paper_id string | title string | published_date string | arxiv_categories list | legal_license string | license_url string | commercial_use_allowed bool | commercial_ip_safety_score int64 | commercial_ip_verdict string | commercial_ip_justification string | reproduction_difficulty_level int64 | reproduction_difficulty_tier string | academic_citations_count int64 | field_weighted_citation_impact_fwci float64 | primary_author string | total_authors_count int64 | authors list | affiliations list | arxiv_url string | code_audit_status string | primary_repo_url string | repo_urls list | github_stars int64 | github_forks int64 | github_last_commit string | github_license string | importance_score float64 | abstract string | title_vector_384d list | abstract_vector_384d list | target_industry string | supported_modalities list | vision_backbone_architecture string | execution_mode string | model_parameter_scale string | min_inference_vram_fp16_gb float64 | min_inference_vram_quantized_gb float64 | recommended_gpu_class string | serving_frameworks_compatible list | tested_benchmarks list | extracted_benchmark_scores list | limitations_and_tradeoffs list | reproduction_recipe string | core_problem_addressed string | key_technical_innovation string | key_quantitative_result string | trend_velocity_tier string | papers_last_30_days int64 | audited_at string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2608.03979v1 | Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent | 2026-08-04T17:45:16Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Zhen Fang | 20 | [
"Zhen Fang",
"Yu Zeng",
"Wenxuan Huang",
"Yiming Zhao",
"Shiting Huang",
"Tianfei Ren",
"Qi Lu",
"Qingnan Ren",
"Qisheng Su",
"Lionel Z. Wang",
"Qingyu Yin",
"Shuang Chen",
"Zehui Chen",
"Lin Chen",
"Zhenfei Yin",
"Yao Hu",
"Shaohui Lin",
"Wanli Ouyang",
"Shaosheng Cao",
"Feng ... | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.03979v1 | VERIFIED_LIVE | https://github.com/Osilly/Vision-DeepResearch | [
"https://github.com/Osilly/Vision-DeepResearch"
] | 676 | 0 | 2026-08-19 | Unspecified | 96.01 | We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypa... | [
-0.03285500034689903,
-0.12996000051498413,
0.021851999685168266,
-0.02513900026679039,
0.049598000943660736,
0.02371699921786785,
-0.1258379966020584,
-0.03720799833536148,
0.0166229996830225,
-0.13708500564098358,
-0.004141000099480152,
-0.0679979994893074,
-0.01759300008416176,
0.069577... | [
-0.024692999199032784,
-0.17244599759578705,
0.02419400028884411,
-0.01706399954855442,
0.11542899906635284,
0.02501700073480606,
-0.049665000289678574,
-0.03896699845790863,
-0.011988000012934208,
-0.0600150004029274,
-0.051295001059770584,
-0.06181500107049942,
-0.03937099874019623,
0.08... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"Visual Grounding & Bounding-Box Detection"
] | 3D Spatio-Temporal Video VAE & Motion Backbone | Local Open-Weights Multimodal VLM | 7B - 8B (Standard Vision LLM - LLaVA-NeXT / Qwen2-VL-7B) | 18 | 6 | Consumer GPU (RTX 4090 / 24GB) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/Osilly/Vision-DeepResearch && cd Vision-DeepResearch && (pip install -e . || pip install -r requirements.txt) | We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. | To address these challenges, we propose Video-DR, featuring a decoupled perception-exploration pipeline with stage-wise tool unlocking that compels exhaustive cross-frame visual grounding prior to web retrieval. | The 30B-A3B variant achieves 59.3%, competitive with Claude-4.5-Sonnet and demonstrating the effectiveness of our training paradigm even at compact scale. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:59.909448 |
2608.15045v1 | MOSS-VL Technical Report | 2026-08-15T05:12:53Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (Apache-2.0) | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Pengyu Wang | 32 | [
"Pengyu Wang",
"Chenkun Tan",
"Shaojun Zhou",
"Qirui Zhou",
"Yanxin Chen",
"Xingyang He",
"Huazheng Zeng",
"Jijun Cheng",
"Chenghao Wang",
"Xiaomeng Qian",
"Pengfei Wang",
"Zhan Huang",
"Shanqing Gao",
"Wei Huang",
"Longjun Cao",
"Wu Ran",
"Jie Liu",
"Changtai Zhu",
"Hongkai Wang... | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.15045v1 | VERIFIED_LIVE | https://github.com/OpenMOSS/MOSS-VL | [
"https://github.com/OpenMOSS/MOSS-VL"
] | 444 | 16 | 2026-08-18 | Apache-2.0 | 92.81 | We present MOSS-VL, an open vision-language model family that treats real-time interaction -- perceiving while it speaks -- as a first-class capability. It is co-designed across the stack: the language decoder attends to vision only through gated cross-attention, so the model can naturally see incoming frames while gen... | [
-0.04636399820446968,
0.049803998321294785,
-0.0805869996547699,
0.09887299686670303,
0.0014619999565184116,
-0.002125999890267849,
0.006971000228077173,
0.03484300151467323,
-0.03723600134253502,
-0.01595900021493435,
-0.015576000325381756,
0.053259000182151794,
-0.017969999462366104,
0.0... | [
-0.024222999811172485,
-0.11948800086975098,
-0.027080999687314034,
-0.0011950000189244747,
0.15243299305438995,
0.004081000108271837,
-0.008402000181376934,
0.008967000059783459,
0.07673799991607666,
-0.08935900032520294,
-0.059411998838186264,
-0.007275999989360571,
-0.09105700254440308,
... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Local Open-Weights Multimodal VLM | 2B - 4B (Edge / Mobile Vision VLM - Qwen2-VL-2B / Moondream) | 9 | 3.2 | Edge / Laptop GPU (RTX 3060 / Apple M-series) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/OpenMOSS/MOSS-VL && cd MOSS-VL && (pip install -e . || pip install -r requirements.txt) | We present MOSS-VL, an open vision-language model family that treats real-time interaction -- perceiving while it speaks -- as a first-class capability. | It is co-designed across the stack: the language decoder attends to vision only through gated cross-attention, so the model can naturally see incoming frames while generating; a synthesized interaction corpus supervises when to speak, when to stay silent, and when to revise; and a staged curriculum concentrates all rea... | We release all five checkpoints, the training curriculum, and the real-time inference code at https://github.com/OpenMOSS/MOSS-VL. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:15.369650 |
2608.05798v1 | KVAE: Family of Tokenizers for Multimodal Generative Models | 2026-08-06T09:34:00Z | [
"cs.CV",
"cs.LG",
"cs.SD"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Andrey Shutkin | 14 | [
"Andrey Shutkin",
"Denis Parkhomenko",
"Ivan Kirillov",
"Kirill Chernyshev",
"Kirill Malakhov",
"Ilia Vasiliev",
"Ilia Trushkin",
"Valeriya Kobenko",
"David Chikovani",
"Alexander Ivanov",
"Azat Saginbaev",
"Egor Silvestrov",
"Ivan Mikheev",
"Konstantin Zakharov"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.05798v1 | VERIFIED_LIVE | https://github.com/kandinskylab/kvae-audio | [
"https://github.com/kandinskylab/kvae-audio",
"https://github.com/kandinskylab/kvae"
] | 117 | 0 | 2026-08-19 | Unspecified | 80.92 | Latent diffusion modeling (LDM), a prominent paradigm, utilizes tokenizers to map input signal to compressed representation. This dependency positions tokenizer as an integral part of generation process itself, since it affects learning speed, quality of synthesized samples and lay foundation for later applications. Th... | [
-0.0165180005133152,
-0.1435610055923462,
0.10700400173664093,
0.029239999130368233,
0.018748000264167786,
0.009362000040709972,
-0.03545700013637543,
-0.015107999555766582,
0.050689999014139175,
-0.05868599936366081,
0.033114999532699585,
-0.1379919946193695,
0.024790000170469284,
-0.0011... | [
-0.007493999786674976,
-0.07028800249099731,
-0.029803000390529633,
-0.05976000055670738,
-0.01612200029194355,
0.022995000705122948,
-0.07742299884557724,
-0.017795000225305557,
0.043497998267412186,
-0.06266099959611893,
