rendering quality and real-time speed.\nTherefore, a new paradigm has been proposed: defining a canonical 3D gaussians and deforming it to individual frames in deformable fields.\nHowever, since the coordinates of canonical 3D gaussians are filled with noise, which can transfer noise into the deformable fields, and there is currently no method that adequately considers the aggregation of 4D information.\nTherefore, we propose Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering (DN-4DGS).\nSpecifically, a Noise Suppression Strategy is introduced to change the distribution of the coordinates of the canonical 3D gaussians and suppress noise.\nAdditionally, a Decoupled Temporal-Spatial Aggregation Module is designed to aggregate information from adjacent points and frames.\nExtensive experiments on various real-world datasets demonstrate that our method achieves state-of-the-art rendering quality under a real-time level. Code is available athttps://github.com/peoplelu/DN-4DGS.",
    "introduction": "Dynamic scene reconstruction from single or multi-view videos is a crucial task in computer vision, with applications such as VR/ARkim2019immersive;pages2018affordable, 3D perceptionlu2024bsnet;li2024mamba24;deng2024diff3detr, movie productionhisatomi2009method, etc. Neural Radiance Fields (NeRF)mildenhall2021nerfoffer a promising approach by representing scenes with implicit functions derived from multi-view inputs. By incorporating time as an additional inputpumarola2021dnerf;li2022neural, NeRF enables dynamic scene rendering. However, the original NeRF model suffers from significant training and rendering costs, attributed to the high number of points sampled per camera ray and volume rendering.\n\n\nFigure 2:Comparison of our render visualization with 4DGaussianwu20234d.The results are rendered on HyperNeRFpark2021hypernerfdataset and use the point cloud provided by HyperNeRF for Gaussian initialization (S⁢p⁢a⁢r⁢s⁢e𝑆𝑝𝑎𝑟𝑠𝑒Sparseitalic_S italic_p italic_a italic_r italic_s italic_eI⁢n⁢i⁢t𝐼𝑛𝑖𝑡Inititalic_I italic_n italic_i italic_t). Image 1: canonical 3D gaussians generated by 4DGaussian. Image 2: deformable 3D gaussians generated by 4DGaussian. Image 3: canonical 3D gaussians generated by our method. Image 4: deformable 3D gaussians after the first stage. Image 5: deformable 3D gaussians after the second stage. Image 6: ground truth. The yellow box emphasizes that through a two-stage deformation process, our method can produce higher-quality rendering results.\n\n\nRecently, the emerging 3D Gaussian Splatting (3DGS)3dgshas significantly increased rendering speed to a real-time level compared to NeRFs by employing a differentiable rasterizer for 3D Gaussian primitives. 3DGS directly optimizes the parameters of 3D gaussians (position, opacity, anisotropic covariance, and spherical harmonics (SH) coefficients) and renders them through projection andα𝛼\\alphaitalic_α-blending. Given the explicit expression nature of 3DGS, recent studieswu20234d;yang2023deformable;huang2023screpresent dynamic scenes by defining a canonical 3D gaussians and deforming it to individual frames in deformable fields. Specifically, during the execution of the deformation field, the coordinatesx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zalong with timet𝑡titalic_tare used as input, and the output corresponds to the changes in Gaussian properties. It is noteworthy that the existing methods, when designing deformable networks, either directly map the 4D coordinates of each input point to the latent space using MLPsyang2023deformable;huang2023sc, or use HexPlanehexplaneto interpolate a series of learnable embeddings to obtain the latent of each pointwu20234d.\n\n\nBoth of these approaches have drawbacks that can not be ignored.\n1) Canonical 3D gaussians are synthesized from multi-frame images of dynamic scenes.\nDue to the presence of dynamic regions and the specific design of A (canonical 3D gaussians) + B (deformable network), canonical 3D gaussians exhibit significant noise, as illustrated in Figure2.\nThis noise is inevitably transferred to the deformable field after the inputx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zis passed through the deformable network. To elaborate on the “Noise”: In the canonical + deformable design, we input the canonical Gaussian coordinatesx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zand timet𝑡titalic_tinto the deformable network. The deformable network essentially performs basic operations (addition, subtraction, multiplication, division) on the coordinatesx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zand timet𝑡titalic_t. Since the point-to-point relationships within the canonical Gaussians are chaotic and erroneous, as shown in Figure2, it is predictable that feeding these erroneous coordinates into the deformable network will transfer this error into the deformation field, introducing inaccuracies in the final deformationsΔ⁢x,Δ⁢y,Δ⁢zΔ𝑥Δ𝑦Δ𝑧\\Delta x,\\Delta y,\\Delta zroman_Δ italic_x , roman_Δ italic_y , roman_Δ italic_z.\n2) There is a lack of feature aggregation for spatial-temporal information, yet due to the presence of noise in canonical 3D gaussians’x⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_z, direct feature aggregation for spatial information would further amplify noise, affecting the learning of the deformable field. Therefore, spatial aggregation after denoising is very crucial.\n\n\nTo address the aforementioned issues, we propose Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering (DN-4DGS),\nprimarily consisting of two components: the Noise Suppression Strategy (NSS) and the Decoupled Temporal-Spatial Aggregation Module (DTS).\nTo address the initial issue, the design of NSS incorporates two deformation operations.\nThe first deformation operation is a standard deformation. It takes the coordinatesx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zof the canonical 3D gaussians and timet𝑡titalic_tas input and outputs corresponding coordinate deformationsΔ⁢x,Δ⁢y,Δ⁢zΔ𝑥Δ𝑦Δ𝑧\\Delta x,\\Delta y,\\Delta zroman_Δ italic_x , roman_Δ italic_y , roman_Δ italic_z.\nThe second deformation builds upon the first by addingΔ⁢x,Δ⁢y,Δ⁢zΔ𝑥Δ𝑦Δ𝑧\\Delta x,\\Delta y,\\Delta zroman_Δ italic_x , roman_Δ italic_y , roman_Δ italic_zto the originalx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_z, creating a modified set of coordinates that is then input into a new feature extraction network. The entire process is illustrated by image 3, 4 and 5 in Figure2.\nThis strategy achieves a successful alteration of the distribution of coordinatesx⁢y⁢z𝑥𝑦𝑧xyzitalic_x italic_y italic_zthrough the initial coordinate deformation, resulting in noise reduction and the generation of a more accurate deformation field.\nIt’s worth noting that during the early stage of training, we only perform the first deformation operation.\nOnly after achieving acceptable results with the first deformation operation do we proceed to the second deformation operation to further enhance accuracy.\nTo address the second problem, we design the DTS.\nThe reason we decouple spatial-temporal aggregation is due to the presence of noise in the coordinates of the canonical 3D gaussians.\nIf we directly perform spatial aggregation on the coordinates of the canonical 3D gaussians, the noise information is inevitably amplified after a series of aggregation operations such as k-nearest neighbors (KNN)qi2017pointnet++, significantly affecting the results of the deformation field.\nTherefore, based on the first design NSS, we conduct spatial aggregation during the second deformation operation.\nConsidering that temporal information is unrelated to the canonical 3D gaussians, temporal aggregation can be directly incorporated into the first deformation operation to enhance feature extraction capabilities.\nIn order to reduce computational overhead and considering that temporal information has already been effectively extracted in the first deformation operation, we do not perform temporal aggregation in the second deformation operation.\nIn conclusion, our main contributions are outlined as follows:\n\n\n(i) We introduce a novel representation called Denoised Deformable Network with Temporal-Spatial Aggregation for high-fidelity and efficient dynamic scene rendering.\n\n\n(ii) We promose the Noise Suppression Strategy, which can change the distribution of the coordinates of the canonical 3D gaussians, suppress noise and generate a more precise deformation field.\n\n\n(iii) We promose the Decoupled Temporal-Spatial Aggregation Module to aggregate information from adjacent points and frames.\n\n\n(iv) Extensive experiments on various real-world datasets demonstrate that our method achieves state-of-the-art rendering quality under a real-time level.\n\n",
    "related_work": "Dynamic NeRF.Novel view synthesis has been a hot topic in academia for several years. NeRFmildenhall2021nerfmodels static scenes implicitly using MLPs, and numerous studiesguo2023forward;li2021neural;park2021nerfies;park2021hypernerf;pumarola2021dnerf;tretschk2021non;xian2021spacehave extended its application to dynamic scenes through a canonical 3D grid structure and a deformation field. HyperNeRFpark2021hypernerfmodels object topology deformation using higher-dimensional inputs, while DyNeRFli2022neuralemploys time-conditioned NeRF to represent a 4D scene. However, these approaches, based on vanilla NeRF, suffer from high computational costs due to ray point sampling and volume rendering.\nTo address this issue, several acceleration methodsliu2022devrf;li2023dynibar;lin2022efficient;lin2023im4d;lombardi2021mixture;peng2023representing;hexplane;fang2022fast;fridovich2023k;tensor4d;wang2023mixedvoxels;wang2023neuralhave been proposed for rendering dynamic scenes. DeVRFliu2022devrfintroduces a grid representation, while IBR-based methodslin2022efficient;lin2023im4dutilize multi-camera information for improved quality and efficiency. TensorRFchen2022tensorfadopts multiple planes as explicit representations for direct dynamic scene modeling. Recent approaches such as K-Planeskplanes, Tensor4Dtensor4d, and HexPlanehexplanehave also been proposed. NeRFPlayersong2023nerfplayerintroduces a unified streaming representation for both grid-based and plane-based methods, utilizing separate models to differentiate static and dynamic scene components, albeit at the cost of slow rendering times. HyperReelattal2023hyperreelsuggests a flexible sampling network coupled with two planes for dynamic scene representation. Despite the improvements in training and rendering speed achieved by these methods, they still fall short of meeting real-time requirements.\n\n\nDynamic Gaussian Splatting.Recently, 3D Gaussian Splatting (3DGS)3dgshas garnered increasing attention from researchers due to its superior rendering quality and real-time rendering speed.\nThe method employs a soft point representation with attributes including position, rotation, density, and radiance, and utilizes differentiable point-based rendering for scene optimization.\nSoon after, several concurrent workswu20234d;yang2023deformable;huang2023sc;zhu2024motiongshave adapted 3D Gaussians for dynamic scenes.\nThese methods represent dynamic scenes by establishing a canonical 3DGS and deforming it to individual frames using deformable fields.\nYang et al.yang2023deformablepredict per-Gaussian offsets using an additional MLP on canonical 3D gaussians, while Wu et al.wu20234dsubstitute the MLP with multi-resolution HexPlaneshexplaneand a lightweight MLP.\nOur work introduces the Noise Suppression Strategy to change the distribution of the coordinates of the canonical 3D gaussians and generate a more precise deformation field.\nAdditionally, for better aggregation of temporal-spatial information, we propose the Decoupled Temporal-Spatial Aggregation Module to consolidate information from adjacent points and frames.\n\n",
    "origin_citations": [
      "Nerf: Representing scenes as neural radiance fields for view synthesis.",
      "Forward flow for novel view synthesis of dynamic scenes.",
      "Neural scene flow fields for space-time view synthesis of dynamic scenes.",
      "Nerfies: Deformable neural radiance fields.",
      "Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields.",
      "D-nerf: Neural radiance fields for dynamic scenes.",
      "Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video.",
      "Space-time neural irradiance fields for free-viewpoint video.",
      "Neural 3d video synthesis from multi-view video.",
      "Devrf: Fast deformable voxel radiance fields for dynamic scenes.",
      "Dynibar: Neural dynamic image-based rendering.",
      "Efficient neural radiance fields for interactive free-viewpoint video.",
      "Im4d: High-fidelity and real-time novel view synthesis for dynamic scenes.",
      "Mixture of volumetric primitives for efficient neural rendering.",
      "Representing volumetric videos as dynamic mlp maps.",
      "Hexplane: A fast representation for dynamic scenes.",
      "Fast dynamic radiance fields with time-aware neural voxels.",
      "K-planes: Explicit radiance fields in space, time, and appearance.",
      "Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering.",
      "Mixed neural voxels for fast multi-view video synthesis.",
      "Neural residual radiance fields for streamably free-viewpoint videos.",
      "Tensorf: Tensorial radiance fields.",
      "Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields.",
      "Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "4d gaussian splatting for real-time dynamic scene rendering.",
      "Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction.",
      "Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes.",
      "Motiongs: Exploring explicit motion guidance for deformable 3d gaussian splatting."
    ],
    "citations": [
      253579,
      477314,
      431388,
      431438,
      392944,
      431037,
      105722,
      431437,
      217861,
      714692,
      244683,
      152912,
      320305,
      218142,
      18047,
      213655,
      145318,
      213434,
      244427,
      223968,
      19107,
      502621,
      333543,
      303364,
      180969,
      320337,
      479493,
      273928
    ],
    "extract_rate": 0.9655172413793104,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 85,
      "Final_Score": 82.9,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 85\nFinal_Score: 83"
    },
    "related_work_scores": {
      "Content_Coherence": 80,
      "Citation_Quality": 92,
      "Synthesis_Analysis": 75,
      "Final_Score": 83.30000000000001,
      "raw_text": "Content_Coherence: 80\nCitation_Quality: 92\nSynthesis_Analysis: 75\nFinal_Score: 82"
    },
    "overall_score": 83.14000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "QQSyNX5s83",
        "forum": "QQSyNX5s83",
        "number": 11359,
        "title": "DN-4DGS: Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering",
        "normalized_title": "dn 4dgs denoised deformable network with temporal spatial aggregation for dynamic scene rendering",
        "abstract": "Dynamic scenes rendering is an intriguing yet challenging problem. Although current methods based on NeRF have achieved satisfactory performance, they still can not reach real-time levels. Recently, 3D Gaussian Splatting (3DGS) has garnered researchers' attention due to their outstanding rendering quality and real-time speed. Therefore, a new paradigm has been proposed: defining a canonical 3D gaussians and deforming it to individual frames in deformable fields. However, since the coordinates of canonical 3D gaussians are filled with noise, which can transfer noise into the deformable fields, and there is currently no method that adequately considers the aggregation of 4D information. Therefore, we propose Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering (DN-4DGS). Specifically, a Noise Suppression Strategy is introduced to change the distribution of the coordinates of the canonical 3D gaussians and suppress noise. Additionally, a Decoupled Temporal-Spatial Aggregation Module is designed to aggregate information from adjacent points and frames. Extensive experiments on various real-world datasets demonstrate that our method achieves state-of-the-art rendering quality under a real-time level. Code is available at https://github.com/peoplelu/DN-4DGS.",
        "authors": [
          "Jiahao Lu",
          "Jiacheng Deng",
          "Ruijie Zhu",
          "Yanzhe Liang",
          "Wenfei Yang",
          "Xu Zhou",
          "Tianzhu Zhang"
        ],
        "decision": null,
        "pdate": 1727287968783,
        "cdate": 1715703913341
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "QQSyNX5s83",
          "forum": "QQSyNX5s83",
          "number": 11359,
          "title": "DN-4DGS: Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering",
          "normalized_title": "dn 4dgs denoised deformable network with temporal spatial aggregation for dynamic scene rendering",
          "abstract": "Dynamic scenes rendering is an intriguing yet challenging problem. Although current methods based on NeRF have achieved satisfactory performance, they still can not reach real-time levels. Recently, 3D Gaussian Splatting (3DGS) has garnered researchers' attention due to their outstanding rendering quality and real-time speed. Therefore, a new paradigm has been proposed: defining a canonical 3D gaussians and deforming it to individual frames in deformable fields. However, since the coordinates of canonical 3D gaussians are filled with noise, which can transfer noise into the deformable fields, and there is currently no method that adequately considers the aggregation of 4D information. Therefore, we propose Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering (DN-4DGS). Specifically, a Noise Suppression Strategy is introduced to change the distribution of the coordinates of the canonical 3D gaussians and suppress noise. Additionally, a Decoupled Temporal-Spatial Aggregation Module is designed to aggregate information from adjacent points and frames. Extensive experiments on various real-world datasets demonstrate that our method achieves state-of-the-art rendering quality under a real-time level. Code is available at https://github.com/peoplelu/DN-4DGS.",
          "authors": [
            "Jiahao Lu",
            "Jiacheng Deng",
            "Ruijie Zhu",
            "Yanzhe Liang",
            "Wenfei Yang",
            "Xu Zhou",
            "Tianzhu Zhang"
          ],
          "decision": null,
          "pdate": 1727287968783,
          "cdate": 1715703913341
        }
      ]
    }
  },
  {
    "paper_id": 26620,
    "title": "Do’s and Don’ts:Learning Desirable Skills with Instruction Videos",
    "abstract": "Unsupervised skill discovery is a learning paradigm that aims to acquire diverse behaviors without explicit rewards.\nHowever, it faces challenges in learning complex behaviors and often leads to learning unsafe or undesirable behaviors.\nFor instance, in various continuous control tasks, current unsupervised skill discovery methods succeed in learning basic locomotions like standing but struggle with learning more complex movements such as walking and running.\nMoreover, they may acquire unsafe behaviors like tripping and rolling or navigate to undesirable locations such as pitfalls or hazardous areas.\nIn response, we presentDoDont(Do’s and Dont’s), an instruction-based skill discovery algorithm composed of two stages.\nFirst, in an instruction learning stage, DoDont leverages action-free instruction videos to train an instruction network to distinguish desirable transitions from undesirable ones.\nThen, in the skill learning stage, the instruction network adjusts the reward function of the skill discovery algorithm to weight the desired behaviors.\nSpecifically, we integrate the instruction network into a distance-maximizing skill discovery algorithm, where the instruction network serves as the distance function.\nEmpirically, with less than 8 instruction videos, DoDont effectively learns desirable behaviors and avoids undesirable ones across complex continuous control tasks.\nCode and videos are available athttps://mynsng.github.io/dodont/",