0.003432000055909157,
-0.0019470000406727195,
-0.011370999738574028,
... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/kandinskylab/kvae-audio && cd kvae-audio && (pip install -e . || pip install -r requirements.txt) | Latent diffusion modeling (LDM), a prominent paradigm, utilizes tokenizers to map input signal to compressed representation. | This report presents series of KVAE tokenizers for audio, image and video, all designed for subsequent text-conditioned generation: KVAE-Audio, a continuous full-band 48 kHz tokenizer with a 50 Hz latent of 64 channels; KVAE-3D -- two causal video tokenizers for 4x16x16 and 4x8x8 compression; KVAE-2D, an image model, c... | We demonstrate that reconstruction (PSNR, LPIPS, PESQ, etc.) and generation results on objective (Frechet Distance, CLIP score, CLAP score, etc.) and subjective (side-by-side evaluation) metrics matches or surpasses frontier opensource tokenizers, such as VAEs from Wan-2.2, HunyuanVideo-1.5, FLUX.2, MovieGen, StableAud... | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:51.479048 |
2608.07468v3 | SimWAM: A Simple World Action Model for End-to-End Autonomous Driving | 2026-08-07T17:59:09Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Zongchuang Zhao | 8 | [
"Zongchuang Zhao",
"Xin Zhou",
"Tianyang Xu",
"Zhengyang Sun",
"Kaixuan Zhou",
"Honglin Li",
"Dingkang Liang",
"Xiang Bai"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.07468v3 | VERIFIED_LIVE | https://github.com/H-EmbodVis/SimWAM | [
"https://github.com/H-EmbodVis/SimWAM"
] | 116 | 0 | 2026-08-19 | Unspecified | 80.88 | World-Action Models (WAMs) improve end-to-end autonomous driving by transferring video dynamics priors to action prediction, but existing methods incur costly test-time future imagination. We present SimWAM, a simple yet effective WAM that leverages future-video prediction as a training-time supervision signal. It co-t... | [
0.03505599871277809,
-0.09069599956274033,
-0.006810000166296959,
-0.0010809999657794833,
0.054683998227119446,
-0.01727999933063984,
0.012144000269472599,
0.03548600152134895,
-0.016665000468492508,
0.032492998987436295,
0.019624000415205956,
-0.06124800071120262,
-0.0024170000106096268,
... | [
-0.0022819999139755964,
-0.1363459974527359,
0.013458999805152416,
0.01700199954211712,
0.06547900289297104,
0.05350400134921074,
0.00120599998626858,
-0.014427999965846539,
0.00544800003990531,
-0.012628000229597092,
-0.03945399820804596,
-0.0721379965543747,
-0.049049001187086105,
0.1187... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"Text-to-Video & Image Synthesis (DiT)"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Local Open-Weights Multimodal VLM | 7B - 8B (Standard Vision LLM - LLaVA-NeXT / Qwen2-VL-7B) | 18 | 6 | Consumer GPU (RTX 4090 / 24GB) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/H-EmbodVis/SimWAM && cd SimWAM && (pip install -e . || pip install -r requirements.txt) | World-Action Models (WAMs) improve end-to-end autonomous driving by transferring video dynamics priors to action prediction, but existing methods incur costly test-time future imagination. | We present SimWAM, a simple yet effective WAM that leverages future-video prediction as a training-time supervision signal. | These results position SimWAM as a simple yet solid baseline that could readily benefit from advances in video generation for efficient autonomous driving. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:42.973087 |
2608.03682v3 | PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud | 2026-08-04T13:53:48Z | [
"cs.AI",
"cs.RO"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Chenghua Wang | 26 | [
"Chenghua Wang",
"Daliang Xu",
"Dongqi Cai",
"Duojin Sun",
"Hao Zhang",
"Haoze Qian",
"Huaiyuan Zhang",
"Jinshuo Cui",
"Junbo Cui",
"Kezhao Zhao",
"Longxi Gao",
"Mengwei Xu",
"Rongjie Yi",
"Ruixin Liu",
"Shangguang Wang",
"Tam Sikyuen",
"Tianyue Zhang",
"Weikai Xie",
"Xuanzhe Liu... | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.03682v3 | VERIFIED_LIVE | https://github.com/mingti-org/phyai | [
"https://github.com/mingti-org/phyai"
] | 100 | 0 | 2026-08-19 | Unspecified | 79.49 | Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. Although these settings share the same checkpoint and action semantics, they often rely on separate inference programs. To unify them, we build P... | [
-0.04840100184082985,
0.03251500055193901,
-0.0348609983921051,
0.055560000240802765,
0.003263999940827489,
-0.05193600058555603,
-0.040449999272823334,
-0.09726899862289429,
-0.027861999347805977,
0.06470999866724014,
0.019742999225854874,
-0.0513250008225441,
-0.030957000330090523,
-0.03... | [
-0.014457999728620052,
-0.03514999896287918,
-0.009964000433683395,
0.011319000273942947,
0.035277001559734344,
-0.02870599925518036,
-0.063571996986866,
-0.029968999326229095,
-0.011510999873280525,
0.07021400332450867,
-0.02035200037062168,
-0.029621999710798264,
-0.0031030001118779182,
... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Local Open-Weights Multimodal VLM | 7B - 8B (Standard Vision LLM - LLaVA-NeXT / Qwen2-VL-7B) | 18 | 6 | Consumer GPU (RTX 4090 / 24GB) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/mingti-org/phyai && cd phyai && (pip install -e . || pip install -r requirements.txt) | Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. | To unify them, we build PhyAI, a Physical AI inference engine with a single runtime that keeps architecture-specific conditioning, solver, cache, and output logic in model adapters while sharing graph execution, kernels, memory management, and parallel services. | Specialized runtimes remain faster in several configurations, so our goal is one runtime with competitive latency rather than the fastest result in every case. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:02.069677 |
2608.13552v2 | PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives | 2026-08-13T17:59:30Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Kaixin Ding | 12 | [
"Kaixin Ding",
"Xi Chen",
"Minghong Cai",
"Zhiyuan Xu",
"Yiyang Wang",
"Yuxiang Lu",
"Junyi Li",
"Shuyang Chen",
"Yuan Gao",
"Xin Tao",
"Pengfei Wan",
"Hengshuang Zhao"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.13552v2 | VERIFIED_LIVE | https://github.com/kxding/PlayWorld | [
"https://github.com/kxding/PlayWorld"
] | 79 | 3 | 2026-08-18 | Unspecified | 77.82 | Video world models simulate future states conditioned on current observations and user actions. Recent systems have demonstrated impressive video consistency and action controllability over long sequences. However, fairly comparing these interactive models remains challenging. In practice, a human player typically eval... | [
0.03906099870800972,
-0.047501999884843826,
-0.041478998959064484,
-0.06667999923229218,
0.04915099963545799,
-0.00853000022470951,
-0.022849999368190765,
-0.042323999106884,
-0.03618700057268143,
0.02731199935078621,
-0.010483999736607075,
-0.05328400060534477,
0.026976000517606735,
0.098... | [
0.02012600004673004,
-0.12379100173711777,
0.006262999959290028,
-0.0688060000538826,
0.09722399711608887,
0.0015780000248923898,
-0.04384399950504303,
-0.02179799973964691,
0.026079000905156136,
-0.013163999654352665,
-0.033604998141527176,
-0.07078699767589569,
-0.009273000061511993,
0.0... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/kxding/PlayWorld && cd PlayWorld && (pip install -e . || pip install -r requirements.txt) | Video world models simulate future states conditioned on current observations and user actions. | Recent systems have demonstrated impressive video consistency and action controllability over long sequences. | Experiments across nine state-of-the-art world models reveal that current models remain unreliable on long-horizon interactive objectives, particularly in maintaining spatial consistency and persistent state evolution. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:20.008814 |
2608.14797v1 | Beyond Tokens: A Survey on Decoding Methods for Large Language and Vision-Language Models | 2026-08-14T18:08:03Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (Apache-2.0) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Haoran Wang | 4 | [
"Haoran Wang",
"Xiongxiao Xu",
"Philip S. Yu",
"Kai Shu"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.14797v1 | VERIFIED_LIVE | https://github.com/wang2226/Awesome-LLM-Decoding | [
"https://github.com/wang2226/Awesome-LLM-Decoding"
] | 76 | 4 | 2026-08-10 | Apache-2.0 | 77.53 | Large language models (LLMs) and large vision-language models (LVLMs) have demonstrated impressive generative capabilities, yet ensuring their outputs align with user intent is still challenging. While most existing approaches address this issue at the training stage, inference-time approaches like decoding methods off... | [
-0.020798999816179276,
-0.06696200370788574,
0.04265400022268295,
-0.05025000125169754,
0.07402800023555756,
0.023036999627947807,
0.024203000590205193,
0.04579100012779236,
0.08288300037384033,
-0.03989500179886818,
-0.018265999853610992,
-0.039737001061439514,
0.04921500012278557,
0.0520... | [
-0.003886000020429492,
-0.10158500075340271,
0.017803000286221504,
-0.03863300010561943,
0.06122099980711937,
0.03908099979162216,
-0.0594789981842041,
-0.024836000055074692,
0.0407710000872612,
-0.0202649999409914,
-0.04571099951863289,
-0.021872999146580696,
0.04085800051689148,
0.013864... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/wang2226/Awesome-LLM-Decoding && cd Awesome-LLM-Decoding && (pip install -e . || pip install -r requirements.txt) | Large language models (LLMs) and large vision-language models (LVLMs) have demonstrated impressive generative capabilities, yet ensuring their outputs align with user intent is still challenging. | In this survey, we identify three emerging paradigms from recent works on decoding methods for LLMs and LVLMs, provide a systematic review of these methods, highlight ongoing challenges, and discuss potential future research directions. | Large language models (LLMs) and large vision-language models (LVLMs) have demonstrated impressive generative capabilities, yet ensuring their outputs align with user intent is still challenging. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:15.935484 |
2608.04385v1 | ReGround: Restoring Visual Grounding in Multi-Step Reasoning through Self-Diagnosis and Visual Re-Examination | 2026-08-05T02:41:53Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Lei Peng | 3 | [
"Lei Peng",
"Shuai Lv",
"Wei Hu"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.04385v1 | VERIFIED_LIVE | https://github.com/sespoir/ReGround | [
"https://github.com/sespoir/ReGround"
] | 43 | 0 | 2026-08-19 | Unspecified | 72.31 | Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. We identify a consistent benchmark-level signature associated with this degradation: across 2,510 re-examined sam... | [
0.053727999329566956,
-0.04477199912071228,
0.06168799847364426,
0.04954100027680397,
0.049908000975847244,
-0.03581500053405762,
0.013344000093638897,
0.09633199870586395,
0.009971999563276768,
-0.029467999935150146,
-0.03525099903345108,
-0.06197600066661835,
0.013775999657809734,
0.1006... | [
0.07952500134706497,
-0.05618099868297577,
0.054885998368263245,
0.04373700171709061,
0.1141429990530014,
-0.0017379999626427889,
0.00585799990221858,
0.06378699839115143,
0.02730100043118,
-0.04520300030708313,
-0.06159200146794319,
-0.057801999151706696,
0.021206000819802284,
0.104024000... | Multimodal AI, Vision-Language Models & Video Generation | [
"Visual Grounding & Bounding-Box Detection"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/sespoir/ReGround && cd ReGround && (pip install -e . || pip install -r requirements.txt) | Vision-Language Models (VLMs) often lose visual grounding during multi-step reasoning: as reasoning chains grow longer, later inference steps rely increasingly on language priors rather than image evidence. | We present ReGround, a two-stage framework that teaches VLMs to self-diagnose grounding failures and selectively re-examine visual evidence, without architectural modifications or external tools. | Experiments on eight benchmarks across two VLM backbones demonstrate consistent gains, especially on visually intensive multi-step reasoning tasks, while incurring only modest inference overhead relative to tool-augmented baselines. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:58.353479 |
2608.14790v2 | Qwen-Video-Edit: Instruction-Based Video Editing by Repurposing an Image Editing Model | 2026-08-14T18:01:29Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 1 | Level 1: Plug-and-Play (Verified Package & Checkpoint Available) | 0 | 1 | Yunpeng Bai | 4 | [
"Yunpeng Bai",
"Yossi Gandelsman",
"Michaël Gharbi",
"Qixing Huang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.14790v2 | VERIFIED_LIVE | https://github.com/yunpeng1998/Qwen-Video-Edit | [
"https://github.com/yunpeng1998/Qwen-Video-Edit"
] | 40 | 0 | 2026-08-18 | Unspecified | 72.06 | Instruction-based video editing is commonly built on video-pretrained generative backbones: a video diffusion transformer is adapted, at considerable cost, to condition on a source video and an editing instruction. In this report we explore a different route and show that a strong instruction-based image editing model ... | [
-0.03796499967575073,
-0.054395999759435654,
0.06294099986553192,
-0.05431399866938591,
0.01095999963581562,
0.036779001355171204,
-0.057121001183986664,
-0.02292099967598915,
0.07121899724006653,
0.02149900048971176,
0.0419900007545948,
0.03709699958562851,
-0.001101000001654029,
0.034237... | [
-0.03557100147008896,
-0.08330000191926956,
0.06709200143814087,
-0.05551199987530708,
0.03846700116991997,
0.027202999219298363,
-0.11611100286245346,
-0.04353800043463707,
0.08643300086259842,
0.010108999907970428,
0.006413999944925308,
0.031096000224351883,
-0.0598520003259182,
0.021877... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"Text-to-Video & Image Synthesis (DiT)"
] | Diffusion Transformer (DiT Latent Patch) | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/yunpeng1998/Qwen-Video-Edit && cd Qwen-Video-Edit && (pip install -e . || pip install -r requirements.txt) | Instruction-based video editing is commonly built on video-pretrained generative backbones: a video diffusion transformer is adapted, at considerable cost, to condition on a source video and an editing instruction. | In this report we explore a different route and show that a strong instruction-based image editing model can edit videos by operating directly on video-VAE latents. | Our results suggest that, despite the large investment in training video latent spaces, per-frame video latents remain close enough to the image domain that mature image editing priors transfer with minimal adaptation. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:15.952398 |
2608.05070v1 | HelloWorld: Enabling Socially Interactive Characters in Video World Models | 2026-08-05T17:14:19Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Liangyang Ouyang | 5 | [
"Liangyang Ouyang",
"Ruicong Liu",
"Xuangeng Chu",
"Kaipeng Zhang",
"Yoichi Sato"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.05070v1 | VERIFIED_LIVE | https://github.com/AlayaLab/HelloWorld | [
"https://github.com/AlayaLab/HelloWorld"
] | 41 | 0 | 2026-08-19 | Unspecified | 71.9 | Despite the remarkable recent progress of video world models, social interaction between users and the characters within these worlds remains unsupported. To fill this gap, we present HelloWorld, a video world model that enables social interaction with in-world characters. With a single button press, users can prompt t... | [
0.030434999614953995,
-0.10504800081253052,
0.029961999505758286,
-0.069473996758461,
0.06281699985265732,
-0.027110999450087547,
0.0009619999909773469,