    "introduction": "Recent advancements in unsupervised pre-training methodologies have led to the creation of large-scale foundational models across diverse domains, including computer vision[3,41]and natural language processing[7,11,22]. These methodologies exploit self-supervised learning objectives to extract meaningful representations without using explicit, supervised learning signals. In attempts to expand this paradigm to reinforcement learning, researchers have explored crafting self-supervised objectives to develop foundational policies capable of learning diverse behaviors without predefined reward signals, termed unsupervised skill discovery (USD)[1,14,9,21,43,31,24,23,27,35,38,39].\n\n\nDespite notable advancements in USD methodologies, acquiring foundational policies in environments with large state and action spaces (e.g., multi-jointed quadrupeds) remains a significant challenge.\nTwo major issues arise when training agents with USD in these complex environments. First, the vast state and action spaces allow the agent to acquire a wide variety of behaviors. While learning simple behaviors like standing may be feasible, acquiring complex behaviors that require intricate joint movements, such as walking or running, can take an exceedingly long time. Second, agents can develop undesirable and risky behaviors during training, such as tripping, rolling, or navigating hazardous areas like pitfalls. These challenges raise a key question:Is the purely unsupervised assumption of USD ideal for learning foundational policies in the real world?\n\n\nAccording to social learning theory, humans learn behaviors through both internal and external motivations[4,5,16]. Internally, they seek to perform diverse behaviors they haven’t tried before. Externally, they adopt socially desirable behaviors and avoid those that are not. In the skill discovery algorithm, this can be mimicked by combining internal motivation (USD objective) with external motivation based on a human’s intention. A straightforward way to incorporate external motivation in USD is to use a hand-designed reward function[26,10]. However, hand-designing reward functions are not scalable for learning diverse and desirable behaviors for several reasons. One of the reasons is the complexity of designing an intricate reward signal for guiding a desired behavior, and another is the difficulty of balancing multiple behavioral intentions into a single reward function.\n\n\nTo address these challenges, we proposeDoDont, a skill discovery algorithm that integrates USD objectives with intended behavioral goals. Instead of relying on a hand-designed reward, DoDontlearns a reward function from a small set of instruction videosthat demonstrate desirable and undesirable behaviors. Videos are chosen because they are inexpensive to collect and do not require action or reward labels[13,46].\n\n\nAs illustrated in Figure1, DoDont starts by collecting instruction videos of desirable (Do’s) and undesirable behaviors (Don’ts).\nWe then train an instruction network that assigns higher values to desirable behaviors and lower values to undesirable ones. This network re-weights the internal USD objective for the skill discovery phase. We utilize a distance-maximizing skill discovery algorithm as our main USD objective[35,38,39]where the instruction network serves as the distance metric.\n\n\nTo validate DoDont, we conduct experiments on various continuous control tasks that require complex locomotion (e.g., Cheetah and Quadruped[45]) or precise manipulation (e.g., Kitchen[17]). Our results show that with fewer than eight instruction videos, DoDont effectively learns complex locomotion skills (e.g., running quadruped), which are challenging to acquire with standard USD algorithms[39].\nMoreover, our instruction network effectively captures human intentions better than hand-designed reward functions in balancing multiple behavioral objectives.\nAdditionally, DoDont learns diverse skills while avoiding undesirable and unsafe behaviors (e.g., backflipping), which are difficult to prevent even with previous skill discovery algorithms that utilize prior knowledge[26].\n\n\nFigure 1:(a)The offline instruction video dataset includes videos of desirable behaviors (Do’s) and undesirable behaviors (Don’ts).(b)Unsupervised skill discovery algorithms tend to learn undesirable behaviors.(c)In DoDont, an instruction network is first trained with the Do’s and Don’ts videos to distinguish desirable and undesirable behaviors. Then, this instruction network adjusts the intrinsic reward of the skill discovery algorithm, promoting desirable skills while avoiding undesirable ones.\n\n",
    "related_work": "2.1Learning diverse behaviors without pre-defined taskNumerous unsupervised skill discovery (USD) methods have been developed to create foundational policies capable of learning diverse behaviors without pre-defined tasks. The most common approach involves maximizing mutual information (MI) between states and skills (I(S;Z)=DK⁢L(p(s,z)||p(s)p(z))I(S;Z)=D_{KL}(p(s,z)||p(s)p(z))italic_I ( italic_S ; italic_Z ) = italic_D start_POSTSUBSCRIPT italic_K italic_L end_POSTSUBSCRIPT ( italic_p ( italic_s , italic_z ) | | italic_p ( italic_s ) italic_p ( italic_z ) ))[14,43,31,9,27]. This is typically done by using an auxiliary neural networkqθ⁢(z|s)subscript𝑞𝜃conditional𝑧𝑠q_{\\theta}(z|s)italic_q start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( italic_z | italic_s )to estimate a lower bound ofI⁢(S;Z)𝐼𝑆𝑍I(S;Z)italic_I ( italic_S ; italic_Z )(I⁢(S;Z)≥𝔼⁢z,s⁢[log⁡qθ⁢(z|s)]𝐼𝑆𝑍𝔼𝑧𝑠delimited-[]subscript𝑞𝜃conditional𝑧𝑠I(S;Z)\\geq\\mathbb{E}{z,s}[\\log q_{\\theta}(z|s)]italic_I ( italic_S ; italic_Z ) ≥ blackboard_E italic_z , italic_s [ roman_log italic_q start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( italic_z | italic_s ) ]). This network acts as a skill discriminator, predicting the skillz𝑧zitalic_zfrom the states𝑠sitalic_s, which encourages the policy to create distinct trajectories for different skills and promotes learning a wide range of skills.\nHowever, these methods often struggle to learn complex, dynamic skills, as the MI objective can be met with simple, static skills, leaving agents unmotivated to explore more intricate behaviors[44,23,30,31,24].To overcome this problem, distance-maximizing skill discovery (DSD) methods have been introduced[35,38,39]. Instead of maximizing MI between states and skills, DSD explicitly maximizes a predefined distance function of statesd:𝒮×𝒮→ℝ0+:𝑑→𝒮𝒮superscriptsubscriptℝ0d:\\mathcal{S}\\times\\mathcal{S}\\rightarrow\\mathbb{R}_{0}^{+}italic_d : caligraphic_S × caligraphic_S → blackboard_R start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT start_POSTSUPERSCRIPT + end_POSTSUPERSCRIPT.\nFor instance, LSD[35]uses Euclidean distance between states to encourage agents to move further distances (dLSD=‖s′−s‖subscript𝑑LSDnormsuperscript𝑠′𝑠d_{\\text{LSD}}=||s^{\\prime}-s||italic_d start_POSTSUBSCRIPT LSD end_POSTSUBSCRIPT = | | italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT - italic_s | |). CSD[38]employs a density model-based distance, promoting agents to visit less frequented states (dCSD=−log⁡qθ⁢(s′|s)subscript𝑑CSDsubscript𝑞𝜃conditionalsuperscript𝑠′𝑠d_{\\text{CSD}}=-\\log q_{\\theta}(s^{\\prime}|s)italic_d start_POSTSUBSCRIPT CSD end_POSTSUBSCRIPT = - roman_log italic_q start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPT | italic_s ), whereqθsubscript𝑞𝜃q_{\\theta}italic_q start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPTis the density model). METRA[39]defines the distance temporally, pushing agents to learn skills that are temporally far apart (dMETRAsubscript𝑑METRAd_{\\text{METRA}}italic_d start_POSTSUBSCRIPT METRA end_POSTSUBSCRIPT= minimum number of environment steps to reachs′superscript𝑠′s^{\\prime}italic_s start_POSTSUPERSCRIPT ′ end_POSTSUPERSCRIPTfroms𝑠sitalic_s). DoDont is built upon the DSD framework, which will be elaborated in Sections3and4.\n\n\n2.2Learning diverse behaviors with pre-defined taskIn contrast to USD, there is a line of research focused on learning policies that exhibit diverse behaviors by integrating intrinsic motivation (e.g., MI-based rewards) with human intention as the form of task-specific reward functions[10,26,55,54]or using demonstration datasets[40]. While traditional RL methods typically aim to find a single optimal policy, these approaches aim to learn multiple policiesπθ⁢(a|s,zi=1:N)subscript𝜋𝜃conditional𝑎𝑠subscript𝑧:𝑖1𝑁\\pi_{\\theta}(a|s,z_{i=1:N})italic_π start_POSTSUBSCRIPT italic_θ end_POSTSUBSCRIPT ( italic_a | italic_s , italic_z start_POSTSUBSCRIPT italic_i = 1 : italic_N end_POSTSUBSCRIPT ). This allows for multiple ways to solve given tasks, resulting in increased robustness to environmental changes and the development of policies with distinct characteristics.SMERL[26]and DGPO[10]combine intrinsic and task rewards, maximizing their sum when task rewards exceed a given threshold. RSPO[55]iteratively finds novel policies by switching between task rewards and intrinsic diversity rewards, ensuring each new policy is distinct from previous ones. DOMiNO[54]addresses a constrained optimization problem by maximizing intrinsic diversity rewards while ensuring that all policies achieve sufficiently high performance (task rewards as constraints). ASE[40]utilizes an action-free demonstration dataset instead of a task reward function, minimizing a behavioral cloning loss while simultaneously maximizing a diversity intrinsic reward. DoDont also employs action-free demonstration datasets to extract human prior knowledge through an instruction network, pioneering the incorporation of both positive and negative behavior examples to guide learning towards desired behaviors and away from undesirable ones.\n\n",
    "origin_citations": [
      "Diversity is all you need: Learning skills without a reward function.",
      "Dynamics-aware unsupervised discovery of skills.",
      "Aps: Active pretraining with successor features.",
      "Explore, discover and learn: Unsupervised discovery of state-covering skills.",
      "Cic: Contrastive intrinsic control for unsupervised skill discovery.",
      "Learning more skills through optimistic exploration.",
      "Direct then diffuse: Incremental unsupervised skill discovery for state covering and goal reaching.",
      "Efficient exploration via state marginal matching.",
      "Learning to discover skills through guidance.",
      "Lipschitz-constrained unsupervised skill discovery.",
      "Controllability-aware unsupervised skill discovery.",
      "Metra: Scalable unsupervised rl with metric-aware abstraction.",
      "Dgpo: discovering multiple strategies with diversity-guided policy optimization.",
      "One solution is not all you need: Few-shot extrapolation via structured maxent rl.",
      "Continuously discovering novel strategies via reward-switching policy optimization.",
      "Discovering policies with domino: Diversity optimization maintaining near optimality.",
      "Ase: large-scale reusable adversarial skill embeddings for physically simulated characters."
    ],
    "citations": [
      389262,
      159461,
      1004429,
      131822,
      91212,
      408495,
      512015,
      268251,
      1090689,
      90918,
      489789,
      320160,
      12782,
      435160,
      65682,
      146237,
      379140
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 78,
      "Final_Score": 80.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 78\nFinal_Score: 81"
    },
    "related_work_scores": {
      "Content_Coherence": 85,
      "Citation_Quality": 82,
      "Synthesis_Analysis": 88,
      "Final_Score": 84.7,
      "raw_text": "Content_Coherence: 85\nCitation_Quality: 82\nSynthesis_Analysis: 88\nFinal_Score: 84"
    },
    "overall_score": 83.14,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "7X5zu6GIuW",
        "forum": "7X5zu6GIuW",
        "number": 10744,
        "title": "Do's and Don'ts: Learning Desirable Skills with Instruction Videos",
        "normalized_title": "do s and don ts learning desirable skills with instruction videos",
        "abstract": "Unsupervised skill discovery is a learning paradigm that aims to acquire diverse behaviors without explicit rewards. However, it faces challenges in learning complex behaviors and often leads to learning unsafe or undesirable behaviors. For instance, in various continuous control tasks, current unsupervised skill discovery methods succeed in learning basic locomotions like standing but struggle with learning more complex movements such as walking and running. Moreover, they may acquire unsafe behaviors like tripping and rolling or navigate to undesirable locations such as pitfalls or hazardous areas. In response, we present **DoDont** (Do’s and Dont’s), an instruction-based skill discovery algorithm composed of two stages. First, in instruction learning stage, DoDont leverages action-free instruction videos to train an instruction network to distinguish desirable transitions from undesirable ones. Then, in the skill learning stage, the instruction network adjusts the reward function of the skill discovery algorithm to weight the desired behaviors. \nSpecifically, we integrate the instruction network into a distance-maximizing skill discovery algorithm, where the instruction network serves as the distance function. Empirically, with less than 8 instruction videos, DoDont effectively learns desirable behaviors and avoids undesirable ones across complex continuous control tasks. Code and videos are available at https://mynsng.github.io/dodont/",
        "authors": [
          "Hyunseung Kim",
          "Byungkun Lee",
          "Hojoon Lee",
          "Dongyoon Hwang",
          "Donghu Kim",
          "Jaegul Choo"
        ],
        "decision": null,
        "pdate": 1727287950173,
        "cdate": 1715697244947
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "7X5zu6GIuW",
          "forum": "7X5zu6GIuW",
          "number": 10744,
          "title": "Do's and Don'ts: Learning Desirable Skills with Instruction Videos",
          "normalized_title": "do s and don ts learning desirable skills with instruction videos",
          "abstract": "Unsupervised skill discovery is a learning paradigm that aims to acquire diverse behaviors without explicit rewards. However, it faces challenges in learning complex behaviors and often leads to learning unsafe or undesirable behaviors. For instance, in various continuous control tasks, current unsupervised skill discovery methods succeed in learning basic locomotions like standing but struggle with learning more complex movements such as walking and running. Moreover, they may acquire unsafe behaviors like tripping and rolling or navigate to undesirable locations such as pitfalls or hazardous areas. In response, we present **DoDont** (Do’s and Dont’s), an instruction-based skill discovery algorithm composed of two stages. First, in instruction learning stage, DoDont leverages action-free instruction videos to train an instruction network to distinguish desirable transitions from undesirable ones. Then, in the skill learning stage, the instruction network adjusts the reward function of the skill discovery algorithm to weight the desired behaviors. \nSpecifically, we integrate the instruction network into a distance-maximizing skill discovery algorithm, where the instruction network serves as the distance function. Empirically, with less than 8 instruction videos, DoDont effectively learns desirable behaviors and avoids undesirable ones across complex continuous control tasks. Code and videos are available at https://mynsng.github.io/dodont/",
          "authors": [
            "Hyunseung Kim",
            "Byungkun Lee",
            "Hojoon Lee",
            "Dongyoon Hwang",
            "Donghu Kim",
            "Jaegul Choo"
          ],
          "decision": null,
          "pdate": 1727287950173,
          "cdate": 1715697244947
        }
      ]
    }
  },
  {
    "paper_id": 25046,
    "title": "SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models",
    "abstract": "Large language models (LLMs) achieve remarkable performance in natural language understanding but require substantial computation resources and memory footprint. Post-training quantization (PTQ) is a powerful compression technique extensively investigated for its effectiveness in reducing memory usage and improving the inference efficiency of LLMs.\nHowever, existing PTQ methods are still not ideal in terms of accuracy and efficiency, especially with below 4 bit-widths.\nStandard PTQ methods using group-wise quantization suffer difficulties in quantizing LLMs accurately to such low-bit, but advanced methods remaining high-precision weights element-wisely are hard to realize its theoretical hardware efficiency.\nThis paper presents a Salience-Driven Mixed-Precision Quantization scheme for LLMs, namelySliM-LLM. The scheme exploits the salience distribution of LLM weights to determine optimal bit-width and quantizers for accurate LLM quantization, while aligning bit-width partition to quantization groups for compact memory usage and fast integer computation on hardware inference.\nSpecifically, the proposed SliM-LLM mainly relies on two novel techniques: (1)Salience-Determined Bit Allocationutilizes the clustering characteristics of salience distribution to allocate the bit-widths of each quantization group. This increases the accuracy of quantized LLMs and maintains the inference efficiency high; (2)Salience-Weighted Quantizer Calibrationoptimizes the parameters of the quantizer by considering the element-wise salience within the group. This balances the maintenance of salient information and minimization of errors. Comprehensive experiments show that SliM-LLM significantly improves the accuracy of various LLMs at ultra-low 2-3 bits,e.g., 2-bit LLaMA-7B achieves a 5.5-times memory-saving compared to the original model on NVIDIA A800 GPUs, and 48% decrease of perplexity compared to the state-of-the-art gradient-free PTQ method. Moreover, SliM-LLM+, which is integrated from the extension of SliM-LLM with gradient-based quantizers, further reduces perplexity by 35.1%. We highlight that the structurally quantized features of SliM-LLM exhibit remarkable versatility and promote improvements in the accuracy of quantized LLMs while keeping inference efficiency on hardware.\nOur code is available athttps://github.com/Aaronhuang-778/SliM-LLM.",
    "introduction": "Large language models (LLMs) have exhibited exceptional performance across a wide array of natural language benchmarks[3,48,19,2].\nNotably, LLaMA[41]and GPT[3]series have significantly contributed to the ongoing evolution of LLMs towards universal language intelligence.\nThe powerful language understanding capabilities of LLMs have been transferred to multi-modal domains[25,1,40,50], laying the foundation for artificial general intelligence[4].\nDespite these significant achievements, the substantial computational and memory requirements of LLMs pose efficiency challenges for real-world applications and deployments, particularly in resource-constrained environments. For example, the latest LLaMA-3-70B111https://github.com/meta-llama/llama3model, with its 70 billion parameters, requires over 150GB of storage and a minimum of two NVIDIA A800 GPUs, each with 80GB of memory, for inference[21].\n\n\nTo reduce the computation burden, post-training quantization (PTQ), as an efficient and effective compression approach[11], has also been explored and proven successful in quantizing the weights of pre-trained LLMs[16,20,26,38,23,6].\nFaced with the dilemma of scaled-up LLMs and the limited computation resources, there is an urgent need for more aggressive compression[20,43].\nHowever, despite considerable efforts, significant performance degradation still occurs in low bit-width scenarios (⩽\\leqslant⩽3-bit).\nTo maintain the performance, unstructured mixed-precision quantization schemes[37,20,12]or specialized transformation computations[6,43,13,7]are necessary. Yet, these approaches impose additional burdens on hardware during the inference of quantized LLMs,\nsuffering memory overhead of element-wise bit-maps and computation overhead of codebook decoding and bit-map addressing (even preventing efficient integer computation).\nMoreover, even though fine-tuning can improve the accuracy of quantized LLMs, it increases the overfitting risk and requires expensive computation resources and a long time[26,5].\nConsequently, ensuring the accuracy of LLMs while maintaining efficiency during deployment remains a significant challenge for current PTQ approaches.\n\n\nFigure 1:(a) The perplexity (↓↓\\downarrow↓) of existing low-bit PTQ methods of LLaMA at 2-bit. Solid-line indicates structured quantization methods. (b) Compare PTQ methods with gradient quantizer at 3-bit. (c) Features of current low-bit quantization methods.Cdenotes codebook-based,Sis statistic-based, andGrepresents gradient-based quantizers.\n\n\nThis paper presents theSalience-DrivenMixed-Precision LLM (SliM-LLM) framework, an accurate and hardware-efficient PTQ method for LLMs (⩽\\leqslant⩽3-bit). SliM-LLM can be seamlessly integrated into existing advanced PTQ pipelines[16,38], as a plug-and-play approach with mixed-precision computing for improved performance (Fig.1). Our approach builds on the observation that not all parameters are equally important[12,20,37]. Specifically, a subset of salient weights significantly influences an LLM’s capabilities and tends to be concentrated in specific channels. During structured group-wise quantization, the uneven distribution of these salient channels leads to differential importance across various groups. Based on this finding, we design a structured mixed-precision quantization approach for LLMs.\nFirst, we develop a novelSalience-Determined Bit Allocation(SBA) method to allocate the optimal bit-width configuration for each structured group based on the salience distribution, minimizing the weight output relative entropy. By implementing bit-width compensation constraints, SBA maintains the average bit-width, while improving the low-bit performance.\nNext, we introduce theSalience-Weighted Quantizer Calibration(SQC), which amplifies the awareness of locally salient weights, preventing the degradation of sensitive information within groups. SQC works collaboratively with SBA, exploiting the local and global salience of weights to preserve the performance of LLMs after quantization. Notably, SliM-LLM does not rely on fine-tuning processes, efficiently deploying weight quantization on various LLMs. Moreover, compared to the unstructured mixed-precision methods[37,12,20], SliM-LLM incurs no additional bits and computation overhead. We also deploy SliM-LLM on the application-level inference tool222https://github.com/AutoGPTQ/AutoGPTQfor LLMs, facilitating mixed-precision inference on graphics processing units (GPUs) with high performance.\n\n\nExperiments show that for various LLM families, SliM-LLM surpasses existing PTQ methods on diverse benchmarks as a plug-and-play unit, particularly in low-bit scenarios. Using GPTQ as the backbone, SliM-LLM improves the perplexity scores of 2-bit LLaMA-13B and LLaMA2-13B on WikiText2[30]from 20.44 and 28.14 to 8.87 and 9.41, denoting performance improvements of over 56%, respectively.\nSliM-LLM even outperforms other unstructured mixed-precision PTQ methods, such as PB-LLM[37], APTQ[18]and LLM-MQ[24], in a deployment-friendly manner, showcasing its superior low-bit accuracy and efficiency. Moreover, we integrate SliM-LLM into OmniQuant[38]and obtain SliM-LLM+through gradient optimization to further improve quantization quality.\n\n",
    "related_work": "Large Language Models(LLMs) have been significantly developed in diverse natural language processing domains, establishing a prominent paradigm in these fields[4,5,51,3,41]. Nevertheless, the exceptional success of LLMs depends on massive parameters and computations, posing significant challenges for deployment in resource-constrained environments. Consequently, research into the compression of LLMs has emerged as a promising field. Existing compression techniques for LLMs primarily include low-bit quantization, pruning, distillation, and low-rank decomposition[46,17,16,44,38,6,52,15,20,33,7]. Among these technologies, low-bit quantization gains remarkable attention, for efficiently reducing the model size without change of network structure[52,51,5].\n\n\nFigure 2:Illustration of our proposed SliM-LLM. TheSalience-Determined Bit Allocation(SBA) optimizes activation-aware structured precision, optimizing the global information distribution in quantization.Salience-Weighted Quantizer Calibration(SQC) detects discretely distributed salient weights, enhancing the local important information in LLMs.\n\n\nQuantization of LLMscan be generally divided into quantization-aware training (QAT)[27]and post-training quantization (PTQ)[44,16,38].\nQAT, by employing a retraining strategy based on quantized perception, better preserves the performance of quantized models. LLM-QAT[27]addresses the data obstacle issue in QAT through data-free distillation. However, for LLMs with huge size of parameters, the cost of retraining is extremely inefficient[5]. Therefore, PTQ has become a more efficient choice for LLMs. For instance, LLM.int8()[27]and ZeroQuant[47]explore the quantization strategies for LLMs in block-wise, which is a low-cost grouping approach that reduces hardware burden. Smoothquant[44]scales weight and activation to decrease the difficulty of quantization. Subsequently, AWQ[26]and OWQ[23]also propose scaling transformations on outlier channels of weight to preserve their information representation capacity. GPTQ[16]reduces the group quantization error of LLMs through Hessian-based error compensation[14], achieving commendable quantization performance at 3-bit. OmniQuant[38]introduces a learnable scaling quantizer to reduce quantization errors in an output-aware manner. To enhance the accuracy of LLMs at 3-bit, APTQ[18]allocates different bit-width to different transformer blocks based on Hessian-trace, enhancing the accuracy of LLMs 3-bit.\nTo achieve LLM quantization at ultra-low bit-width, recent novel efforts such as QuIP[6], QuIP#[43], and AQLM[13]promote quantization performance at 2-bit through learnable codebooks or additional fine-tuning. Meanwhile, approaches like SpQR[12], PB-LLM[37], and BiLLM[20]employ finer-grained partitioning for grouped quantization with unstructured mixed-precision for weights, further improving the PTQ performance. However, existing low-bit methods still rely on special structures and fine-grained grouping to ensure accuracy, which increases the difficulty of hardware deployment. Additionally, the extra fine-tuning training may pose a risk of domain-specific overfitting and undermine the efficiency of PTQ[26].\n\n",
    "origin_citations": [
      "Sparks of artificial general intelligence: Early experiments with GPT-4.",
      "A survey on evaluation of large language models.",
      "A survey of large language models.",
      "Language models are few-shot learners.",
      "Llama: Open and efficient foundation language models.",
      "Qa-lora: Quantization-aware low-rank adaptation of large language models.",
      "Compressing large-scale transformer-based models: A case study on bert.",
      "GPTQ: Accurate post-training quantization for generative pre-trained transformers.",
      "Smoothquant: Accurate and efficient post-training quantization for large language models.",
      "Omniquant: Omnidirectionally calibrated quantization for large language models.",
      "Quip: 2-bit quantization of large language models with guarantees.",
      "A survey on model compression for large language models.",
      "SparseGPT: Massive language models can be accurately pruned in one-shot.",
      "BiLLM: Pushing the limit of post-training quantization for llms.",
      "Accurate LoRA-Finetuning Quantization of LLMs via Information Retention.",
      "DB-LLM: Accurate dual-binarization for efficient llms.",
      "LLM-QAT: Data-Free Quantization Aware Training for Large Language Models.",
      "Efficient and affordable post-training quantization for large-scale transformers.",
      "AWQ: Activation-aware weight quantization for LLM compression and acceleration.",
      "OWQ: Lessons learned from activation outliers for weight quantization in large language models.",
      "Optimal brain compression: A framework for accurate post-training quantization and pruning.",
      "APTQ: Attention-aware post-training mixed-precision quantization for large language models.",
      "Quip#: Even better LLM quantization with hadamard incoherence and lattice codebooks.",
      "Extreme compression of large language models via additive quantization.",
      "SpQR: A sparse-quantized representation for near-lossless LLM weight compression.",
      "PB-LLM: Partially binarized large language models."
    ],
    "citations": [
      7638,
      186685,
      8461,
      679,
      1552,
      478590,
      164788,
      945903,
      245034,
      404785,
      468609,
      704619,
      303937,
      100017,
      99199,
      466369,
      257203,
      308051,
      1088952,
      67497,
      465495,
      99754,
      483704,
      306939,
      477369
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 88,
      "Publication_Potential": 92,
      "Final_Score": 89.19999999999999,
      "raw_text": "Content_Quality: 88\nPublication_Potential: 92\nFinal_Score: 89"
    },
    "related_work_scores": {
      "Content_Coherence": 74,
      "Citation_Quality": 88,
      "Synthesis_Analysis": 72,
      "Final_Score": 79.0,
      "raw_text": "Content_Coherence: 74\nCitation_Quality: 88\nSynthesis_Analysis: 72\nFinal_Score: 79"
    },
    "overall_score": 83.08,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2025/Conference",
        "venue": "ICML",
        "year": "2025",
        "openreview_id": "PO9bBEPNWy",
        "forum": "PO9bBEPNWy",
        "number": 9428,
        "title": "SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models",
        "normalized_title": "slim llm salience driven mixed precision quantization for large language models",
        "abstract": "Post-training quantization (PTQ) is an effective technique for compressing large language models (LLMs). However, while uniform-precision quantization is computationally efficient, it often compromises model performance. To address this, we propose SliM-LLM, a salience-driven mixed-precision quantization framework that allocates bit-widths at the group-wise with high accuracy. Our approach leverages the observation that important weights follow a structured distribution and introduces two key components: 1) Salience-Determined Bit Allocation adaptively assigns bit-widths to groups within each layer based on their salience; and 2) Salience-Weighted Quantizer Calibration optimizes quantizer parameters by incorporating element-level salience, retain essential information. With its structured group-wise partitioning, SliM-LLM provides a hardware-friendly solution that matches the efficiency of uniform quantization methods while significantly improving accuracy. Experiments show that SliM-LLM achieves superior performance across various LLMs at low bit-widths. For example, a 2-bit quantized LLaMA-7B model reduces memory usage by nearly 6x compared to the floating-point baseline, decreases perplexity by 48% compared to state-of-the-art gradient-free PTQ methods, and maintains GPU inference speed. Additionally, the extended version, SliM-LLM+, which incorporates gradient-based quantization, further reduces perplexity by 35.1%. Our code is\navailable at https://github.com/Aaronhuang-778/SliM-LLM.",
        "authors": [
          "Wei Huang",
          "Haotong Qin",
          "Yangdong Liu",
          "Yawei Li",
          "Qinshuo Liu",
          "Xianglong Liu",
          "Luca Benini",
          "Michele Magno",
          "Shiming Zhang",
          "XIAOJUAN QI"
        ],
        "decision": null,
        "pdate": 1746105473892,
        "cdate": 1737615020838
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2025/Conference",
          "venue": "ICML",
          "year": "2025",
          "openreview_id": "PO9bBEPNWy",
          "forum": "PO9bBEPNWy",
          "number": 9428,
          "title": "SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models",
          "normalized_title": "slim llm salience driven mixed precision quantization for large language models",
          "abstract": "Post-training quantization (PTQ) is an effective technique for compressing large language models (LLMs). However, while uniform-precision quantization is computationally efficient, it often compromises model performance. To address this, we propose SliM-LLM, a salience-driven mixed-precision quantization framework that allocates bit-widths at the group-wise with high accuracy. Our approach leverages the observation that important weights follow a structured distribution and introduces two key components: 1) Salience-Determined Bit Allocation adaptively assigns bit-widths to groups within each layer based on their salience; and 2) Salience-Weighted Quantizer Calibration optimizes quantizer parameters by incorporating element-level salience, retain essential information. With its structured group-wise partitioning, SliM-LLM provides a hardware-friendly solution that matches the efficiency of uniform quantization methods while significantly improving accuracy. Experiments show that SliM-LLM achieves superior performance across various LLMs at low bit-widths. For example, a 2-bit quantized LLaMA-7B model reduces memory usage by nearly 6x compared to the floating-point baseline, decreases perplexity by 48% compared to state-of-the-art gradient-free PTQ methods, and maintains GPU inference speed. Additionally, the extended version, SliM-LLM+, which incorporates gradient-based quantization, further reduces perplexity by 35.1%. Our code is\navailable at https://github.com/Aaronhuang-778/SliM-LLM.",
          "authors": [
            "Wei Huang",
            "Haotong Qin",
            "Yangdong Liu",
            "Yawei Li",
            "Qinshuo Liu",
            "Xianglong Liu",
            "Luca Benini",
            "Michele Magno",
            "Shiming Zhang",
            "XIAOJUAN QI"
          ],
          "decision": null,
          "pdate": 1746105473892,
          "cdate": 1737615020838
        }
      ]
    }
  },
  {
    "paper_id": 49473,
    "title": "R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts",
    "abstract": "In large multimodal models (LMMs), the perception of non-language modalities (e.g., visual representations) is usually not on par with the large language models (LLMs)’ powerful reasoning capabilities, deterring LMMs’ performance on challenging downstream tasks.\nThis weakness has been recently mitigated by replacing the vision encoder with a mixture-of-experts (MoE), which provides rich, multi-granularity, and diverse representations required by diverse downstream tasks.\nThe performance of multimodal MoE largely depends on its router, which reweights and mixes the representations of different experts for each input.\nHowever, we find that the end-to-end trained router does not always produce the optimal routing weights for every test sample. To bridge the gap, we propose a novel and efficient method “Re-Routing inTest-Time (R2-T2)” that locally optimizes the vector of routing weights in test-time by moving it toward those vectors of the correctly predicted samples in a neighborhood of the test sample. We propose three R2-T2 strategies with different optimization objectives and neighbor-search spaces. R2-T2 consistently and greatly improves state-of-the-art LMMs’ performance on challenging benchmarks of diverse tasks, without training any base-model parameters.",