-0.049761999398469925,
0.047258999198675156,
-0.015968000516295433,
0.05586700141429901,
-0.03854700177907944,
-0.020749999210238457,
0.0... | [
-0.03597399964928627,
-0.08938000351190567,
0.06340699642896652,
-0.06621699780225754,
0.06783799827098846,
-0.06444799900054932,
0.022863000631332397,
-0.025259999558329582,
0.03506999835371971,
-0.03610000014305115,
0.003674000035971403,
-0.04445299878716469,
-0.03610499948263168,
0.0361... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"Text-to-Video & Image Synthesis (DiT)",
"3D / Spatial Scene & World State Simulation"
] | Diffusion Transformer (DiT Latent Patch) | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/AlayaLab/HelloWorld && cd HelloWorld && (pip install -e . || pip install -r requirements.txt) | Despite the remarkable recent progress of video world models, social interaction between users and the characters within these worlds remains unsupported. | To fill this gap, we present HelloWorld, a video world model that enables social interaction with in-world characters. | Experiments demonstrate that HelloWorld surpasses a variety of baselines in interaction quality, while maintaining state-of-the-art picture aesthetics and camera-pose following. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:53.783225 |
2608.16793v1 | PixRestore: Unified Image Restoration via Pixel Diffusion Transformer | 2026-08-17T16:49:55Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Lingchen Sun | 9 | [
"Lingchen Sun",
"Rongyuan Wu",
"Xiangtao Kong",
"Jixin Zhao",
"Qiaosi Yi",
"Yujing Sun",
"Shuaizheng Liu",
"Zhengqiang Zhang",
"Lei Zhang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.16793v1 | VERIFIED_LIVE | https://github.com/csslc/PixRestore | [
"https://github.com/csslc/PixRestore"
] | 30 | 1 | 2026-08-18 | Unspecified | 69.75 | Unified image restoration (UIR) aims to recover high-quality (HQ) content from low-quality (LQ) images with different degradations using a single model. Most recent methods adapt large pretrained text-to-image (T2I) latent diffusion models for their strong capacity and generative priors. However, the variational autoen... | [
-0.07858700305223465,
-0.022866999730467796,
0.041593998670578,
-0.050923001021146774,
-0.02053000032901764,
-0.06644300371408463,
-0.03203599900007248,
-0.10042800009250641,
-0.029384000226855278,
0.005563999991863966,
0.0007600000244565308,
0.0820619985461235,
0.06300199776887894,
-0.020... | [
-0.08050200343132019,
-0.06168000027537346,
0.11209800094366074,
0.024636000394821167,
0.03853200003504753,
-0.044808998703956604,
-0.0907829999923706,
-0.08975499868392944,
0.020191999152302742,
-0.054878998547792435,
-0.012512999586760998,
-0.018884999677538872,
0.016899999231100082,
0.0... | Multimodal AI, Vision-Language Models & Video Generation | [
"Text-to-Video & Image Synthesis (DiT)"
] | Diffusion Transformer (DiT Latent Patch) | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/csslc/PixRestore && cd PixRestore && (pip install -e . || pip install -r requirements.txt) | Unified image restoration (UIR) aims to recover high-quality (HQ) content from low-quality (LQ) images with different degradations using a single model. | However, the variational autoencoder (VAE) in latent T2I models may discard restoration-sensitive details, while the open-ended synthesis prior can introduce content-inconsistent artifacts. | Experiments on public benchmarks and real-world test sets show that, with only about 50M parameters and single-step inference, PixRestore achieves the best overall fidelity, perceptual quality, and robustness to degradations among competing UIR models while being far more efficient. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:09.298099 |
2608.10413v1 | DriveVLA-M0: Failure-Aware Memory Augmentation for Autonomous Driving | 2026-08-11T03:01:46Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (Apache-2.0) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Zebin Xing | 13 | [
"Zebin Xing",
"Yupeng Zheng",
"Qiang Chen",
"Linbo Wang",
"Yichen Zhang",
"Pengxuan Yang",
"Junli Wang",
"Deheng Qian",
"Xiaoqing Ye",
"Junyu Han",
"Yifeng Pan",
"Qichao Zhang",
"Dongbin Zhao"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.10413v1 | VERIFIED_LIVE | https://github.com/ZebinX/DriveVLA-M0 | [
"https://github.com/ZebinX/DriveVLA-M0"
] | 29 | 1 | 2026-08-16 | Apache-2.0 | 69.22 | Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for end-to-end autonomous driving by enabling unified reasoning across perception, language, and planning. However, existing approaches lack mechanisms to exploit past failures or adapt to distribution shifts, causing the model to persist... | [
0.05179800093173981,
-0.08721599727869034,
0.001551999943330884,
0.05456100031733513,
-0.01654299907386303,
0.00954500027000904,
0.021838000044226646,
-0.015925999730825424,
-0.09061899781227112,
-0.04709399864077568,
0.0257750004529953,
-0.004891999997198582,
-0.018379999324679375,
0.0016... | [
0.029170000925660133,
-0.11272700130939484,
0.019291000440716743,
0.04038599878549576,
0.06891900300979614,
0.023831000551581383,
-0.015709999948740005,
-0.00609100004658103,
0.004393999930471182,
-0.04307200014591217,
-0.0357539989054203,
-0.011017999611794949,
0.020604999735951424,
0.103... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/ZebinX/DriveVLA-M0 && cd DriveVLA-M0 && (pip install -e . || pip install -r requirements.txt) | Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for end-to-end autonomous driving by enabling unified reasoning across perception, language, and planning. | In this paper, we propose DriveVLA-M0, a retrieval-augmented VLA with failure-aware latent memory. | Furthermore, we show that DriveVLA-M0 scales effectively with additional memory, enabling training-free performance gains through memory expansion. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:31.428994 |
2608.06197v1 | EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning | 2026-08-06T15:54:36Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Zishan Xu | 12 | [
"Zishan Xu",
"Zhiyuan Yao",
"Yuxin Chen",
"Yifu Guo",
"Zhengxi Lu",
"Yuquan Lu",
"Jinyang Huang",
"Yan Xu",
"Yasheng Wang",
"Weinan Zhang",
"Xingshan Zeng",
"Weiwen Liu"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.06197v1 | VERIFIED_LIVE | https://github.com/Within-yao/EnvACE | [
"https://github.com/Within-yao/EnvACE"
] | 23 | 0 | 2026-08-19 | Unspecified | 67.08 | Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. We introduce EnvACE, an agentic reinforcement learning method that rep... | [
-0.0005329999839887023,
-0.02832699939608574,
0.0967399999499321,
-0.01589200086891651,
0.08180899918079376,
-0.05100100114941597,
0.007230999879539013,
-0.033243998885154724,
-0.018296999856829643,
0.06213900074362755,
-0.04331500083208084,
-0.12609900534152985,
0.00940100010484457,
0.055... | [
-0.0012610000558197498,
-0.04828599840402603,
0.008449999615550041,
-0.01862799935042858,
0.06197100132703781,
-0.010714000090956688,
-0.04791399836540222,
0.02627900056540966,
0.011820999905467033,
0.02959500066936016,
-0.05529500171542168,
-0.07740999758243561,
0.030830999836325645,
0.04... | Multimodal AI, Vision-Language Models & Video Generation | [
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/Within-yao/EnvACE && cd EnvACE && (pip install -e . || pip install -r requirements.txt) | Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. | We introduce EnvACE, an agentic reinforcement learning method that replaces external environment interaction during training with world rehearsal. | Controlled studies further show that world rehearsal consistently improves policy learning across model scales. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:47.987018 |
2608.03327v2 | Screenshots or Tools? Eliciting Tool Use and Managing Multimodal Context in Hybrid GUI-MCP Computer-Use Agents | 2026-08-04T08:35:51Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Siqi Fan | 9 | [