    "introduction": "Mixture-of-Experts (MoE) have achieved remarkable success in scaling up the size and capacity of large language and multimodal models (LLMs and LMMs)(Shazeer et al.,2017)without (significantly) increasing the inference cost. Specifically, it allows us to increase the total number of experts, which provides finer-grained expertise and skills, yet selecting a constant number of experts for each input(Lepikhin et al.,2020). In MoE, the sparse selection of experts is achieved through a router, which determines the weight of each candidate expert based on the input so only experts with nonzero weights are selected(Fedus et al.,2022). MoE then aggregates the outputs of the selected experts according to their weights. Hence, the router and its produced routing weights play important roles in MoE’s inference cost and output quality.\nAs the most widely studied LMM, many vision language models (VLM) adopt an architecture composed of a vision encoder and an LLM(Zhu et al.,2023), which are both pre-trained and then aligned by further finetuning so the LLM can include the vision encoder’s output in its input as additional tokens. The alignment is usually obtained through a lightweight projection layer or Q-former (a Transformer model) converting the vision encoder’s output to LLM tokens. Despite the broad usage of this architecture, the capability of a vision encoder is usually much more limited than the LLMs (i.e., the “modality imbalance”)(Schrodi et al.,2024). So the visual features cannot cover all the information required by different reasoning tasks performed by LLMs. Moreover, the alignment module may lead to an information bottleneck from the visual perception to the reasoning(Yao et al.,2024).\nRecent advances in LMMs replace a single vision encoder with a mixture of encoders(Lin et al.,2024; Lee et al.,2025; Zong et al.,2024; Shi et al.,2024), which turns out to be an effective and low-cost approach to mitigatemodality imbalance and alignment bottleneck. In multimodal MoE, each expert is an encoder or a mixer of sensory inputs that focuses on a specific type of features, e.g., object classes, text in images, spatial relations, dense captions, segmentation, etc., so the LLM can select the information acquired by any given downstream task from the concatenated or fused features from the MoE, through a router that is trained in an end-to-end manner to produce the weights of all the experts adaptive to the input task.\nAlthough multimodal MoE achieves remarkable success in enhancing the performance of existing LMMs, the choice of experts or the routing weights for individual instances are not always optimal due to the limitations of the router’s design and the diversity of potential downstream tasks compared to the tasks used to train the router. The suboptimality of routing substantially limits the performance and generalization of multimodal MoE on unseen tasks.\nAs illustrated in Figure2, the base model initially selects a sub-optimal expert (e.g.,𝐈langsubscript𝐈lang\\mathbf{I}_{\\textsc{lang}}bold_I start_POSTSUBSCRIPT lang end_POSTSUBSCRIPT) for a spatial reasoning task, leading to incorrect predictions.\nThis has been verified on recent multimodal MoE models. As shown in Table2, compared to the original routing weights of base models, the optimal (oracle) routing weights improve the accuracy by≥10%absentpercent10\\geq 10\\%≥ 10 %on most evaluated LMM benchmarks. To avoid the expensive cost of re-training a router on a much larger dataset, in this paper, we investigatehow to improve the routing weights in test-time without training any model parameters.\nSince routing weights encode the choices of experts with essential knowledge and key skills acquired by the input task, and motivated by the assumption that knowledge and skills are usually transferable across different tasks, we posit that the routing weights of successful tasks can provide critical clues for optimizing the routing weights of a new task.\nSpecifically, we leverage the similarity in a task embedding space, which may reflect the knowledge or skill sharing between tasks, and modify the routing weight vector of a test task by imitating its nearby successful tasks. While the task embedding space, optimization objective, and the number of update steps can vary and their design choices may result in different performances, this innovative mechanism ofoptimizing routing weights or “re-routing” in test-time (R2-T2)focuses on correcting the mistakes made by the routers in existing multimodal MoE, e.g., extracting object detection features for a task mainly depending on the text information in an input image, and thus turns various failed cases into success. Rather than finetuning the whole model, R2-T2 is training-free and aims to maximize the potential of MoE in the reasoning tasks by LMMs.\nFollowing the above idea, we explored several novel strategies for test-time routing weight optimization. They all modify the routing weights of a test task/sample based on a representative set of tasks/samples on which the multimodal MoE achieves correct or high-quality outputs.\nWhile the oracle routing weights are achieved by minimizing the test sample’s loss, for a practical approach, we propose to replace the oracle loss with a surrogate, i.e., a weighted average of losses of nearby reference samples, and apply multiple steps of “neighborhood gradient descent (NGD)” to minimize the surrogate.\nIn addition, we investigate kernel regression and mode finding, which do not require gradient descent. The former moves the routing weights to a kernel-weighted sum of nearby reference tasks’ routing weights in a task embedding space, while the latter moves the routing weights to the nearest mode on the distribution of reference tasks’ routing weights.\nEvaluating these strategies on two recent multimodal MoE models across eight challenging benchmarks, we find that R2-T2 significantly outperforms models twice its size, as shown in Figure1.\nOur analysis reveals that NGD progressively refines routing, increasing correct predictions while mitigating the original router’s over-reliance on a single expert.\nCase studies confirm that test-time re-routing enhances domain-specific reasoning, demonstrating R2-T2’s ability to adapt multimodal MoE models without additional training, unlocking greater generalization and robustness.\nOur main contributions can be summarized below:•We proposed a novel problem of R2-T2 that bridges a significant performance gap on multimodal MoE.•We developed three practical R2-T2 strategies that shed several critical insights into expert re-routing.•Our R2-T2 considerably advances the performance of multimodal MoE on several recent benchmarks of challenging tasks for LMMs.",
    "related_work": "Large Multimodel Modelshas emerged as a powerful paradigm for integrating language and non-language modalities, such as images(Radford et al.,2021), audio(Ao et al.,2021), and video(Zellers et al.,2021), to perform complex reasoning tasks.\nRecent advancements have been driven by the fusion of pretrained LLMs with multimodal encoders(Peng et al.,2023; Tsimpoukelli et al.,2021; Alayrac et al.,2022), enabling the models to process and generate cross-modal content effectively.\nWorks such as Flamingo(Alayrac et al.,2022)and BLIP-2(Li et al.,2023a)demonstrated the potential of aligning vision and language modalities through carefully designed bridging modules.\nHowever, these models often fall short in richness or alignment with the reasoning capabilities of LLMs(Bubeck et al.,2023; Bommasani et al.,2021).\nTo address this, techniques have been proposed, such as contrastive pretraining(Radford et al.,2021; Yuan et al.,2021)and feature fusion mechanisms(Lu et al.,2019).\nYet, efficiently capturing diverse modal interactions across different tasks remains a bottleneck(Baltrušaitis et al.,2018), highlighting the need for more adaptive mechanisms in multimodal reasoning.\nMixture-of-Expertshas become a prominent architectural choice to enhance the scalability and efficiency of large-scale neural networks(Shazeer et al.,2017).\nBy dynamically selecting a subset of specialized expert modules for each input(Li et al.,2023b), MoE reduces computational overhead while maintaining high expressive power(Shazeer et al.,2017; Zoph et al.,2022).\nIn the context of LLMs, MoE has been shown to improve both training efficiency and generalization across tasks(Artetxe & Schwenk,2019).\nWorks such as Switch Transformers(Fedus et al.,2022)and GShard(Lepikhin et al.,2020)have demonstrated the effectiveness of MoE in scaling up model capacity without prohibitive increases in training costs.\nIn multimodal settings, MoE has been explored to address the modality alignment problem(Goyal et al.,2021), where different experts handle distinct modalities or specific tasks.\nHowever, the optimal utilization of experts heavily relies on the effectiveness of routing mechanisms, which remains an active area of research.\nRouters and Routing Strategiesare the cornerstone of any MoE-based architecture, responsible for determining which experts are activated for each input(Li & Zhou,2024).\nTraditional routers, such as softmax gating functions(Shazeer et al.,2017), compute a weighted combination of experts based on input embeddings.\nDespite their simplicity, these routing strategies often face challenges in achieving optimal expert assignment(Lepikhin et al.,2020; Zoph et al.,2022), particularly in unseen or highly diverse test scenarios.\nRecent works have proposed advanced routing strategies, including routing via reinforcement learning(Rosenbaum et al.,2017), early-exit(Li et al.,2023c), and task-specific allocation(Shi et al.,2024).\nHowever, these approaches typically focus on training-time optimization, leaving test-time adaptability largely unexplored.\nR2-T2 introduces an efficient method to refine routing weights dynamically during inference, ensuring better alignment with task-specific requirements and improving overall model robustness across diverse multimodal benchmarks.\nTest-Time Optimizationhas been explored by adapting models dynamically during inference to improve generalization.\nFor example,(Wang et al.,2022)propose test-time adaptation, which fine-tunes model parameters on test data distributions using entropy minimization or self-supervised learning.\nSimilarly,(Sun et al.,2020)introduce test-time training, where models are updated via auxiliary tasks (e.g., rotation prediction) during inference.\nHowever, these methods require modifying the base model’s parameters, leading to significant computational overhead and potential instability when deployed on resource-constrained systems.\nUnlike prior test-time optimization methods that update model weights, R2-T2 solely optimizes the routing weights of a frozen MoE model without retraining any model parameters.",
    "origin_citations": [
      "Learning transferable visual models from natural language supervision.",
      "Speecht5: Unified-modal encoder-decoder pre-training for spoken language processing.",
      "Merlot: Multimodal neural script knowledge models.",
      "Kosmos-2: Grounding multimodal large language models to the world.",
      "Multimodal few-shot learning with frozen language models.",
      "Flamingo: a visual language model for few-shot learning.",
      "Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.",
      "Sparks of artificial general intelligence: Early experiments with gpt-4.",
      "On the opportunities and risks of foundation models.",
      "Multimodal contrastive training for visual representation learning.",
      "Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks.",
      "Multimodal machine learning: A survey and taxonomy.",
      "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.",
      "Simple: Specialized model-sample matching for domain generalization.",
      "Designing effective sparse expert models.",
      "Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond.",
      "Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.",
      "Gshard: Scaling giant models with conditional computation and automatic sharding.",
      "Coordination among neural modules through a shared global workspace.",
      "Your mixture-of-experts llm is secretly an embedding model for free.",
      "Routing networks: Adaptive selection of non-linear functions for multi-task learning.",
      "Towards inference efficient deep ensemble learning.",
      "Eagle: Exploring the design space for multimodal llms with mixture of encoders.",
      "Continual test-time domain adaptation.",
      "Test-time training with self-supervision for generalization under distribution shifts, 2020."
    ],
    "citations": [
      1639,
      340744,
      5238,
      3019,
      3003,
      7565,
      212324,
      7638,
      1550,
      163747,
      1278,
      1030,
      680,
      1087856,
      6555,
      707,
      8454,
      218469,
      1006995,
      786,
      212440,
      782239,
      500764,
      5092
    ],
    "extract_rate": 0.96,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 88,
      "Final_Score": 84.5,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 88\nFinal_Score: 85"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 83,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.1,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 83\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 83.06,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2025/Conference",
        "venue": "ICML",
        "year": "2025",
        "openreview_id": "oqPcOMafOF",
        "forum": "oqPcOMafOF",
        "number": 411,
        "title": "R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts",
        "normalized_title": "r2 t2 re routing in test time for multimodal mixture of experts",
        "abstract": "In large multimodal models (LMMs), the perception of non-language modalities (e.g., visual representations) is usually not on par with the large language models (LLMs)' powerful reasoning capabilities, deterring LMMs' performance on challenging downstream tasks. \nThis weakness has been recently mitigated by replacing the vision encoder with a mixture-of-experts (MoE), which provides rich, multi-granularity, and diverse representations required by different downstream tasks. The performance of multimodal MoE largely depends on its router, which reweights and mixes the representations of different experts for each input. However, we find that the end-to-end trained router does not always produce the optimal routing weights for every test sample. To bridge the gap, we propose a novel and efficient method ''**R**e-**R**outing in **T**est-**T**ime (R2-T2)'' that locally optimizes the vector of routing weights in test-time by moving it toward those vectors of the correctly predicted samples in a neighborhood of the test sample. We propose three R2-T2 strategies with different optimization objectives and neighbor-search spaces. R2-T2 consistently and significantly improves state-of-the-art LMMs' performance on challenging multimodal benchmarks of diverse tasks, without training any parameters in the base model. Our code can be accessed here.",
        "authors": [
          "Zhongyang Li",
          "Ziyue Li",
          "Tianyi Zhou"
        ],
        "decision": null,
        "pdate": 1746105088104,
        "cdate": 1736462326697
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2025/Conference",
          "venue": "ICML",
          "year": "2025",
          "openreview_id": "oqPcOMafOF",
          "forum": "oqPcOMafOF",
          "number": 411,
          "title": "R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts",
          "normalized_title": "r2 t2 re routing in test time for multimodal mixture of experts",
          "abstract": "In large multimodal models (LMMs), the perception of non-language modalities (e.g., visual representations) is usually not on par with the large language models (LLMs)' powerful reasoning capabilities, deterring LMMs' performance on challenging downstream tasks. \nThis weakness has been recently mitigated by replacing the vision encoder with a mixture-of-experts (MoE), which provides rich, multi-granularity, and diverse representations required by different downstream tasks. The performance of multimodal MoE largely depends on its router, which reweights and mixes the representations of different experts for each input. However, we find that the end-to-end trained router does not always produce the optimal routing weights for every test sample. To bridge the gap, we propose a novel and efficient method ''**R**e-**R**outing in **T**est-**T**ime (R2-T2)'' that locally optimizes the vector of routing weights in test-time by moving it toward those vectors of the correctly predicted samples in a neighborhood of the test sample. We propose three R2-T2 strategies with different optimization objectives and neighbor-search spaces. R2-T2 consistently and significantly improves state-of-the-art LMMs' performance on challenging multimodal benchmarks of diverse tasks, without training any parameters in the base model. Our code can be accessed here.",
          "authors": [
            "Zhongyang Li",
            "Ziyue Li",
            "Tianyi Zhou"
          ],
          "decision": null,
          "pdate": 1746105088104,
          "cdate": 1736462326697
        }
      ]
    }
  },
  {
    "paper_id": 6331,
    "title": "The Shaped Transformer:Attention Models in the Infinite\nDepth-and-Width Limit",
    "abstract": "In deep learning theory, the covariance matrix of the representations serves as a proxy to examine the network’s trainability.\nMotivated by the success of Transformers, we study the covariance matrix of a modified Softmax-based attention model with skip connections in the proportional limit of infinite-depth-and-width.\nWe show that at initialization the limiting distribution can be described by a stochastic differential equation (SDE) indexed by the depth-to-width ratio.\nTo achieve a well-defined stochastic limit, the Transformer’s attention mechanism is modified by centering the Softmax output at identity, and scaling the Softmax logits by a width-dependent temperature parameter.\nWe examine the stability of the network through the corresponding SDE, showing how the scale of both the drift and diffusion can be elegantly controlled with the aid of residual connections.\nThe existence of a stable SDE implies that the covariance structure is well-behaved, even for very large depth and width, thus preventing the notorious issues of rank degeneracy in deep attention models.\nFinally, we show, through simulations, that the SDE provides a surprisingly good description of the corresponding finite-size model. We coin the nameshaped Transformerfor these architectural modifications.",
    "introduction": "Pre-trained large language models have experienced a remarkable increase in popularity due to their eerily human-like ability to puzzle through complex reasoning tasks, solve coding challenges, and produce pages of logically sound text[1].\nArguably, the Transformer is the foundation of these successes[2].\nRecent research has found evidence for scaling laws, linking the performance of these architectures to their parameter counts and the quantity of training data, fueling the desire to train deeper and wider models on ever larger datasets\nin order to unlock new levels of performance[3,4,5,6,7].\n\n\nBundled with the increased expressivity of deep architectures, however, is increased numerical instability, both in the forward pass and gradients, which hinders training.\nOne of the clearest examples of instability is the so-called rank collapse phenomenon[8,9]– the observation that,\nin Softmax-based attention models,\nthe network’s representation of different tokens tend to perfectly align at large depth.\nThe resulting poorly conditioned covariance and correlation between tokens leads to exploding and/or vanishing gradients at initialization, disrupting gradient updates of the affected parameters.\nThis situation violates a well-known guiding principle\nfrom the literature of deep signal propagation:\na stable covariance is a necessary condition for stable training[10,11,12,13,14,15].\nIn fact, the instability of Transformers is evident when considering the critical role of hyperparameter tuning and the judicious use of normalization layers.\nIn this work, we study Transformers in a novel infinite limit, rectify sources of instability with a novel modification, and derive the SDEs characterizing the covariance and output distribution.\n\n\nFigure 1:Our shaped Transformer prevents token representations from becoming perfectly aligned, i.e. rank collapse.\nLeft: mean correlationρℓα⁢βsubscriptsuperscript𝜌𝛼𝛽ℓ\\rho^{\\alpha\\beta}_{\\ell}italic_ρ start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPTof Transformers (Equation11) with and without shaped attention (Equation9) and Pre-LN[16].\nRight: kernel density estimate and histogram of correlations from covariance SDE inTheorem4.2and shaped attention NN.\nHere we note correlation converging to1111implies a poorly conditioned covariance matrix.\nSimulated withn=200,d=150,γ=1/8,τ0=1,ρ0α⁢β=0.2formulae-sequence𝑛200formulae-sequence𝑑150formulae-sequence𝛾18formulae-sequencesubscript𝜏01subscriptsuperscript𝜌𝛼𝛽00.2n=200,d=150,\\gamma=1/\\sqrt{8},\\tau_{0}=1,\\rho^{\\alpha\\beta}_{0}=0.2italic_n = 200 , italic_d = 150 , italic_γ = 1 / square-root start_ARG 8 end_ARG , italic_τ start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 1 , italic_ρ start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT 0 end_POSTSUBSCRIPT = 0.2, SDE step size0.010.010.010.01, and212superscript2122^{12}2 start_POSTSUPERSCRIPT 12 end_POSTSUPERSCRIPTsamples.\n\n\nScaling limits have been used successfully to provide guidance on architecture[17,18,19]and tuning hyperparameters settings[20]. Our work represents a contribution in this direction.\nThe ability to use such limits to diagnose instabilities depends on their tractability and faithfulness to real-world (finite) networks.\nIn this regard, not all limits are created equal.\nIn particular, the faithfulness of scaling limits depends critically on how other parameters are scaled with width. One of the simplest (and thus most popular) limits to work with – the “NTK” limit[21,22,23,24,25]– treats the depth of the network as fixed.