"Siqi Fan",
"Minghao Li",
"Xiaoqian Ma",
"Wenhui Tan",
"Xiusheng Huang",
"Juntong Wu",
"Liujie Zhang",
"Shuo Shang",
"Weihang Chen"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.03327v2 | VERIFIED_LIVE | https://github.com/redai-infra/hybrid-routing-agent | [
"https://github.com/redai-infra/hybrid-routing-agent"
] | 15 | 0 | 2026-08-19 | Unspecified | 63.48 | Hybrid computer-use agents can act through screenshots or call text tools. We find that having a tool available does not settle which way the effect goes. Under one identical GUI-MCP harness on the OSWorld-MCP benchmark (309 tasks), the same MCP tools improve a reasoning model by +4.0pp and degrade a non-reasoning mode... | [
-0.027134999632835388,
-0.037324998527765274,
-0.013845999725162983,
-0.0738620012998581,
0.11500599980354309,
-0.05238499864935875,
0.06213099882006645,
0.07430599629878998,
-0.02126999944448471,
-0.045882001519203186,
0.06548900157213211,
-0.09335999935865402,
0.11013399809598923,
0.0357... | [
-0.03325600177049637,
-0.05619499832391739,
-0.03083599917590618,
0.029346000403165817,
0.059484999626874924,
-0.08056599646806717,
-0.046500999480485916,
0.07450900226831436,
0.03680000081658363,
0.01498000044375658,
0.02113400027155876,
-0.016081999987363815,
0.06969500333070755,
0.01896... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/redai-infra/hybrid-routing-agent && cd hybrid-routing-agent && (pip install -e . || pip install -r requirements.txt) | Hybrid computer-use agents can act through screenshots or call text tools. | We find that having a tool available does not settle which way the effect goes. | Dropping it and halving image history cuts input tokens by about a third, at a small accuracy cost. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:04.033546 |
2607.28627v1 | ReToken: One Token to Improve Vision-Language Models for Visual Retrieval | 2026-07-30T17:59:56Z | [
"cs.CV",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Yao Xiao | 6 | [
"Yao Xiao",
"Reuben Tan",
"Zhen Zhu",
"Yuqun Wu",
"Jianfeng Gao",
"Derek Hoiem"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2607.28627v1 | VERIFIED_LIVE | https://github.com/avaxiao/ReToken | [
"https://github.com/avaxiao/ReToken"
] | 14 | 0 | 2026-08-19 | Unspecified | 62.72 | Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a... | [
0.004939999897032976,
-0.05598999932408333,
0.06957100331783295,
0.01107999961823225,
0.051631998270750046,
0.06633999943733215,
0.06931599974632263,
0.0377810001373291,
0.06373699754476547,
-0.058859001845121384,
-0.012910000048577785,
-0.06832700222730637,
0.07064799964427948,
0.08505800... | [
-0.006957000121474266,
-0.07440699636936188,
0.013383000157773495,
0.011831999756395817,
0.07759000360965729,
0.06781599670648575,
0.01928599923849106,
-0.0069229998625814915,
0.02151699922978878,
-0.07662200182676315,
-0.021102000027894974,
-0.06335499882698059,
0.0004239999980200082,
0.0... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Local Open-Weights Multimodal VLM | 7B - 8B (Standard Vision LLM - LLaVA-NeXT / Qwen2-VL-7B) | 18 | 6 | Consumer GPU (RTX 4090 / 24GB) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Temporal Video Token Context Explosion"
] | git clone https://github.com/avaxiao/ReToken && cd ReToken && (pip install -e . || pip install -r requirements.txt) | Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. | We present ReToken, a single learnable embedding trained as an explicit retrieval target that selects a sparse set of query-relevant visual tokens from a pre-filled visual KV cache. | Trained on only a small image-QA dataset, ReToken yields consistent gains across image and video benchmarks: on Visual Haystacks it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points (>20% relative), and on LVBench it transfers zero-shot to long video for an 8.0-point gain with Qwen3VL-8B. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:22.881663 |
2608.07003v1 | HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models | 2026-08-07T09:19:16Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Yu Xue | 8 | [
"Yu Xue",
"Haoxuan Qu",
"Zhuoling Li",
"Hongbin Xu",
"Jianxiong Yin",
"Simon See",
"Hossein Rahmani",
"Jun Liu"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.07003v1 | VERIFIED_LIVE | https://github.com/zylwithxy/HRDiT | [
"https://github.com/zylwithxy/HRDiT"
] | 13 | 0 | 2026-08-19 | Unspecified | 62.44 | Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models ... | [
-0.06129400059580803,
-0.05636399984359741,
0.010529999621212482,
0.03883200138807297,
0.04841100051999092,
-0.06066200137138367,
-0.09337200224399567,
-0.034175001084804535,
-0.01659799925982952,
-0.020426999777555466,
0.05049299821257591,
0.027170000597834587,
0.024947000667452812,
-0.02... | [
-0.0640999972820282,
-0.0462150014936924,
0.07549100369215012,
0.029651999473571777,
0.06313499808311462,
-0.11343500018119812,
-0.07216999679803848,
-0.005911000072956085,
0.005226000212132931,
-0.0015920000150799751,
-0.01369599997997284,
0.046473998576402664,
0.01996299996972084,
0.0181... | Multimodal AI, Vision-Language Models & Video Generation | [
"Text-to-Video & Image Synthesis (DiT)",
"3D / Spatial Scene & World State Simulation"
] | Diffusion Transformer (DiT Latent Patch) | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/zylwithxy/HRDiT && cd HRDiT && (pip install -e . || pip install -r requirements.txt) | Training-free text-to-high-resolution image generation has recently attracted growing research attention. | To address these challenges, we propose a novel method tailored to adapt off-the-shelf DiT models for high-resolution image synthesis. | Extensive experiments show the efficacy of our method. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:44.918728 |
2608.17975v1 | aDSL: Agentic 3D Creation via Joint Agent-Program Design | 2026-08-18T16:27:44Z | [
"cs.GR",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Rui-Huan Wang | 6 | [
"Rui-Huan Wang",
"Si-Tong Wei",
"Jia-Qi He",
"Heng-Yi Wei",
"Baoquan Chen",
"Peng-Shuai Wang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.17975v1 | VERIFIED_LIVE | https://github.com/sig-pku/aDSL | [
"https://github.com/sig-pku/aDSL"
] | 12 | 0 | 2026-08-19 | Unspecified | 62.24 | Programmatic representations provide a compelling paradigm for 3D content creation, enabling fine-grained edits, interpretability, and explicit structural control. Yet, agentic workflows that rely on large language models (LLMs) to author 3D programs remain brittle, often failing to translate high-level intent into con... | [
-0.0524820014834404,
-0.03368699923157692,
-0.08506900072097778,
-0.045740000903606415,
-0.004404999781399965,
-0.09196999669075012,
-0.015057000331580639,
-0.060869000852108,
-0.02407900057733059,
0.04853000119328499,
-0.04651299864053726,
-0.011163000017404556,
0.016566000878810883,
0.04... | [
0.013825999572873116,
-0.05067399889230728,
-0.014618000015616417,
-0.011029000394046307,
0.014272999949753284,
-0.059946998953819275,
-0.05122699961066246,
-0.02449999935925007,
-0.01181000005453825,
0.006508000195026398,
-0.0916529968380928,
-0.01659799925982952,
0.040832001715898514,
0.... | Multimodal AI, Vision-Language Models & Video Generation | [
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Spatial Coordinate Inaccuracy & Grounding Drift"
] | git clone https://github.com/sig-pku/aDSL && cd aDSL && (pip install -e . || pip install -r requirements.txt) | Programmatic representations provide a compelling paradigm for 3D content creation, enabling fine-grained edits, interpretability, and explicit structural control. | In this paper, we jointly design an Agent-centric Domain-Specific Language (aDSL) and a role-specialized multi-agent system to close this gap. | Our method outperforms prior LLM-based baselines on text-to-shape and image-to-shape tasks while preserving explicit structure, editability, and interpretability. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:07.077601 |