\nAs a result, at initialization, this limit does not accumulate sufficient random fluctuations over the depth of the network, leading to deterministic covariance matrices that do not agree with those of standard (finite) networks.\nSuch networks have another defect: they are incapable of learning features in the limit[26].\nVarious other limits have been studied, towards identifying tractable yet faithful models of initialization and/or training.\nThese include mean field limits[27,28,29,30]and\nthe perturbative regime[31,32,33,34,35,36,37].\n\n\nThis work operates in a relatively new regime – theproportionalinfinite depth-and-width limit –\nwhere depthd𝑑ditalic_dand widthn𝑛nitalic_ndiverge as the ratiod/n𝑑𝑛d/nitalic_d / italic_ntends to a positive constant.\nThis limit, first analyzed by[38,  ], has been the recent subject of study in the context of neural network[39,40,41,42,19].\nA related line of work also studied the Lyapunov exponent for products of random matrices[43,44,45,46].\nThis regime retains the network’s stochasticity and, at initialization, has been shown to closely resemble the behaviour of finite architectures, yet still yield a relatively simple limiting description, expressible in terms of stochastic differential equations[41,19].\nIn this work, we fully characterize the initial output distributions of a network with skip connections and Softmax-based attention mechanisms, in the proportional infinite-depth-and-width limit.\n\n\nInspired by the idea of shaping activation functions[17,18,19,47], our theoretical approach finds an adequately modified attention mechanism via its SDE limit.\nOur modification involves making the attention matrix closer to the identity, and appropriately choosing the temperature parameterτ𝜏\\tauitalic_τ, which re-scales the logits of the Softmax.\nSimilar to shaping activation functions, the temperature scaling we devise linearizes and reduces the saturation of the Softmax, a known source of training instability in Transformers[48].\nIn order to model the feedforward layer of a Transformer’s block,\nwe extend existing results[19]to derive an SDE for the proportional limit of shaped-ReLU feedforward multi-layer perceptrons (MLPs) with skip connections. Combined, we fully characterize the output distribution of a Transformer with shaped non-linearities (Corollary4.3).\n\n\n(a)Value WeightsWℓVsubscriptsuperscript𝑊𝑉ℓW^{V}_{\\ell}italic_W start_POSTSUPERSCRIPT italic_V end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT(b)Query WeightsWℓQsubscriptsuperscript𝑊𝑄ℓW^{Q}_{\\ell}italic_W start_POSTSUPERSCRIPT italic_Q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPTFigure 2:Comparing gradients norms at initialization for different parameters as a function of depth, with and without shaped attention. The architecture is the same as inFigure1but with autoregressive causal masking, and the task is next-token prediction on code data. Left: Value weightsWℓVsubscriptsuperscript𝑊𝑉ℓW^{V}_{\\ell}italic_W start_POSTSUPERSCRIPT italic_V end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPTfor shaped attention, standard Pre-LN, and the original Post-LN block[2]. Right: the same gradient norm plot but for Query weightsWlQsubscriptsuperscript𝑊𝑄𝑙W^{Q}_{l}italic_W start_POSTSUPERSCRIPT italic_Q end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_l end_POSTSUBSCRIPT. We find that shaping the attention mechanism successfully prevents gradients from vanishing, while unshaped Transformers suffer from rapidly vanishing gradients. Interestingly, only the Post-LN query gradients vanish, but value gradients are stable across depths, which is consistent with the findings of[9,  ]. On the other hand, shaped attention has stable gradients for both parameters inside and outside the Softmax nonlinearity.\n\n\nNotably, our modification successfully prevents a poorly conditioned covariance matrix, whereas the vanilla Softmax-based attention model without LayerNorm[49]fails in this regard, and the corresponding Pre-LN architecture provides only marginal improvements (seeFigure1).\nGiven that our modification is inspired by previous work on shaping activation functions, we coin the termsshaped attentionfor the proposed attention mechanism andshaped Transformerfor the overall architecture that includes the MLP block and residual connections.\nThrough simulations (e.g.,Figure1), we show that the limiting neural covariance SDE approximates the distribution of finite-size Transformers with shaped attention mechanism surprisingly well.\nWe also provide preliminary training experiments for our proposed shaped attention architecture on standard language modeling tasks, demonstrating the feasibility of the new architecture in practice (seeSection5andAppendixD).\n\n\nIn summary, our contributions are as follows:1.We study the effect of skip connections in the\nproportional limit, showing that under a precise relation between the scaling parameters of the shortcut and residual branches, the feature covariance converges to the solution of a weighted version of the neural covariance SDE for MLPs (Theorem3.2).\nThe dependence on the depth-to-width ratio implies the existence of a stable non-commutative limit for residual networks, complementing the commutative limit studied in[50,  ].2.We proposeshaped attention, where we modify the Softmax-based attention mechanism to be a perturbation of the identity.\nWe demonstrate that shaped attention successfully prevents the degeneracy of correlation in contrast to existing Transformer architectures (Figure1). The enhanced stability in the forward pass is reflected in the gradients, which are also stable with depth, as we empirically show inFigure2.3.For the proposed shaped attention architecture, we derive the neural covariance SDE characterizing the initial distribution in the proportional limit\n(Theorem4.2).\nConsequently, we provide the first characterization of Transformer-type architectures, i.e. the shaped Transformer, in the large depth-and-width regime (Corollary4.3).4.We provide simulations to validate the theory and to interpret the effects of network hyperparamaters on the covariance matrix of the shaped Transformer.\nSpecifically, we study finite time stability of the SDE and provide explicit guidance on hyperparameters to prevent numerical instability.The paper is organized as follows:\nInSection2, we provide the basic setup and some background on existing results.\nInSection3, we generalize the SDE results of[19,  ]to include skip connections. This serves as a model to understand the effect of skip connections in isolation from the attention model.\nInSection4,\nwe present our main result, first pinpointing the origins of instability in the Softmax, then showing how the modifications underlyingshaped attentionallow us to derive a non-trivial SDE limit.\nFinally, inSection5, we discuss the implications of our results and some future directions.\nProofs of all theorems and additional experiments are deferred to the Appendix.\n\n",
    "related_work": "Setup.LetXℓ∈ℝm×nsubscript𝑋ℓsuperscriptℝ𝑚𝑛X_{\\ell}\\in\\mathbb{R}^{m\\times n}italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_m × italic_n end_POSTSUPERSCRIPTbe the data matrix representing a sequence ofm𝑚mitalic_mtokens embedded inn𝑛nitalic_ndimensions at layerℓ∈[d]ℓdelimited-[]𝑑\\ell\\in[d]roman_ℓ ∈ [ italic_d ], whered𝑑ditalic_dis the depth of the network. We elide the explicit dependence onℓℓ\\ellroman_ℓwhen it is clear from the context, and use superscript Greek letters to indicate specific tokens’ representations, for instancexℓα∈ℝnsubscriptsuperscript𝑥𝛼ℓsuperscriptℝ𝑛x^{\\alpha}_{\\ell}\\in\\mathbb{R}^{n}italic_x start_POSTSUPERSCRIPT italic_α end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_n end_POSTSUPERSCRIPTis theα𝛼\\alphaitalic_α-th row ofXℓsubscript𝑋ℓX_{\\ell}italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT. We consider the following attention model with residual connections:Xℓ+1=λ⁢Xℓ+γ⁢Aℓ⁢Xℓ⁢1n⁢WℓVsubscript𝑋ℓ1𝜆subscript𝑋ℓ𝛾subscript𝐴ℓsubscript𝑋ℓ1𝑛subscriptsuperscript𝑊𝑉ℓX_{\\ell+1}=\\lambda X_{\\ell}+\\gamma A_{\\ell}X_{\\ell}\\ \\frac{1}{\\sqrt{n}}W^{V}_{%\n\\ell}\\,italic_X start_POSTSUBSCRIPT roman_ℓ + 1 end_POSTSUBSCRIPT = italic_λ italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT + italic_γ italic_A start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT divide start_ARG 1 end_ARG start_ARG square-root start_ARG italic_n end_ARG end_ARG italic_W start_POSTSUPERSCRIPT italic_V end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT(1)whereγ,λ∈[0,1]𝛾𝜆01\\gamma,\\lambda\\in[0,1]italic_γ , italic_λ ∈ [ 0 , 1 ]are parameters that control the strength of the shortcut and residual branch, respectively,WℓV∈ℝn×nsubscriptsuperscript𝑊𝑉ℓsuperscriptℝ𝑛𝑛W^{V}_{\\ell}\\in\\mathbb{R}^{n\\times n}italic_W start_POSTSUPERSCRIPT italic_V end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_n × italic_n end_POSTSUPERSCRIPTis the weight matrix of the values, andAℓ∈ℝm×msubscript𝐴ℓsuperscriptℝ𝑚𝑚A_{\\ell}\\in\\mathbb{R}^{m\\times m}italic_A start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_m × italic_m end_POSTSUPERSCRIPTis the attention matrix.\nWe consider Softmax-based scaled dot-product attention, whereAℓsubscript𝐴ℓA_{\\ell}italic_A start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPThas the form:Aℓ=Softmax⁢(1τ⁢Xℓ⁢1n⁢WℓQ⁢1n⁢WℓK,⊤⁢Xℓ⊤),subscript𝐴ℓSoftmax1𝜏subscript𝑋ℓ1𝑛superscriptsubscript𝑊ℓ𝑄1𝑛superscriptsubscript𝑊ℓ𝐾topsuperscriptsubscript𝑋ℓtopA_{\\ell}=\\text{Softmax}\\left(\\frac{1}{\\tau}X_{\\ell}\\ \\frac{1}{\\sqrt{n}}W_{\\ell%\n}^{Q}\\ \\frac{1}{\\sqrt{n}}W_{\\ell}^{K,\\top}\\ X_{\\ell}^{\\top}\\right),italic_A start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT = Softmax ( divide start_ARG 1 end_ARG start_ARG italic_τ end_ARG italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT divide start_ARG 1 end_ARG start_ARG square-root start_ARG italic_n end_ARG end_ARG italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_Q end_POSTSUPERSCRIPT divide start_ARG 1 end_ARG start_ARG square-root start_ARG italic_n end_ARG end_ARG italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_K , ⊤ end_POSTSUPERSCRIPT italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT ⊤ end_POSTSUPERSCRIPT ) ,(2)where the Softmax is applied row-wise,WℓQ,WℓK∈ℝn×nksuperscriptsubscript𝑊ℓ𝑄superscriptsubscript𝑊ℓ𝐾superscriptℝ𝑛subscript𝑛𝑘W_{\\ell}^{Q},W_{\\ell}^{K}\\in\\mathbb{R}^{n\\times n_{k}}italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_Q end_POSTSUPERSCRIPT , italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_K end_POSTSUPERSCRIPT ∈ blackboard_R start_POSTSUPERSCRIPT italic_n × italic_n start_POSTSUBSCRIPT italic_k end_POSTSUBSCRIPT end_POSTSUPERSCRIPTare additional random weights, andτ𝜏\\tauitalic_τis a temperature parameter, which controls the entropy of the distribution.\nHere we let all the weight matricesWℓQ,WℓK,WℓVsuperscriptsubscript𝑊ℓ𝑄superscriptsubscript𝑊ℓ𝐾subscriptsuperscript𝑊𝑉ℓW_{\\ell}^{Q},W_{\\ell}^{K},W^{V}_{\\ell}italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_Q end_POSTSUPERSCRIPT , italic_W start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT start_POSTSUPERSCRIPT italic_K end_POSTSUPERSCRIPT , italic_W start_POSTSUPERSCRIPT italic_V end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPThave𝒩⁢(0,1)𝒩01\\mathcal{N}(0,1)caligraphic_N ( 0 , 1 )-iid entries.\nIn the case whereλ,γ=1𝜆𝛾1\\lambda,\\gamma=1italic_λ , italic_γ = 1, with the application of LayerNorm on the residual branch[16], and withτ=nk𝜏subscript𝑛𝑘\\tau=\\sqrt{n_{k}}italic_τ = square-root start_ARG italic_n start_POSTSUBSCRIPT italic_k end_POSTSUBSCRIPT end_ARG, we recover the attention block of the vanilla \"Pre-LN\" Transformer architecture[2].\nHere we note that we pull the conventionaln−1/2superscript𝑛12n^{-1/2}italic_n start_POSTSUPERSCRIPT - 1 / 2 end_POSTSUPERSCRIPTfactor outside of the weight matrices, which preserves the forward pass, and yields equivalent training dynamics up to a reparameterization of the learning rate[26].\nIn this work, we consider unnormalized architectures, and control the variance propagation with the conditionλ2+γ2=1superscript𝜆2superscript𝛾21\\lambda^{2}+\\gamma^{2}=1italic_λ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT + italic_γ start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT = 1[41].\nWe are interested in studying the so-calledneural covariancefor the attention model (Equation1) in the proportional limit.\n\n\nNeural Covariance.In deep learning theory, researchers have long sought to understand how networks internally represent different inputs and how different architectural choices affect these representations. The approach followed by work on signal propagation has been to study how the relative alignment of different inputs evolves across the network, as measured by the neural covarianceVℓα⁢β:=1n⁢⟨xℓα,xℓβ⟩assignsubscriptsuperscript𝑉𝛼𝛽ℓ1𝑛subscriptsuperscript𝑥𝛼ℓsubscriptsuperscript𝑥𝛽ℓV^{\\alpha\\beta}_{\\ell}:=\\frac{1}{n}\\langle x^{\\alpha}_{\\ell},x^{\\beta}_{\\ell}\\rangleitalic_V start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT := divide start_ARG 1 end_ARG start_ARG italic_n end_ARG ⟨ italic_x start_POSTSUPERSCRIPT italic_α end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT , italic_x start_POSTSUPERSCRIPT italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ⟩(orρα⁢β:=(Vℓα⁢α⁢Vℓβ⁢β)−1/2⁢Vℓα⁢βassignsuperscript𝜌𝛼𝛽superscriptsubscriptsuperscript𝑉𝛼𝛼ℓsubscriptsuperscript𝑉𝛽𝛽ℓ12subscriptsuperscript𝑉𝛼𝛽ℓ\\rho^{\\alpha\\beta}:=(V^{\\alpha\\alpha}_{\\ell}V^{\\beta\\beta}_{\\ell})^{-1/2}V^{%\n\\alpha\\beta}_{\\ell}italic_ρ start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT := ( italic_V start_POSTSUPERSCRIPT italic_α italic_α end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT italic_V start_POSTSUPERSCRIPT italic_β italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ) start_POSTSUPERSCRIPT - 1 / 2 end_POSTSUPERSCRIPT italic_V start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPTif interested only in the correlation).\nAt initialization, characterizations of this covariance structure have been exploited to infer important properties of neural networks[10,11].\nAs an example, in the sequential infinite-width-then-depth limit, the correlationρdα⁢βsubscriptsuperscript𝜌𝛼𝛽𝑑\\rho^{\\alpha\\beta}_{d}italic_ρ start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPTof MLPs is known to converge to a fixed point independent of the input[11,17].\nIn this regime, the model is not able to discriminate different data points, which severely hinders training, as the gradient step for the deep layers is taken in the same direction regardless of the input.\nIn the context of Softmax-based attention models,[8,  ]proved that the feature matrixXℓsubscript𝑋ℓX_{\\ell}italic_X start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPTloses rank doubly exponentially fast with depth, and[9,  ]showed how this leads to vanishing gradients of the queries and keys parameters, thus further highlighting how the stability of forward and backward passes are deeply entangled (see alsoFigure2).\n\n\nStabilizing the Effect of Non-Linear Layers.Central to the issue of degeneracy of the neural covariance are commonly used non-linear activation functions that severely deviate from the identity. The recent line of work of Deep Kernel Shaping (DKS)[17,18,19]addresses the issue by considering the cumulative amount of non-linearity throughout layers, andshapingthe activation function by making it closer to the identity map.\nInspired by this line of work,[47,  ]devise an initialization for Transformers that avoid the rank collapse problem without the aid of skip connections or LayerNorm.In an alternative approach, the line of work behind Stable ResNets[50,51,52,53]considers scaling the residual branches byγ=1/depth𝛾1depth\\gamma=1/\\sqrt{\\text{depth}}italic_γ = 1 / square-root start_ARG depth end_ARG, and postulates this scaling is sufficient to stabilize the neural covariance with minimal assumptions on the activation function.[9,  ]adopts this scaling to give precise formulas on the expected covariance of a Transformer at initialization.\nIn this work, we considerγ𝛾\\gammaitalic_γconstant in width and depth, and derive a complementary limiting result.\n\n\nThe Proportional Infinite-Depth-and-Width Limit.In the context of feed-forward MLPs, the output distribution with respect to a single input was studied in[38,42], where it was shown that for the ReLU nonlinearity, the norm of the activationsVα⁢αsuperscript𝑉𝛼𝛼V^{\\alpha\\alpha}italic_V start_POSTSUPERSCRIPT italic_α italic_α end_POSTSUPERSCRIPTconverges to a log-normal random variable.\nTo resolve the degeneracy of covariance and provide a characterization of output distributions formultiple inputs,[19,  ]shapes the ReLU by setting its slope1/width1width1/\\sqrt{\\text{width}}1 / square-root start_ARG width end_ARG-away from linearity.\nIn the proportional limit, the effect of the non-linearity accumulates over thed𝑑ditalic_dlayers, and the covariance matrixVℓ=[Vℓα⁢β]α⁢βsubscript𝑉ℓsubscriptdelimited-[]subscriptsuperscript𝑉𝛼𝛽ℓ𝛼𝛽V_{\\ell}=[V^{\\alpha\\beta}_{\\ell}]_{\\alpha\\beta}italic_V start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT = [ italic_V start_POSTSUPERSCRIPT italic_α italic_β end_POSTSUPERSCRIPT start_POSTSUBSCRIPT roman_ℓ end_POSTSUBSCRIPT ] start_POSTSUBSCRIPT italic_α italic_β end_POSTSUBSCRIPTconverges weakly to the solution of the SDEd⁢Vt=bReLU⁢(Vt)⁢d⁢t+Σlin1/2⁢(Vt)⁢d⁢Bt,𝑑subscript𝑉𝑡subscript𝑏ReLUsubscript𝑉𝑡𝑑𝑡superscriptsubscriptΣlin12subscript𝑉𝑡𝑑subscript𝐵𝑡dV_{t}=b_{\\text{ReLU}}(V_{t})\\,dt+\\Sigma_{\\text{lin}}^{1/2}(V_{t})\\,dB_{t}\\,,italic_d italic_V start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT = italic_b start_POSTSUBSCRIPT ReLU end_POSTSUBSCRIPT ( italic_V start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) italic_d italic_t + roman_Σ start_POSTSUBSCRIPT lin end_POSTSUBSCRIPT start_POSTSUPERSCRIPT 1 / 2 end_POSTSUPERSCRIPT ( italic_V start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ) italic_d italic_B start_POSTSUBSCRIPT italic_t end_POSTSUBSCRIPT ,(3)where the formulae for coefficientsbReLU,Σlinsubscript𝑏ReLUsubscriptΣlinb_{\\text{ReLU}},\\Sigma_{\\text{lin}}italic_b start_POSTSUBSCRIPT ReLU end_POSTSUBSCRIPT , roman_Σ start_POSTSUBSCRIPT lin end_POSTSUBSCRIPTcan be found inTheorem3.2.We note that the output neuron distributions are directly recovered as a conditional Gaussian with covarianceVTsubscript𝑉𝑇V_{T}italic_V start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPTforT=dn𝑇𝑑𝑛T=\\frac{d}{n}italic_T = divide start_ARG italic_d end_ARG start_ARG italic_n end_ARG, in a similar spirit as the neural network Gaussian process (NNGP) results[21,22,23].\nFor example, thei𝑖iitalic_i-th outputXout,isubscript𝑋out𝑖X_{\\text{out},i}italic_X start_POSTSUBSCRIPT out , italic_i end_POSTSUBSCRIPTconditioned onVdsubscript𝑉𝑑V_{d}italic_V start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPTare asymptotically iid.𝒩⁢(0,VT)𝒩0subscript𝑉𝑇\\mathcal{N}(0,V_{T})caligraphic_N ( 0 , italic_V start_POSTSUBSCRIPT italic_T end_POSTSUBSCRIPT )asd,n→∞→𝑑𝑛d,n\\to\\inftyitalic_d , italic_n → ∞.\nThe reader is referred toAppendixAfor more technical background on the covariance SDE and the convergence result.While the existing results are limited to initialization, we remind the reader that this is a necessary step before we can study training dynamics.\nIn particular, the NNGP techniques developed for infinite-width networks at initialization were directly used to study the training dynamics in the same limit[25,54].\nWe will provide further discussions on this topic inSection5.\n\n",
    "origin_citations": [