2608.03450v1 | Balancing Efficiency and Efficacy: Training-Free Attention-Guided Switching Between Explicit and Latent Thoughts for MLLMs | 2026-08-04T10:46:10Z | [
"cs.MM",
"cs.AI",
"cs.CL",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Haoqian Kang | 8 | [
"Haoqian Kang",
"Liupeng Li",
"Kuofeng Gao",
"Jinpeng Wang",
"Zhenyu Lu",
"Bin Chen",
"Ke Chen",
"Yaowei Wang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.03450v1 | VERIFIED_LIVE | https://github.com/swordAndSnow/MM26-AGS | [
"https://github.com/swordAndSnow/MM26-AGS"
] | 11 | 0 | 2026-08-19 | Unspecified | 60.98 | Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. Explicit text-based Chain-of-Thought (CoT) is computationally expensive and prone to visual hallucinations, while existing latent reasoning methods typically require costly training. Furthe... | [
0.08791200071573257,
-0.06617700308561325,
0.009309000335633755,
-0.03137600049376488,
0.06417600065469742,
-0.0011820000363513827,
0.010243999771773815,
0.027285000309348106,
0.11102700233459473,
0.006622000131756067,
-0.01816900074481964,
0.03351600095629692,
-0.034115999937057495,
-0.04... | [
0.08212900161743164,
-0.15686200559139252,
0.03324100002646446,
0.018138999119400978,
0.12961700558662415,
-0.01218200009316206,
0.0031109999399632215,
0.03602200001478195,
0.08959800004959106,
0.00171999994199723,
-0.02774300053715706,
-0.06456299871206284,
0.04222099855542183,
0.03323300... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Fine-Grained Visual Hallucination & Object Fabrications"
] | git clone https://github.com/swordAndSnow/MM26-AGS && cd MM26-AGS && (pip install -e . || pip install -r requirements.txt) | Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. | Furthermore, directly adapting training-free LLM reasoning mechanisms to the multimodal setting yields unstable performance. | Extensive experiments demonstrate that our method achieves state-of-the-art performance, significantly improving both accuracy and inference efficiency by reducing autoregressive steps and latency. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:02.944739 |
2607.29122v1 | A Frozen Pixel-Space Diffusion Model Can Guide Itself with Its Own Samples | 2026-07-31T07:52:08Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Zixuan Fu | 6 | [
"Zixuan Fu",
"Chong Wang",
"Lanqing Guo",
"Kailai Zhou",
"Jiahao Nie",
"Bihan Wen"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2607.29122v1 | VERIFIED_LIVE | https://github.com/zfu006/SSG | [
"https://github.com/zfu006/SSG"
] | 10 | 0 | 2026-08-19 | Unspecified | 60.07 | Pixel-space diffusion models aim to learn an end-to-end generator directly over raw pixels. This is challenging because a single model must capture both global structure and local texture in the same high-dimensional space. While recent work improves pixel diffusion through alternative prediction targets, training obje... | [
-0.03700900077819824,
-0.13098299503326416,
0.03660000115633011,
0.018130000680685043,
0.04410399869084358,
-0.058472998440265656,
0.013515999540686607,
-0.035881999880075455,
0.04791099950671196,
-0.03410400077700615,
0.05806500092148781,
0.03609900176525116,
0.009974000044167042,
-0.0031... | [
-0.05290300026535988,
-0.0855410024523735,
0.08337999880313873,
0.037032999098300934,
0.0819849967956543,
-0.04835600033402443,
-0.027282999828457832,
-0.05116099864244461,
-0.0389459989964962,
-0.0634860023856163,
-0.014204000122845173,
-0.0371210016310215,
-0.037790000438690186,
0.037783... | Multimodal AI, Vision-Language Models & Video Generation | [
"Text-to-Video & Image Synthesis (DiT)"
] | Diffusion Transformer (DiT Latent Patch) | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/zfu006/SSG && cd SSG && (pip install -e . || pip install -r requirements.txt) | Pixel-space diffusion models aim to learn an end-to-end generator directly over raw pixels. | While recent work improves pixel diffusion through alternative prediction targets, training objectives, and architectures, these advances typically require training a new model from scratch. | Across multiple pixel diffusion models on ImageNet, our \textbf{Synthetic Self-Guidance (SSG)} consistently improves generation while adapter training requires less than 1$\%$ of full-model training compute: it reduces FID by over 50$\%$ across the evaluated JiT variants without classifier-free guidance (CFG) and furth... | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:22.011629 |
2608.09818v1 | MedPixel: A Unified Pixel-Language Model for Medical Reasoning and Segmentation | 2026-08-10T16:37:24Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (Apache-2.0) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Haoyu Yang | 8 | [
"Haoyu Yang",
"Meixing Shi",
"Zengjie Chen",
"Haoran Sun",
"Haitao Leng",
"Xiaoming Shi",
"Yuxiang Cai",
"Yankai Jiang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.09818v1 | VERIFIED_LIVE | https://github.com/yhy-whu/Medpixel | [
"https://github.com/yhy-whu/Medpixel"
] | 6 | 0 | 2026-08-10 | Apache-2.0 | 56.54 | Reliable medical image understanding requires models to connect clinical language and visual reasoning with pixel-level grounding. Yet medical vision-language models often lack precise localization, whereas medical segmenters typically rely on explicit target categories or precise spatial prompts. This divide is reinfo... | [
0.036687999963760376,
-0.028512999415397644,
0.041958000510931015,
-0.08306299895048141,
-0.024999000132083893,
-0.07812599837779999,
0.10664699971675873,
0.05178900063037872,
0.03547799959778786,
-0.048611998558044434,
-0.04907799884676933,
-0.04436499997973442,
0.03383199870586395,
0.095... | [
0.045490000396966934,
-0.04082399979233742,
0.06613299995660782,
-0.0560230016708374,
0.023517999798059464,
-0.03850099816918373,
0.04029199853539467,
0.03528299927711487,
0.0399789996445179,
-0.12555299699306488,
-0.07335499674081802,
-0.08544500172138214,
0.014654000289738178,
0.08407899... | Multimodal AI, Vision-Language Models & Video Generation | [
"Visual Grounding & Bounding-Box Detection",
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/yhy-whu/Medpixel && cd Medpixel && (pip install -e . || pip install -r requirements.txt) | Reliable medical image understanding requires models to connect clinical language and visual reasoning with pixel-level grounding. | Yet medical vision-language models often lack precise localization, whereas medical segmenters typically rely on explicit target categories or precise spatial prompts. | Across this task spectrum, MedPixel achieves strong performance in both pixel-level prediction and response generation, together with effective zero-shot transfer to external grounding benchmarks and robustness to imperfect spatial prompts. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:32.762717 |
2608.13602v2 | Omni-LiveAvatar: Minute-Level Real-Time Streaming Joint Audio-Video Avatar Generation | 2026-08-07T15:34:15Z | [
"cs.MM",
"cs.CV",
"cs.SD"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Lunjie Zhu | 10 | [
"Lunjie Zhu",
"Xingtong Ge",
"Fangyu Lin",
"Yi Zhang",
"Zhening Liu",
"Mengfei Li",
"Yumeng Zhang",
"Guanglu Song",
"Yu Liu",
"Jun Zhang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.13602v2 | VERIFIED_LIVE | https://github.com/Aoko955/Omni-LiveAvatar | [
"https://github.com/Aoko955/Omni-LiveAvatar"
] | 6 | 0 | 2026-08-19 | Unspecified | 56.42 | Joint audio-video generative models serve as foundation for immersive and interactive digital-human generation. Nevertheless, most existing models rely on bidirectional attention and multi-step denoising and can generate only short clips, making them unsuitable for real-time interaction over extended durations. We pres... | [
-0.0013269999762997031,
0.005615999922156334,
0.000859999970998615,
-0.053348999470472336,