      "“On layer normalization in the transformer architecture”",
      "“Attention is all you need”, 2017",
      "“Feature learning in infinite-width neural networks”, 2020",
      "“The future is log-Gaussian: ResNets and their infinite-depth-and-width limit at initialization”",
      "“Exponential expressivity in deep neural networks through transient chaos”",
      "“Deep Information Propagation”",
      "“Rapid training of deep neural networks without skip connections or normalization layers using deep kernel shaping”, 2021",
      "“Attention is not all you need: Pure attention loses rank doubly exponentially with depth”",
      "“Signal Propagation in Transformers: Theoretical Perspectives and the Role of Rank Collapse”, 2022",
      "“Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers”, 2022",
      "“The neural covariance SDE: Shaped infinite depth-and-width networks at initialization”",
      "“Deep Transformers without Shortcuts: Modifying Self-attention for Faithful Signal Propagation”, 2023",
      "“Width and Depth Limits Commute in Residual Networks”, 2023",
      "“Stable resnet”",
      "“Dynamical isometry is achieved in residual networks in a universal way for any activation function”",
      "“On the infinite-depth limit of finite-width neural networks”, 2022",
      "“Products of many large random matrices and gradients in deep neural networks”",
      "“Precise characterization of the prior predictive distribution of deep ReLU networks”",
      "“Bayesian learning for neural networks”, 1995",
      "“Deep Neural Networks as Gaussian Processes”",
      "“Gaussian process behaviour in wide deep neural networks”, 2018",
      "“Neural tangent kernel: Convergence and generalization in neural networks”",
      "“Tensor programs ii: Neural tangent kernel for any architecture”"
    ],
    "citations": [
      1511,
      38,
      918110,
      194979,
      31781,
      62899,
      343283,
      217547,
      412843,
      706410,
      413010,
      227151,
      491955,
      435881,
      390932,
      461784,
      1088322,
      193777,
      528106,
      986882,
      3963,
      167759
    ],
    "extract_rate": 0.9565217391304348,
    "article_scores": {
      "Content_Quality": 85,
      "Publication_Potential": 90,
      "Final_Score": 86.5,
      "raw_text": "Content_Quality: 85\nPublication_Potential: 90\nFinal_Score: 86.5"
    },
    "related_work_scores": {
      "Content_Coherence": 78,
      "Citation_Quality": 87,
      "Synthesis_Analysis": 75,
      "Final_Score": 80.7,
      "raw_text": "Content_Coherence: 78\nCitation_Quality: 87\nSynthesis_Analysis: 75\nFinal_Score: 81"
    },
    "overall_score": 83.02000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2023/Conference",
        "venue": "NeurIPS",
        "year": "2023",
        "openreview_id": "PqfPjS9JRX",
        "forum": "PqfPjS9JRX",
        "number": 11329,
        "title": "The Shaped Transformer: Attention Models in the Infinite Depth-and-Width Limit",
        "normalized_title": "the shaped transformer attention models in the infinite depth and width limit",
        "abstract": "In deep learning theory, the covariance matrix of the representations serves as a\nproxy to examine the network’s trainability. Motivated by the success of Transform-\ners, we study the covariance matrix of a modified Softmax-based attention model\nwith skip connections in the proportional limit of infinite-depth-and-width. We\nshow that at initialization the limiting distribution can be described by a stochastic\ndifferential equation (SDE) indexed by the depth-to-width ratio. To achieve a\nwell-defined stochastic limit, the Transformer’s attention mechanism is modified\nby centering the Softmax output at identity, and scaling the Softmax logits by a\nwidth-dependent temperature parameter. We examine the stability of the network\nthrough the corresponding SDE, showing how the scale of both the drift and diffu-\nsion can be elegantly controlled with the aid of residual connections. The existence\nof a stable SDE implies that the covariance structure is well-behaved, even for very\nlarge depth and width, thus preventing the notorious issues of rank degeneracy\nin deep attention models. Finally, we show, through simulations, that the SDE\nprovides a surprisingly good description of the corresponding finite-size model.\nWe coin the name shaped Transformer for these architectural modifications.",
        "authors": [
          "Lorenzo Noci",
          "Chuning Li",
          "Mufan Bill Li",
          "Bobby He",
          "Thomas Hofmann",
          "Chris J. Maddison",
          "Daniel M. Roy"
        ],
        "decision": null,
        "pdate": 1695326072857,
        "cdate": 1683795502959
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2023/Conference",
          "venue": "NeurIPS",
          "year": "2023",
          "openreview_id": "PqfPjS9JRX",
          "forum": "PqfPjS9JRX",
          "number": 11329,
          "title": "The Shaped Transformer: Attention Models in the Infinite Depth-and-Width Limit",
          "normalized_title": "the shaped transformer attention models in the infinite depth and width limit",
          "abstract": "In deep learning theory, the covariance matrix of the representations serves as a\nproxy to examine the network’s trainability. Motivated by the success of Transform-\ners, we study the covariance matrix of a modified Softmax-based attention model\nwith skip connections in the proportional limit of infinite-depth-and-width. We\nshow that at initialization the limiting distribution can be described by a stochastic\ndifferential equation (SDE) indexed by the depth-to-width ratio. To achieve a\nwell-defined stochastic limit, the Transformer’s attention mechanism is modified\nby centering the Softmax output at identity, and scaling the Softmax logits by a\nwidth-dependent temperature parameter. We examine the stability of the network\nthrough the corresponding SDE, showing how the scale of both the drift and diffu-\nsion can be elegantly controlled with the aid of residual connections. The existence\nof a stable SDE implies that the covariance structure is well-behaved, even for very\nlarge depth and width, thus preventing the notorious issues of rank degeneracy\nin deep attention models. Finally, we show, through simulations, that the SDE\nprovides a surprisingly good description of the corresponding finite-size model.\nWe coin the name shaped Transformer for these architectural modifications.",
          "authors": [
            "Lorenzo Noci",
            "Chuning Li",
            "Mufan Bill Li",
            "Bobby He",
            "Thomas Hofmann",
            "Chris J. Maddison",
            "Daniel M. Roy"
          ],
          "decision": null,
          "pdate": 1695326072857,
          "cdate": 1683795502959
        }
      ]
    }
  },
  {
    "paper_id": 16454,
    "title": "Selecting Large Language Model to Fine-tune via Rectified Scaling Law",
    "abstract": "The ever-growing ecosystem of LLMs has posed a challenge in selecting the most appropriate pre-trained model to fine-tune amidst a sea of options.\nGiven constrained resources, fine-tuning all models and making selections afterward is unrealistic.\nIn this work, we formulate this resource-constrained selection task into predicting fine-tuning performance and illustrate its natural connection with Scaling Law.\nUnlike pre-training, we find that the fine-tuning scaling curve includes not just the well-known “power phase” but also the previously unobserved “pre-power phase”. We also explain why existing Scaling Law fails to capture this phase transition phenomenon both theoretically and empirically.\nTo address this, we introduce the concept of “pre-learned data size” into our Rectified Scaling Law, which overcomes theoretical limitations and fits experimental results much better.\nBy leveraging our law, we propose a novel LLM selection algorithm that selects the near-optimal model with hundreds of times less resource consumption, while other methods may provide negatively correlated selection. The project page is available atrectified-scaling-law.github.io.",
    "introduction": "Recent years have witnessed the unprecedented development of large language models (LLMs)(Touvron et al.,2023; Achiam et al.,2023), as well as the benefits they bring to numerous downstream tasks(Liu et al.,2019; Devlin et al.,2018). Among all progresses, one important technique isfine-tuning, which re-trains a pre-trained model on specific datasets to convert the model into a task-specific expert(Ke et al.,2023b,a). It has been widely demonstrated that fine-tuning can substantially improve the performance of downstream applications(Raffel et al.,2020; Alt et al.,2019). The common workflow of fine-tuning a LLM starts with selecting an appropriate pre-trained model. Thanks to the ever-growing ecosystem of LLMs like HuggingFace, we are able to choose from countless models for specific downstream task fine-tuning.\n\n\nFigure 1:(a) The Pearson correlation between the true full-fine-tuning performance and the predicted performance of three intuitive methods, given different resource constraints denoted byγ𝛾\\gammaitalic_γ. These baseline methods cannot predict performance well especially under demanding constraints (smallγ𝛾\\gammaitalic_γ), and could even provide negatively correlated predictions.\n(b) The phase transition phenomenon observed in the scaling of fine-tuning lossL𝐿Litalic_Lwith training sample sizeD𝐷Ditalic_D. In addition to the widely studied power phase where(L,D)𝐿𝐷(L,D)( italic_L , italic_D )are linearly correlated under the log-log scale, we discover the pre-power phase whenD𝐷Ditalic_Dis small. Previous laws fail to fit both phases, while our proposed law fits quite well.\n(c) Our LLM selection algorithm that extrapolates full-fine-tuning performance based on the new law.\n\n\nHowever, the explosion of open-sourced models also poses a “mixed blessing”: how can we select the model with optimal performance after fine-tuning? Given various resource constraints on time, computation and storage(Hoffmann et al.,2022a), it is unrealistic to fine-tune all candidates and make selections afterward. It is also unstable and unpredictable to rely on empirical human impressions to select LLM for a new task, such as selecting the largest one, the most well-known one, or even the one with the highest zero-shot performance on targeted tasks(Brown et al.,2020).\nIn addition, most existing model selection methods(Vu et al.,2020; Dwivedi et al.,2020)fail to solve LLM fine-tuning tasks because they were designed for classification and regression tasks, which is incompatible with generative LLMs(Bai et al.,2023).\nThis brings us to the problem of LLM selection for fine-tuning from a unified perspective, especially in a resource-constrained manner.\n\n\nTo better address this challenge, we formulateLLM Selection in the context of fine-tuningfor the first time. Our framework models the challenge as a resource-constrained task to predict thefull-fine-tuningperformance of a model,i.e.formulae-sequence𝑖𝑒i.e.italic_i . italic_e ., the performance after fine-tuning on the entire downstream task dataset.\nBy measuring the error between the predicted and the true full-fine-tuning performance, we further show that intuitive selection methods based on model size, zero-shot performance, or fine-tuned performance on a small subset, all fail to give a good full-fine-tuning performance prediction (Figure1(a)). The correlation between their prediction and the ground-truth performance is surprisingly low.\n\n\nWe point out that the challenge in predicting full-fine-tuning performance with limited resources naturally draws parallels to the study ofLLM Scaling Law(Kaplan et al.,2020), which has been successfully applied to predict the LLM pre-training performance with at most10,000×10,000\\times10 , 000 ×less compute(Achiam et al.,2023). Similarly,can we leverage Scaling Law to efficiently and accurately predict the performance of fine-tuning as well?\n\n\nIn this paper, we conduct thorough experiments on scaling behavior in fine-tuning using30303030models with sizes varying from100100100100million to7777billion. As shown inFigure1(b), we find a previously unobserved phase transition pattern called “pre-power phase” on the low-data regimes where the slope gradually decreases before the widely studied “power phase” where the test loss and number of samplesD𝐷Ditalic_Dis roughly linearly correlated. The transition is crucial for fine-tuning, as typical fine-tuning datasets can vary from hundreds to millions of samples, covering both phases.\nWe theoretically explain this phenomenon via the concept ofpre-learned data size, which represents the equivalent amount of downstream task samples that the model has pre-learned from the pre-training corpus. Inspired by this, we establish Rectified Scaling Law of LLM fine-tuning (a.k.a.formulae-sequence𝑎𝑘𝑎a.k.a.italic_a . italic_k . italic_a .“Fine-tuning Scaling Law”) by incorporating this concept (Equation7), which fits all experimental results much better than all existing laws, aligning with our theoretical judgments.\n\n\nBased on the Rectified Scaling Law of LLM fine-tuning, we design a novel LLM selection algorithm called “Accept then Stop” (AtS,Figure1(c)). Starting from the maximum allowed constraints, it keeps accepting fine-tuning results on a series of size-decreasing subsets, stops once it distinguishes the transition pattern, and uses all accepted results to linearly extrapolate the full-fine-tuning performance. The designed algorithm demonstrates outstanding LLM selection performance under extensive experimental settings, and selects the near-optimal model with hundreds of times less resource consumption, under which other approaches can provide negatively correlated selection results. Extensive ablation experiments also prove its robustness and stability.\n\n\nIn summary, we first formulate LLM selection framework with great compatibility, and draw its connection with the study of Scaling Law for model fine-tuning (Section2). We demonstrate why previous laws fail to fit fine-tuning performance both theoretically and empirically, and establish a new Scaling Law that fits much better (Section3). We propose a novel LLM selection algorithm based on the established law that significantly outperforms all other baselines under extensive experimental settings (Section4). Together, our work makes a first step towards LLM selection for fine-tuning, and towards better understanding of Scaling Law in practical downstream applications.\n\n",
    "related_work": "Model selection.Early model selection methods require that all models share identical architectures and differ only in pre-trained datasets(Cui et al.,2018; Tran et al.,2019). Those similarity-based methods(Vu et al.,2020; Dwivedi et al.,2020)fine-tune a model on a target dataset, and use the feature similarity between this model and candidate models to predict the fine-tuning performance for each model.(Ye et al.,2021)extends the feature-based method to model selection under the out-of-distribution setting. Another line of works design training-free metrics to examine whether pre-trained features are easily transferred to target tasks(P’andy et al.,2021; Ibrahim et al.,2021). More recently, there has been attempts to formulate the problem as learning to recommend(Li et al.,2023b)or rank(Zhang et al.,2023). One reason for not adopting existing model selection methods outside LLM is that they focus mainly on classification or regression tasks(Deshpande et al.,2021; Li et al.,2023a).\nThese methods either rely on features of inputs(Lin et al.,2023)or consider a fixed label set(Nguyen et al.,2020), which is not appropriate in the open-world text generation setting and could lead to the one-to-many problems(Bao et al.,2019).\nThe ever-growth of open-sourced LLM models urgently calls for the investigation of LLM selection.\n\n\nScaling Law.Laws between model performance and variables like model size or data size during pre-training have been widely studied(Rosenfeld et al.,2019; Aghajanyan et al.,2023; Fernandes et al.,2023; Frantar et al.,2023), and are applied to estimate an optimal allocation of compute for pre-training LLMs(Kaplan et al.,2020; Hoffmann et al.,2022b). Recently, more fine-grained Scaling Laws have been proposed, such as data-constrained scaling(Muennighoff et al.,2023)and hyper-parameter scaling(Bi et al.,2024). For LLM fine-tuning,Hernandez et al. (2021)compared the scaling effect between transfer learning and pre-training, andTay et al. (2021)observed the inconsistency of model size scaling between pre-training and fine-tuning. A concurrent work(Zhang et al.,2024)suggested a multiplicative law in fine-tuning scaling. However, none of these studies identified the pre-power phase in the fine-tuning process under low-data regimes, and their models fail to capture this phase transition pattern.\nWithin the broader context of deep learning,Rosenfeld et al. (2019); Alabdulmohsin et al. (2022); Caballero et al. (2023)posited the necessity of a transition phase bridging the initial random-guess point and the power-law region in from-scratch training processes.\nTheir primary approach involved modeling different phases separately and integrating them using a smooth function, which essentially introduced more parameters for Scaling Law.\nIn contrast, our proposed Rectified Scaling Law focuses on the fine-tuning of LLMs, and parameterizes the transition with a single term representing the pre-learned data size. This rectification is not only simple and intuitive but also empirically validated through solid experiments.\n\n",
    "origin_citations": [
      "Large scale fine-grained categorization and domain-specific transfer learning.",
      "Transferability and hardness of supervised classification tasks.",
      "Exploring and predicting transferability across nlp tasks.",
      "Duality diagram similarity: a generic framework for initialization selection in task transfer learning.",
      "Towards a theoretical framework of out-of-distribution generalization.",
      "Transferability estimation using bhattacharyya class separability.",
      "Newer is not always better: Rethinking transferability metrics, their peculiarities, stability and performance.",
      "Guided recommendation for model fine-tuning.",
      "Model spider: Learning to rank pre-trained models efficiently.",
      "A linearized framework and a new benchmark for model selection for fine-tuning.",
      "Class incremental learning via likelihood ratio based task prediction.",
      "Leep: A new measure to evaluate transferability of learned representations.",
      "Plato: Pre-trained dialogue generation model with discrete latent variable.",
      "A constructive prediction of the generalization error across scales, 2019.",
      "Scaling laws for generative mixed-modal language models.",
      "Scaling laws for multilingual neural machine translation.",
      "Scaling laws for sparsely-connected foundation models.",
      "Scaling laws for neural language models.",
      "Training compute-optimal large language models.",
      "Scaling data-constrained language models.",
      "Deepseek llm: Scaling open-source language models with longtermism.",
      "Scaling laws for transfer.",
      "Scale efficiently: Insights from pre-training and fine-tuning transformers.",
      "When scaling meets llm finetuning: The effect of data, model and finetuning method.",
      "Revisiting neural scaling laws in language and vision.",
      "Broken neural scaling laws, 2023."
    ],
    "citations": [
      499289,
      42232,
      418954,
      162186,
      2691,
      338124,
      340901,
      306531,
      324946,
      478428,
      164634,
      1899,
      456092,
      302811,
      227440,
      848693,
      472,
      7460,
      258244,
      485079,
      120477,
      365882,
      463601,
      228673,
      334059
    ],
    "extract_rate": 0.9615384615384616,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 88,
      "Final_Score": 83.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 74,
      "Citation_Quality": 87,
      "Synthesis_Analysis": 85,
      "Final_Score": 82.5,
      "raw_text": "Content_Coherence: 74\nCitation_Quality: 87\nSynthesis_Analysis: 85\nFinal_Score: 82"
    },
    "overall_score": 83.02000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICML.cc/2024/Conference",
        "venue": "ICML",
        "year": "2024",
        "openreview_id": "Bq2THeNXRr",
        "forum": "Bq2THeNXRr",
        "number": 9501,
        "title": "Selecting Large Language Model to Fine-tune via Rectified Scaling Law",
        "normalized_title": "selecting large language model to fine tune via rectified scaling law",
        "abstract": "The ever-growing ecosystem of LLMs has posed a challenge in selecting the most appropriate pre-trained model to fine-tune amidst a sea of options. Given constrained resources, fine-tuning all models and making selections afterward is unrealistic. In this work, we formulate this resource-constrained selection task into predicting fine-tuning performance and illustrate its natural connection with Scaling Law. Unlike pre-training, we find that the fine-tuning scaling curve includes not just the well-known \"power phase\" but also the previously unobserved \"pre-power phase\". We also explain why existing Scaling Law fails to capture this phase transition phenomenon both theoretically and empirically. To address this, we introduce the concept of \"pre-learned data size\" into our Rectified Scaling Law, which overcomes theoretical limitations and fits experimental results much better. By leveraging our law, we propose a novel LLM selection algorithm that selects the near-optimal model with hundreds of times less resource consumption, while other methods may provide negatively correlated selection. The project page is available at rectified-scaling-law.github.io.",
        "authors": [
          "Haowei Lin",
          "Baizhou Huang",
          "Haotian Ye",
          "Qinyu Chen",
          "Zihao Wang",
          "Sujian Li",
          "Jianzhu Ma",
          "Xiaojun Wan",
          "James Zou",
          "Yitao Liang"
        ],
        "decision": null,
        "pdate": 1714610495782,
        "cdate": 1706872323276
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICML.cc/2024/Conference",
          "venue": "ICML",
          "year": "2024",
          "openreview_id": "Bq2THeNXRr",
          "forum": "Bq2THeNXRr",
          "number": 9501,
          "title": "Selecting Large Language Model to Fine-tune via Rectified Scaling Law",
          "normalized_title": "selecting large language model to fine tune via rectified scaling law",
          "abstract": "The ever-growing ecosystem of LLMs has posed a challenge in selecting the most appropriate pre-trained model to fine-tune amidst a sea of options. Given constrained resources, fine-tuning all models and making selections afterward is unrealistic. In this work, we formulate this resource-constrained selection task into predicting fine-tuning performance and illustrate its natural connection with Scaling Law. Unlike pre-training, we find that the fine-tuning scaling curve includes not just the well-known \"power phase\" but also the previously unobserved \"pre-power phase\". We also explain why existing Scaling Law fails to capture this phase transition phenomenon both theoretically and empirically. To address this, we introduce the concept of \"pre-learned data size\" into our Rectified Scaling Law, which overcomes theoretical limitations and fits experimental results much better. By leveraging our law, we propose a novel LLM selection algorithm that selects the near-optimal model with hundreds of times less resource consumption, while other methods may provide negatively correlated selection. The project page is available at rectified-scaling-law.github.io.",
          "authors": [
            "Haowei Lin",
            "Baizhou Huang",
            "Haotian Ye",
            "Qinyu Chen",
            "Zihao Wang",
            "Sujian Li",
            "Jianzhu Ma",
            "Xiaojun Wan",
            "James Zou",
            "Yitao Liang"
          ],
          "decision": null,
          "pdate": 1714610495782,
          "cdate": 1706872323276
        }
      ]
    }
  },
  {
    "paper_id": 31003,
    "title": "SVG: 3D Stereoscopic Video Generation via Denoising Frame Matrix",
    "abstract": "Video generation models have demonstrated great capabilities of producing impressive monocular videos, however, the generation of 3D stereoscopic video remains under-explored. We propose a pose-free and training-free approach for generating 3D stereoscopic videos using an off-the-shelf monocular video generation model. Our method warps a generated monocular video into camera views on stereoscopic baseline using estimated video depth, and employs a novelframe matrixvideo inpainting framework. The framework leverages the video generation model to inpaint frames observed from different timestamps and views. This effective approach generates consistent and semantically coherent stereoscopic videos without scene optimization or model fine-tuning. Moreover, we develop a disocclusion boundary re-injection scheme that further improves the quality of video inpainting by alleviating the negative effects propagated from disoccluded areas in the latent space. We validate the efficacy of our proposed method by conducting experiments on videos from various generative models, including Sora[4], Lumiere[2], WALT[8], and Zeroscope[42]. The experiments demonstrate that our method has a significant improvement over previous methods. The code will be released athttps://daipengwa.github.io/SVG_ProjectPage/",
    "introduction": "As VR/AR technology advances, the demand for creating stereoscopic content and delivering immersive 3D experiences to users continues to grow.\nDue to visual sensitivity, binocular stereoscopic content should feature flawless 3D and semantic consistency between both eye views, as well as seamless temporal consistency across frames.\nWhile monocular video generation models have been extensively researched and methods are now capable of synthesizing high-fidelity videos that adhere to complex text prompts[4], there has not been much progress in the realm of generating 3D stereoscopic videos at the scene level.\nOne reason for this gap lies in the substantial amount of monocular video data that is readily available, contrasted with the scarcity of stereo video data for training models to generate stereoscopic videos directly.\n\n\nAn emergent solution is to convert generated monocular videos into stereoscopic videos using novel view synthesis[24;27].\nHowever, these methods usually overly rely on camera pose estimation, which is a challenging task on its own either using SFM[39]or joint optimization[27], and as a result tend to be unstable, particularly in dynamic scenes where cameras experience subtle motions or when the content is dominated by dynamic objects with temporally varying appearances, both of which are prevalent in generated videos.\nConsequently, these methods fail in optimizing 3D scenes and offer low-quality solutions to the task (see Fig.3).\nMoreover, these approaches are based on reconstruction, lacking the generative ability to hallucinate occluded regions in the novel views that do not appear in any of the remaining video frames.\n\n\nIn this paper, we propose an alternative pose-free and training-free framework, for the sake of robustness and generalization capability,\nthat operates solely by exploiting inference of an off-the-shelf video generation model[42]to generate high quality 3D stereoscopic videos.\nOur initial attempt follows a typical 2D to 3D image uplifting methodology[14]and extends it into the video domain.\nSpecifically, we first generate a monocular video as the left view,\nwhich is then reprojected into the right view using per-frame estimated monocular depths[46], where we apply temporal-spatial smoothing to improve the consistency of the estimated depth.\nSubsequently, we leverage an off-the-shelf video generation model’s[42]ability to generate natural videos, by adding noise and denoising the warped video frames to inpaint the disoccluded regions, inspired by diffusion-based image inpainting[1].\n\n\nHowever, this naive pipeline does not produce appealing results:\ninpainting the right-view video frames independently, without referencing the left view, typically generates semantically mismatched content.\nTo address this problem, we propose a novel representation, called theframe matrix, which contains frame sequences observed from a number of viewpoints evenly distributed along the baseline between two eyes. The frame sequences along the view direction (rows of the matrix) form videos with camera motion, while the frame sequences along the time direction (columns of the matrix) form videos with scene motions (see Fig.1second column). Since the video diffusion model has video prior for both scene and camera motions, we propose to jointly update the entireframe matrixfrom both directions. In each denoising step, we use resample techniques[28]by alternatively denoising frame sequences along the view and the time directions. Finally, we obtain a semantically consistent and temporally smooth 3D stereoscopic video by taking the leftmost and the rightmost frame sequences to represent the left-eye view and the right-eye view, respectively.\n\n\nFurthermore, we note that the inevitable resolution downsampling operation in most video generation models with latent encoding[4;2;42;8]is detrimental to the video inpainting task. During encoding, the dark pixels created by disocclusion can degrade the features near the disocclusion boundary, leading to undesirable artifacts (see Fig.5). Instead of following the inpainting scheme proposed in previous work[1], which encodes the latent feature only once, we iteratively update both the disoccluded regions in the image space and the latent feature map with generated content during the diffusion process. This approach re-injects the generated content into the disocclusion boundary, which mitigates the negative impact of dark disocclusion and effectively prevents the artifacts.\n\n\nTo validate the efficacy of our proposed method, we generate stereoscopic video from monocular videos generated by Sora[4], Lumiere[2], WALT[8], and Zeroscope[42]. Both qualitative and quantitative evaluations suggest that our approach outperforms other baselines in 3D stereoscopic video generation. Our contributions are summarized as follows:•We design a novel pipeline to generate 3D stereoscopic videos. Unlike previous work, our method does not need camera pose estimation or fine-tuning on specific datasets.•We propose a novelframe matrixrepresentation that regularizes the diffusion-based video inpainting to generate semantically consistent and temporally smooth content.•We propose a re-injection scheme that drastically reduces the negative influence of disoccluded regions in latent space and produces high-quality results.•We conduct comprehensive experiments that show the superiority of our approach over previous methods for 3D stereoscopic video generation.\n\n",
    "related_work": "Video Generation.Video generation[42;4;2;8;9;11;13;40]has achieved tremendous progress since the advent of the diffusion model[12]. Taking into account the dataset requirements and scarcity of tagged videos, a prominent approach for video generation is to extend pre-trained image generation models[37;38;36]by inserting additional temporal layers and then fine-tuning them on video data[7;3;44]. To further improve the compute efficiency and enable long clip processing, WALT[8]and Lumiere[2]proposed to compress the video in both the temporal and spatial dimensions. More recently, Sora[4]adopted a transformer diffusion architecture[34]and was trained on large-scale video datasets to produce impressive video generation results.\nDifferent from previous video generation models focusing on producing higher-quality and longer monocular videos, our method orthogonally explores the possibility of leveraging pre-trained video generation models for stereoscopic 3D video generation.\n\n\nNovel View Synthesis.Great progress has been made for novel view synthesis in both static and dynamic scenes captured by single or multiple cameras[30;47;20;17;31]. Mildenhallet al.[30]proposed to encode the static scene into neural radiance fields (NeRF), which were then used for novel view synthesis through volume rendering. For more challenging scenes with dynamic content, follow-up works additionally optimized a deformation field[32;15;33]or scene flow fields[23]to handle the motion of dynamic objects. Instead of encoding the scene into a NeRF, DynIBaR[24]leveraged nearby frames for rendering novel view images, and dynamic objects were handled by optimized motion fields. Different from methods requiring pre-computed camera poses, RoDynRF[27]jointly optimized the NeRF and camera poses from scratch. Concurrently, FVS[19]achieves novel view video synthesis using a plane-based scene representation. Although these approaches produce high-quality renderings, they are limited to scenes where the camera pose can be accurately estimated and have limited synthesis capability. In contrast, we design a method that explicitly avoids having to estimate camera poses and possesses the ability to hallucinate unseen content.\n\n\n3D Content Creation and Inpainting.Automated 3D content creation[14;5;6;48]is another related area, with emerging approaches such as inpainting[11]or multi-view generators[26;43]. Recently, Text2Room[14]proposed creating a 3D room by warping an image into novel views and using a text-guided inpainter to deal with disocclusions. WonderJourney[48]made this process automatic by including a large language model in the loop. Similar to creating static scenes, we could use pretrained video inpainter[50;22]for dynamic 3D content creation, however, these models suffer from generalization problems in creating high-quality, consistent 3D content. Lastly, Deep3D[45]is trained using 3D movies, with the goal of converting 2D videos into stereoscopic videos. However, the training data is not publicly available and it lacks the flexibility to modify videos for creative purposes, such as different stereo baselines. In this paper, we explore the possibilities of using video generation models for 3D video creation without training on specific, hard-to-obtain datasets.\n\n",
    "origin_citations": [
      "Modelscope text-to-video technical report.",
      "Video generation models as world simulators.",
      "Lumiere: A space-time diffusion model for video generation.",
      "Photorealistic video generation with diffusion models.",
      "Flexible diffusion modeling of long videos.",
      "Imagen video: High definition video generation with diffusion models.",
      "Video diffusion models.",
      "Make-a-video: Text-to-video generation without text-video data.",
      "Denoising diffusion probabilistic models.",
      "High-resolution image synthesis with latent diffusion models.",
      "Photorealistic text-to-image diffusion models with deep language understanding.",
      "Hierarchical text-conditional image generation with clip latents.",
      "Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.",
      "Align your latents: High-resolution video synthesis with latent diffusion models.",
      "Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation.",
      "Scalable diffusion models with transformers.",
      "Nerf: Representing scenes as neural radiance fields for view synthesis.",
      "Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera.",
      "Neural 3d video synthesis from multi-view video.",
      "3d gaussian splatting for real-time radiance field rendering.",
      "Instant neural graphics primitives with a multiresolution hash encoding.",
      "Nerfies: Deformable neural radiance fields.",
      "Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes.",
      "Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields.",
      "Neural scene flow fields for space-time view synthesis of dynamic scenes.",
      "Dynibar: Neural dynamic image-based rendering.",
      "Robust dynamic radiance fields.",
      "Fast view synthesis of casual videos.",
      "Text2room: Extracting textured 3d meshes from 2d text-to-image models.",
      "Neural point cloud rendering via multi-plane projection.",
      "Gaussianflow: Splatting gaussian dynamics for 4d content creation.",
      "Wonderjourney: Going from anywhere to everywhere.",
      "Zero-1-to-3: Zero-shot one image to 3d object.",
      "Stereodiffusion: Training-free stereo image generation using latent diffusion models, 2024.",
      "Propainter: Improving propagation and transformer for video inpainting.",
      "Towards an end-to-end framework for flow-guided video inpainting.",
      "Deep3d: Fully automatic 2d-to-3d video conversion with deep convolutional neural networks."