0.08629000186920166,
-0.002767000114545226,
-0.01549800019711256,
-0.06334800273180008,
0.009061000309884548,
-0.0002880000101868063,
-0.030186999589204788,
-0.069582000374794,
-0.04971100017428398,
... | [
0.0005530000198632479,
-0.07665800303220749,
0.014289000071585178,
-0.028217999264597893,
0.03346699848771095,
-0.015490000136196613,
-0.06759899854660034,
-0.08956100046634674,
0.029675999656319618,
-0.05431300029158592,
-0.05886499956250191,
-0.03783800080418587,
-0.07466600090265274,
-0... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | 3D Spatio-Temporal Video VAE & Motion Backbone | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/Aoko955/Omni-LiveAvatar && cd Omni-LiveAvatar && (pip install -e . || pip install -r requirements.txt) | Joint audio-video generative models serve as foundation for immersive and interactive digital-human generation. | We present Omni-LiveAvatar, the first framework for minute-level, real-time streaming joint audio-video avatar generation. | In terms of speed, it achieves a 33$\times$ generation speedup over its teacher, LTX-2, on a single NVIDIA H200 GPU; in terms of generation quality, it outperforms accelerated baselines across visual quality, audio quality, cross-modal synchronization, and human fidelity. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:43.447605 |
2608.11741v1 | JieZi: A Large-Scale Expert-Audited Dataset and Benchmark for Ancient Chinese Character Exegesis | 2026-08-12T07:30:00Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Ran Li | 6 | [
"Ran Li",
"Huiguo He",
"Jiahuan Cao",
"Junle Liu",
"Hiuyi Cheng",
"Lianwen Jin"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.11741v1 | VERIFIED_LIVE | https://github.com/Ran00w/JieZi | [
"https://github.com/Ran00w/JieZi"
] | 5 | 0 | 2026-06-05 | NOASSERTION | 55.28 | The scholarly exegesis of ancient Chinese characters demands integrating visual observation, linguistic analysis, and historical context. However, existing computational approaches focus narrowly on subtasks such as character recognition and retrieval, lacking the structured datasets and benchmarks required for compreh... | [
-0.0656609982252121,
0.07212399691343307,
0.020634999498724937,
0.02802100032567978,
-0.002015999983996153,
0.028821999207139015,
-0.013687999919056892,
-0.05679300054907799,
0.07944100350141525,
0.05244399979710579,
0.054322998970746994,
-0.07128900289535522,
0.03078100085258484,
-0.01392... | [
-0.0770919993519783,
0.04307299852371216,
0.014671999961137772,
-0.004898999817669392,
-0.02134699933230877,
-0.01406800001859665,
-0.04257399961352348,
0.005872999783605337,
0.05288799852132797,
0.012377999722957611,
-0.006972999777644873,
-0.08047299832105637,
0.022533999755978584,
0.029... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/Ran00w/JieZi && cd JieZi && (pip install -e . || pip install -r requirements.txt) | The scholarly exegesis of ancient Chinese characters demands integrating visual observation, linguistic analysis, and historical context. | To address this limitation, we introduce Ancient Chinese Character Exegesis (ACCE), a vision-language question answering (VQA) task that models the scholarly exegesis process. | Fine-tuning on JieZi-Dataset substantially improves performance across all four levels. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:26.545396 |
2608.02039v2 | RSVideo: Are Your Vision-Language Models Ready for Remote Sensing Videos? | 2026-08-03T10:34:42Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Hongjie Zhou | 8 | [
"Hongjie Zhou",
"Shiqin Wang",
"Haoyang Chen",
"Haonan Guo",
"Di Wang",
"Juhua Liu",
"Fu Lin",
"Yong Luo"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.02039v2 | VERIFIED_LIVE | https://github.com/HongjieZhou0329/RSVideo | [
"https://github.com/HongjieZhou0329/RSVideo"
] | 5 | 0 | 2026-08-19 | Unspecified | 54.92 | Remote-sensing videos enable real-time observation of changes in target attributes, short-term activities, and scene evolution. They record motion, actions, interactions, and scene changes that cannot be captured by isolated images. Existing models primarily target single images or discrete temporal observations spanni... | [
-0.004389000125229359,
-0.18035399913787842,
0.0057219997979700565,
-0.04577700048685074,
0.14652299880981445,
0.05948000028729439,
-0.031062999740242958,
-0.04483199864625931,
0.040800001472234726,
0.01003399956971407,
-0.05061899870634079,
-0.04254399985074997,
-0.01040900032967329,
0.09... | [
0.008727000094950199,
-0.13258999586105347,
0.03040499985218048,
-0.0213600005954504,
0.22091999650001526,
0.06651300191879272,
0.01387799996882677,
-0.07196400314569473,
0.06446299701929092,
-0.014808000065386295,
-0.0516900010406971,
-0.07011300325393677,
-0.04879400134086609,
0.10671299... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"3D / Spatial Scene & World State Simulation"
] | 3D Spatio-Temporal Video VAE & Motion Backbone | Local Open-Weights Multimodal VLM | 7B - 8B (Standard Vision LLM - LLaVA-NeXT / Qwen2-VL-7B) | 18 | 6 | Consumer GPU (RTX 4090 / 24GB) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"accuracy of 40.63%"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/HongjieZhou0329/RSVideo && cd RSVideo && (pip install -e . || pip install -r requirements.txt) | Remote-sensing videos enable real-time observation of changes in target attributes, short-term activities, and scene evolution. | However, a unified evaluation setting for assessing vision-language models on continuous remote-sensing video understanding remains lacking. | RSVideo achieves a maximum absolute improvement of 9.01% with InternVL3.5-14B and attains the highest accuracy of 40.63% with Qwen3.6-27B across 26 open-source vision-language backbones. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:09.487528 |
2608.15698v1 | ConceptFormer: Learning Adaptive Latent Concepts for Query-Document Alignment in Visual Document Retrieval | 2026-08-16T12:07:12Z | [
"cs.CV",
"cs.IR"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (MIT) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Peng Chunyi | 12 | [
"Peng Chunyi",
"Xu Zhipeng",
"Yan Yukun",
"Liu Zhenghao",
"Yu Shi",
"Mei Sen",
"Sun Yubo",
"Zhang Yongheng",
"Zhou Jie",
"Gu Yu",
"Yu Ge",
"Sun Maosong"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.15698v1 | VERIFIED_LIVE | https://github.com/NEUIR/ConceptFormer | [
"https://github.com/NEUIR/ConceptFormer"
] | 4 | 0 | 2026-08-18 | MIT | 53.86 | Visual document retrieval is a critical component of multimodal retrieval-augmented generation, aiming to identify query-relevant pages from document collections where evidence is distributed across text, layout, charts, and visual structures. Recent efforts toward finer-grained supervision primarily rely on textual de... | [
-0.011452999897301197,
-0.07226300239562988,
-0.05278300121426582,
0.027348000556230545,
0.0035099999513477087,
0.027242999523878098,
-0.0001049999991664663,
0.02638700045645237,
-0.0016469999682158232,
0.056689999997615814,
0.006467000115662813,
0.017681000754237175,
0.056811001151800156,
... | [
0.020817000418901443,
-0.08903100341558456,
-0.032368000596761703,
0.050884999334812164,
0.06717599928379059,
0.0230529997497797,
-0.0004440000047907233,
0.015150000341236591,
-0.005714999977499247,
-0.006653999909758568,
-0.016242999583482742,
-0.03987500071525574,
0.0692949965596199,
0.0... | Multimodal AI, Vision-Language Models & Video Generation | [