    ],
    "citations": [
      179813,
      124902,
      271516,
      147199,
      461035,
      5927,
      176197,
      4827,
      2240,
      5926,
      2558,
      185878,
      54397,
      139303,
      140016,
      253579,
      207151,
      217861,
      180969,
      58406,
      431438,
      273928,
      392944,
      431388,
      244683,
      303363,
      273783,
      494765,
      203523,
      426714,
      273000,
      495077,
      290051,
      129077,
      65153,
      245902
    ],
    "extract_rate": 0.972972972972973,
    "article_scores": {
      "Content_Quality": 83,
      "Publication_Potential": 88,
      "Final_Score": 84.5,
      "raw_text": "Content_Quality: 83\nPublication_Potential: 88\nFinal_Score: 85"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 85,
      "Synthesis_Analysis": 78,
      "Final_Score": 82.0,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 85\nSynthesis_Analysis: 78\nFinal_Score: 82"
    },
    "overall_score": 83.0,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "ICLR.cc/2025/Conference",
        "venue": "ICLR",
        "year": "2025",
        "openreview_id": "sx2jXZuhIx",
        "forum": "sx2jXZuhIx",
        "number": 5854,
        "title": "SVG: 3D Stereoscopic Video Generation via Denoising Frame Matrix",
        "normalized_title": "svg 3d stereoscopic video generation via denoising frame matrix",
        "abstract": "Video generation models have demonstrated great capability of producing impressive monocular videos, however, the generation of 3D stereoscopic video remains under-explored. We propose a pose-free and training-free approach for generating 3D stereoscopic videos using an off-the-shelf monocular video generation model. Our method warps a generated monocular video into camera views on stereoscopic baseline using estimated video depth, and employs a novel frame matrix video inpainting framework. The framework leverages the video generation model to inpaint frames observed from different timestamps and views. This effective approach generates consistent and semantically coherent stereoscopic videos without scene optimization or model fine-tuning. Moreover, we develop a disocclusion boundary re-injection scheme that further improves the quality of video inpainting by alleviating the negative effects propagated from disoccluded areas in the latent space. We validate the efficacy of our proposed method by conducting experiments on videos from various generative models, including Sora [4], Lumiere [2], WALT [8], and Zeroscope [12]. The experiments demonstrate that our method has a significant improvement over previous methods. Project page at https://daipengwa.github.io/SVG_ProjectPage/",
        "authors": [
          "Peng Dai",
          "Feitong Tan",
          "Qiangeng Xu",
          "David Futschik",
          "Ruofei Du",
          "Sean Fanello",
          "XIAOJUAN QI",
          "Yinda Zhang"
        ],
        "decision": null,
        "pdate": 1737562430015,
        "cdate": 1727325287619
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "ICLR.cc/2025/Conference",
          "venue": "ICLR",
          "year": "2025",
          "openreview_id": "sx2jXZuhIx",
          "forum": "sx2jXZuhIx",
          "number": 5854,
          "title": "SVG: 3D Stereoscopic Video Generation via Denoising Frame Matrix",
          "normalized_title": "svg 3d stereoscopic video generation via denoising frame matrix",
          "abstract": "Video generation models have demonstrated great capability of producing impressive monocular videos, however, the generation of 3D stereoscopic video remains under-explored. We propose a pose-free and training-free approach for generating 3D stereoscopic videos using an off-the-shelf monocular video generation model. Our method warps a generated monocular video into camera views on stereoscopic baseline using estimated video depth, and employs a novel frame matrix video inpainting framework. The framework leverages the video generation model to inpaint frames observed from different timestamps and views. This effective approach generates consistent and semantically coherent stereoscopic videos without scene optimization or model fine-tuning. Moreover, we develop a disocclusion boundary re-injection scheme that further improves the quality of video inpainting by alleviating the negative effects propagated from disoccluded areas in the latent space. We validate the efficacy of our proposed method by conducting experiments on videos from various generative models, including Sora [4], Lumiere [2], WALT [8], and Zeroscope [12]. The experiments demonstrate that our method has a significant improvement over previous methods. Project page at https://daipengwa.github.io/SVG_ProjectPage/",
          "authors": [
            "Peng Dai",
            "Feitong Tan",
            "Qiangeng Xu",
            "David Futschik",
            "Ruofei Du",
            "Sean Fanello",
            "XIAOJUAN QI",
            "Yinda Zhang"
          ],
          "decision": null,
          "pdate": 1737562430015,
          "cdate": 1727325287619
        }
      ]
    }
  },
  {
    "paper_id": 44009,
    "title": "SyncVIS: Synchronized Video Instance Segmentation",
    "abstract": "Recent DETR-based methods have advanced the development of Video Instance Segmentation (VIS) through transformers’ efficiency and capability in modeling spatial and temporal information. Despite harvesting remarkable progress, existing works follow asynchronous designs, which model video sequences via either video-level queries only or adopting query-sensitive cascade structures, resulting in difficulties when handling complex and challenging video scenarios. In this work, we analyze the cause of this phenomenon and the limitations of the current solutions, and propose to conduct synchronized modeling via a new framework named SyncVIS. Specifically, SyncVIS explicitly introduces video-level query embeddings and designs two key modules to synchronize video-level query with frame-level query embeddings: a synchronized video-frame modeling paradigm and a synchronized embedding optimization strategy. The former attempts to promote the mutual learning of frame- and video-level embeddings with each other and the latter divides large video sequences into small clips for easier optimization. Extensive experimental evaluations are conducted on the challenging YouTube-VIS 2019 & 2021 & 2022, and OVIS benchmarks, and SyncVIS achieves state-of-the-art results, which demonstrates the effectiveness and generality of the proposed approach. The code is available athttps://github.com/rkzheng99/SyncVIS.",
    "introduction": "Video Instance Segmentation (VIS) is a fundamental while challenging vision task that aims to detect, segment, and track object instances inside videos based on a set of predefined object categories at the same time. With the prosperous video media, VIS has attracted various attention due to its numerous vital applications in areas such as video understanding, video editing, autonomous driving, etc.\n\n\nBenefiting from favorable long-range modeling among frames, query-based offline VIS methods[6,31,15,29,17,36]like Mask2Former-VIS[6], and SeqFormer[31]begin to dominate the VIS. Inspired by the object detection method DETR[5], they learn a group of queries that can track and segment potential instances simultaneously across the multiple frames of a video. On the other hand, online VIS approaches like IDOL[32]also exploit the temporal consistency of query embeddings and associate instances via linking the corresponding query embeddings frame by frame. Albeit the success gained by those methods, we find they barely capitalize multi-frame inputs. In practice, the Mask2Former-VIS[6]would significantly perform worse if more input frames are given during training (evidenced in Fig.3). This is paradoxical to our common sense that more frames could facilitate deep learning models obtaining more motion information of instances.\n\n\nFigure 1:Comparison of video instance segmentation paradigms. Previous methods (left part) like VITA[15]adoptasynchronousquery-sensitive structures to model instance appearances and trajectories. Our model (right part) employs frame and video embeddings in a query-robustsynchronousmanner, and they synchronize with each other through the transformer decoder to generate the refined video-level query embeddings for the prediction. Also, we employ a synchronized embedding optimization strategy ‘Sync. Optim.’ instead of the classic optimization approach.\n\n\nFor this problem, many researchers[15,31,17]point out that the video-level queries are vitally hard to track the instances well if receiving many frames in training. That is because the trajectory complexity will increase in polynomials along with the number of frames. Therefore, state-of-the-art methods like SeqFormer[31]and VITA[15]usually decouple the trajectory into spatial and temporal dimensions, which are modeled by frame-level and video-level queries, respectively. Specifically, they utilize the frame-level queries to segment each frame independently and then associate these frame-level queries with video-level queries, which are responsible for the final video-level prediction. The well-trained frame-level queries guarantee the quality in the spatial dimension and thus decrease the burden of video queries. However, we argue that two issues remain in these asynchronous designs (as illustrated at the left of Fig.1). First, with the asynchronous structure, the wellness of video-level queries heavily relies on the learning of former frame-level queries, inside which some motion information may be lost because it is an image encoding stage (rather than video encoding), which leads to the sensitivity of queries to the learning quality of former stages. Second, previous works have not solved the bipartite matching among more frames (rather than single frame), and thus the optimization complexity of trajectories remains exorbitant. Both two issues block the further development of query-based methods for video instance segmentation.\n\n\nTo this end, we propose to model video and frame queries synchronously with a new framework named SyncVIS to address the above-mentioned issues. Built upon DETR-style structures[6,32], our SyncVIS has two key components: the synchronized video-frame modeling paradigm and the synchronized embedding optimization strategy. Both designs put effort into unifying the frame- and video-level predictions in synchronization. The synchronized video-frame modeling paradigm makes frame- and video-level embeddings interact with each other in a query-robust parallel manner, rather than a query-sensitive cascade structure. Then the synchronized embedding optimization strategy adds a video-level buffer state to generate more tractable intermediate bipartite matching optimization compared with only frame-level losses. Fig.1demonstrates the schematic difference between the asynchronous state-of-the-art method and our synchronous approach. Our model is schematically simple but practically more effective, with exquisite designs as follows.\n\n\nIn the synchronized video-frame modeling paradigm, we employ frame and video-level embeddings in the transformer decoder to model object segmentation and tracking synchronously. Specifically, frame-level embeddings are assigned to each sampled frame, and responsible for modeling the appearance of instances, and video-level embeddings are a set of shared instance queries for all sampled frames, which are used to characterize the general motion (In the DETR-style architecture, when video queries are associated with features across time via the decoder, they can effectively model instance-level motion through the cascade structure. In Mask2Former-VIS, the use of video queries alone enables the capture of instance motion). Frame-level embeddings are kept on each frame to attend to instances locally. In each decoder layer, the video-level embeddings are aggregated to refine frame-level embeddings on the corresponding frame. The refined frame-level embeddings, in turn, are aggregated into video-level embeddings. By repeating this synchronization in decoder layers, SyncVIS incorporates the semantics and movement of instances in each frame.\nIn the synchronized embedding optimization strategy, we focus more on video-level bipartite matching. Concretely, we decouple the input video into several clips to synchronize video and frame, and the total number of clips is related to the combinatorial number. Then, we calculate each clip loss independently by video-level bipartite matching, so that video embeddings can maintain their association ability.\n\n\nWe evaluate our SyncVIS on four popular VIS benchmarks, including YouTube-VIS 2019 & 2021 & 2022[34], and OVIS-2021[27]. The experiments show the effectiveness of our method with significant improvement over the current state-of-the-art methods VITA[15], DVIS[38], and CTVIS[37]. Our contributions are as follows:•We analyze the limitations of existing video instance segmentation methods and propose a framework named SyncVIS with synchronized video-frame modeling. It can well characterize instances’ trajectories under complex and challenging video scenarios.•We develop two critical modules: a synchronized video-frame modeling paradigm and a synchronized embedding optimization strategy. The former adopts a synchronized paradigm to alleviate error accumulation in cascade structures. The latter divides large video sequences into small clips for easier optimization.•We conduct extensive experimental evaluations on challenging VIS benchmarks, including YouTube-VIS 2019 & 2021 &2022, and OVIS 2021, and the achieved state-of-the-art results demonstrate the effectiveness and generality of the proposed approach.\n\n",
    "related_work": "Online video instance segmentation.Most online VIS methods adopt the tracking-by-detection paradigm, integrating a tracking branch into image instance segmentation models. These methods predict detection and segmentation within a local range using a few frames and associate these outputs using matching algorithms. MaskTrack R-CNN[34]incorporates a tracking branch to Mask R-CNN[12]. Many subsequent approaches[4,35,21], follow this pipeline, measuring the similarities between frame-level predictions and associating them with different matching modules. CrossVIS[35]uses the instance feature in the current frame to pixel-wisely localize the same instance in another frame. MinVIS[16]implements a query-based image instance segmentation model[7]on individual frames and associate query embeddings via bipartite matching.\n\n\nContrarily, some previous works[9,19,10,14], draw inspiration from Video Object Segmentation[25], Multi-Object Tracking[8,24,41,2,26,39], and Multi-Object Tracking and Segmentation[28]. GenVIS[14]adopts a novel target label assignment strategy and builds instance prototype memory in query-based sequential learning. IDOL[32], based on Deformable-DETR[42], introduces a contrastive learning head that acquires discriminative instance embeddings for association[11]. CTVIS[37]improved upon IDOL by constructing a consistent paradigm for both training and inference. However, online VIS methods usually adopt frame-level query and ignore the video-level associations across non-adjacent frames, which is problematic when handling complex long videos.\n\n\nOffline video instance segmentation.Offline methods predict instance masks and trajectories through the whole video in one step using the whole video as input. STEm-Seg[1]proposes a single-stage model which learns and clusters the spatio-temporal embeddings. MaskProp[3]and Propose-Reduce[19]improve association and mask quality by mask propagation. Efficient-VIS[30]uses a tracklet query paired with a tracklet proposal to represent object instances. VisTR[29]successfully adapts DETR[5]to VIS, using instance queries to model the whole video. IFC[17]proposes inter-frame communication transformers, using memory tokens to model associations across frames.\nBy adapting Mask2Former[7]to 3D spatio-temporal features, Mask2Former-VIS[6]becomes the state-of-the-art by exploiting its mask-oriented representation. TeViT[36]introduces a new approach based on transformers instead of the CNN backbone and associates temporal information efficiently. SeqFormer[31]decomposes the shared instance queries into frame-level box ones and utilizes video-level instance queries to relate different frames. Recently, VITA[15]uses object tokens to represent the whole video and employs video queries to decode semantics from object tokens. TMT-VIS[40]manages to jointly train multiple datasets to improve performance via different taxonomy information. However, these methods typically implement only video query or utilize asynchronous structures, and the final query-sensitive approaches have difficulties dealing with complex scenarios.\n\n\nFigure 2:Overview of the proposed synchronous video-frame modeling framework SyncVIS. The developed synchronized video-frame modeling paradigm enables video-level embeddings and frame-level ones to synchronize with each other in each stage of the decoder. SyncVIS also suggests a new synchronized embedding optimization strategy. As shown in the right part, SyncVIS decouples the input video frames into several sub-clips and feeds each sub-clip into the mask and classification head. By applying these modules, SyncVIS can incorporate both semantics and movement of instances in each frame in a synchronous manner for superior characterizing ability.\n\n",
    "origin_citations": [
      "Video instance segmentation.",
      "Mask r-cnn.",
      "Sipmask: Spatial information preservation for fast image and video instance segmentation.",
      "Crossover learning for fast online video instance segmentation.",
      "Sg-net: Spatial granularity network for one-stage video instance segmentation.",
      "Minvis: A minimal video instance segmentation framework without video-based training.",
      "Masked-attention mask transformer for universal image segmentation.",
      "Compfeat: Comprehensive feature aggregation for video instance segmentation.",
      "Video instance segmentation with a propose-reduce paradigm.",
      "Visolo: Grid-based space-time aggregation for efficient online video instance segmentation.",
      "A generalized framework for video instance segmentation.",
      "Video object segmentation using space-time memory networks.",
      "Motchallenge: A benchmark for single-camera multiple target tracking.",
      "Trackformer: Multi-object tracking with transformers.",
      "Global tracking transformers.",
      "Tracking without bells and whistles.",
      "Quasi-dense similarity learning for multiple object tracking.",
      "Fairmot: On the fairness of detection and re-identification in multiple object tracking.",
      "Mots: Multi-object tracking and segmentation.",
      "In defense of online models for video instance segmentation.",
      "Deformable detr: Deformable transformers for end-to-end object detection.",
      "Momentum contrast for unsupervised visual representation learning.",
      "Ctvis: Consistent training for online video instance segmentation.",
      "Stem-seg: Spatio-temporal embeddings for instance segmentation in videos.",
      "Classifying, segmenting, and tracking object instances in video with mask propagation.",
      "Efficient video instance segmentation via tracklet query and proposal.",
      "End-to-end video instance segmentation with transformers.",
      "End-to-end object detection with transformers.",
      "Video instance segmentation using inter-frame communication transformers.",
      "Mask2former for video instance segmentation.",
      "Temporally efficient vision transformer for video instance segmentation.",
      "Seqformer: Sequential transformer for video instance segmentation.",
      "Vita: Video instance segmentation via object token association.",
      "Tmt-vis: Taxonomy-aware multi-dataset joint training for video instance segmentation."
    ],
    "citations": [
      5799,
      520,
      6746,
      6752,
      6771,
      235755,
      152903,
      6760,
      6750,
      151752,
      245663,
      2652,
      575535,
      39390,
      500931,
      218892,
      350215,
      206922,
      6772,
      421027,
      7373,
      122,
      468915,
      6756,
      5798,
      518863,
      8838,
      7360,
      195326,
      536997,
      370184,
      922370,
      412276,
      271542
    ],
    "extract_rate": 1.0,
    "article_scores": {
      "Content_Quality": 82,
      "Publication_Potential": 88,
      "Final_Score": 83.8,
      "raw_text": "Content_Quality: 82\nPublication_Potential: 88\nFinal_Score: 84"
    },
    "related_work_scores": {
      "Content_Coherence": 82,
      "Citation_Quality": 86,
      "Synthesis_Analysis": 78,
      "Final_Score": 82.4,
      "raw_text": "Content_Coherence: 82\nCitation_Quality: 86\nSynthesis_Analysis: 78\nFinal_Score: 82"
    },
    "overall_score": 82.96000000000001,
    "peer_review": {
      "match_type": "normalized_title_exact",
      "match_score": 1.0,
      "title_similarity": 1.0,
      "abstract_similarity": null,
      "primary_venue": {
        "source": "openreview",
        "venue_id": "NeurIPS.cc/2024/Conference",
        "venue": "NeurIPS",
        "year": "2024",
        "openreview_id": "tTpVHsqTKf",
        "forum": "tTpVHsqTKf",
        "number": 10058,
        "title": "SyncVIS: Synchronized Video Instance Segmentation",
        "normalized_title": "syncvis synchronized video instance segmentation",
        "abstract": "Recent DETR-based methods have advanced the development of Video Instance Segmentation (VIS) through transformers' efficiency and capability in modeling spatial and temporal information. Despite harvesting remarkable progress, existing works follow asynchronous designs, which model video sequences via either video-level queries only or adopting query-sensitive cascade structures, resulting in difficulties when handling complex and challenging video scenarios. In this work, we analyze the cause of this phenomenon and the limitations of the current solutions, and propose to conduct synchronized modeling via a new framework named SyncVIS. Specifically, SyncVIS explicitly introduces video-level query embeddings and designs two key modules to synchronize video-level query with frame-level query embeddings: a synchronized video-frame modeling paradigm and a synchronized embedding optimization strategy. The former attempts to promote the mutual learning of frame- and video-level embeddings with each other and the latter divides large video sequences into small clips for easier optimization. Extensive experimental evaluations are conducted on the challenging YouTube-VIS 2019 & 2021 & 2022, and OVIS benchmarks, and SyncVIS achieves state-of-the-art results, which demonstrates the effectiveness and generality of the proposed approach. The code is available at https://github.com/rkzheng99/SyncVIS.",
        "authors": [
          "rongkun Zheng",
          "Lu Qi",
          "Xi Chen",
          "Yi Wang",
          "Kun Wang",
          "Yu Qiao",
          "Hengshuang Zhao"
        ],
        "decision": null,
        "pdate": 1727287928600,
        "cdate": 1715689746959
      },
      "all_venues": [
        {
          "source": "openreview",
          "venue_id": "NeurIPS.cc/2024/Conference",
          "venue": "NeurIPS",
          "year": "2024",
          "openreview_id": "tTpVHsqTKf",
          "forum": "tTpVHsqTKf",
          "number": 10058,
          "title": "SyncVIS: Synchronized Video Instance Segmentation",
          "normalized_title": "syncvis synchronized video instance segmentation",
          "abstract": "Recent DETR-based methods have advanced the development of Video Instance Segmentation (VIS) through transformers' efficiency and capability in modeling spatial and temporal information. Despite harvesting remarkable progress, existing works follow asynchronous designs, which model video sequences via either video-level queries only or adopting query-sensitive cascade structures, resulting in difficulties when handling complex and challenging video scenarios. In this work, we analyze the cause of this phenomenon and the limitations of the current solutions, and propose to conduct synchronized modeling via a new framework named SyncVIS. Specifically, SyncVIS explicitly introduces video-level query embeddings and designs two key modules to synchronize video-level query with frame-level query embeddings: a synchronized video-frame modeling paradigm and a synchronized embedding optimization strategy. The former attempts to promote the mutual learning of frame- and video-level embeddings with each other and the latter divides large video sequences into small clips for easier optimization. Extensive experimental evaluations are conducted on the challenging YouTube-VIS 2019 & 2021 & 2022, and OVIS benchmarks, and SyncVIS achieves state-of-the-art results, which demonstrates the effectiveness and generality of the proposed approach. The code is available at https://github.com/rkzheng99/SyncVIS.",
          "authors": [
            "rongkun Zheng",
            "Lu Qi",
            "Xi Chen",
            "Yi Wang",
            "Kun Wang",
            "Yu Qiao",
            "Hengshuang Zhao"
          ],
          "decision": null,
          "pdate": 1727287928600,
          "cdate": 1715689746959
        }
      ]
    }
  }
]