"High-Res OCR & Document Intelligence"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Dense Document & High-Resolution OCR Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/NEUIR/ConceptFormer && cd ConceptFormer && (pip install -e . || pip install -r requirements.txt) | Visual document retrieval is a critical component of multimodal retrieval-augmented generation, aiming to identify query-relevant pages from document collections where evidence is distributed across text, layout, charts, and visual structures. | Recent efforts toward finer-grained supervision primarily rely on textual descriptions or localized visual regions as evidence proxies. | Experiments on diverse visual document retrieval benchmarks demonstrate that ConceptFormer achieves 16.7\% and 22.1\% relative improvements in average NDCG@10 over the strongest visual retrieval baseline and the strongest OCR-based text retrieval baseline, respectively. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:12.830881 |
2608.02218v1 | PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs | 2026-08-03T13:39:25Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Haojie Hu | 6 | [
"Haojie Hu",
"Chenhao Dang",
"Yaojia Liu",
"Hengrui Kang",
"Conghui He",
"Weijia Li"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.02218v1 | VERIFIED_LIVE | https://github.com/Shannon4Science/PosterMELD | [
"https://github.com/Shannon4Science/PosterMELD"
] | 4 | 0 | 2026-08-19 | Unspecified | 53.34 | Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. Existing systems hide request-level failures by scoring only completed outputs; direct image generation is not element-editable, while coding-agent workflows are costly. PosterMELD is a template-conditioned multi-agent p... | [
-0.05608300119638443,
0.04427599906921387,
-0.08706299960613251,
0.039372000843286514,
0.026381999254226685,
0.04525600001215935,
-0.07601500302553177,
-0.06485500186681747,
-0.04398500174283981,
0.015333999879658222,
-0.00785099994391203,
-0.01472499966621399,
0.09749899804592133,
-0.0389... | [
-0.09359999746084213,
0.04468600079417229,
-0.045472998172044754,
0.028551999479532242,
0.07478199899196625,
-0.025484999641776085,
-0.08399400115013123,
0.037932999432086945,
-0.010129000060260296,
0.04466399922966957,
-0.04366200044751167,
0.04337399825453758,
-0.014902000315487385,
-0.0... | Multimodal AI, Vision-Language Models & Video Generation | [
"Single-Image Visual Question Answering & Reasoning"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/Shannon4Science/PosterMELD && cd PosterMELD && (pip install -e . || pip install -r requirements.txt) | Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. | Existing systems hide request-level failures by scoring only completed outputs; direct image generation is not element-editable, while coding-agent workflows are costly. | Code and resources are available at https://github.com/Shannon4Science/PosterMELD. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:08.586532 |
2608.01113v1 | CoT-Edit: Let CoT Guide Instruction Video Editing | 2026-08-02T09:20:15Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Sen Liang | 5 | [
"Sen Liang",
"Fengbin Guan",
"Youliang Zhang",
"Xin Li",
"Zhibo Chen"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.01113v1 | VERIFIED_LIVE | https://github.com/flying-sky999/CoT-Edit | [
"https://github.com/flying-sky999/CoT-Edit"
] | 4 | 0 | 2026-08-19 | Unspecified | 53.3 | Text-driven instruction-based video editing in complex scenes remains challenging: purely textual prompts often fail to capture precise spatial relationships and physical constraints, resulting in target ambiguity and physically implausible outcomes. To address this, we propose a plan--guide--edit framework that explic... | [
0.013017999939620495,
0.022420000284910202,
0.03604400157928467,
-0.057172998785972595,
0.07346200197935104,
0.10893700271844864,
-0.05818599835038185,
0.027103999629616737,
0.05384200066328049,
0.004271000158041716,
-0.013132000342011452,
0.0575609989464283,
-0.08582700043916702,
0.047417... | [
0.09849700331687927,
-0.06481699645519257,
0.0807960033416748,
0.008547999896109104,
0.08273299783468246,
0.0030869999900460243,
-0.08976799994707108,
-0.03307399898767471,
0.049215998500585556,
-0.016385000199079514,
-0.03285500034689903,
-0.014326999895274639,
-0.04791799932718277,
0.081... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding",
"Visual Grounding & Bounding-Box Detection",
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Video Spatio-Temporal Benchmark & Frame Analysis"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/flying-sky999/CoT-Edit && cd CoT-Edit && (pip install -e . || pip install -r requirements.txt) | Text-driven instruction-based video editing in complex scenes remains challenging: purely textual prompts often fail to capture precise spatial relationships and physical constraints, resulting in target ambiguity and physically implausible outcomes. | To address this, we propose a plan--guide--edit framework that explicitly bridges semantic intent and spatial execution. | Trained first in a modular manner and then jointly, our framework achieves superior performance with reduced data requirements, delivering precise localization in scenes with multiple similar objects and physically consistent object additions, and extensive experiments demonstrate state-of-the-art performance over mult... | Explosive (>50/mo) | 1,442 | 2026-08-19T18:56:15.034420 |
2608.16628v1 | Hypergraph-based Multimodal Retrieval-Augmented Generation with Incremental Refinement | 2026-08-17T14:30:09Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (Apache-2.0) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Shenao Chen | 8 | [
"Shenao Chen",
"Yidan Xu",
"Xiangmin Han",
"Rundong Xue",
"Duanpo Wu",
"Yuhan Gao",
"Chenggang Yan",
"Yue Gao"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.16628v1 | VERIFIED_LIVE | https://github.com/ShenAoChen2001/MMHRAG | [
"https://github.com/ShenAoChen2001/MMHRAG"
] | 3 | 0 | 2026-08-19 | Apache-2.0 | 51.96 | Modern Multimodal Retrieval-Augmented Generation (M-RAG) systems are fundamentally limited by the binary connectivity paradigm of traditional simple graphs, which fails to capture the intricate, high-order correlations among heterogeneous entities, such as the N-ary relationships between a visual chart, its scattered t... | [
-0.0123070003464818,
-0.10009399801492691,
0.02733200043439865,
0.012908999808132648,
0.08329399675130844,
0.03213400021195412,
-0.005520000122487545,
-0.04938799887895584,
0.012112000025808811,
-0.058173999190330505,
0.03613800182938576,
-0.018648000434041023,
0.04545700177550316,
0.03485... | [
-0.019423000514507294,
-0.07013499736785889,
0.032471999526023865,
0.04093499854207039,
0.08827300369739532,
-0.014317000284790993,
-0.07542800158262253,
-0.07938700169324875,
-0.04985500127077103,
-0.0013909999979659915,
-0.0019829999655485153,
0.01609399914741516,
0.07774200290441513,
0.... | Multimodal AI, Vision-Language Models & Video Generation | [
"High-Res OCR & Document Intelligence"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Dense Document & High-Resolution OCR Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/ShenAoChen2001/MMHRAG && cd MMHRAG && (pip install -e . || pip install -r requirements.txt) | Modern Multimodal Retrieval-Augmented Generation (M-RAG) systems are fundamentally limited by the binary connectivity paradigm of traditional simple graphs, which fails to capture the intricate, high-order correlations among heterogeneous entities, such as the N-ary relationships between a visual chart, its scattered t... | In this paper, we propose Hyper-M2RAG, a novel framework that redefines multimodal document retrieval through High-order Hypergraph Representation Learning. | Extensive evaluations on multimodal benchmarking datasets demonstrate that Hyper-M2RAG significantly outperforms state-of-the-art methods in both retrieval precision and generation coherence. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:09.795375 |
2608.13045v1 | P2Fusion: Prompt-based Progressive Infrared-Visible Image Fusion via Dual-Prior Distillation | 2026-08-13T10:09:49Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (MIT) | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Yi Shi | 10 | [
"Yi Shi",
"Huichao Xie",
"Yuqing Wang",
"Mingyu Wang",
"Kaihui Yang",
"Yu Liu",
"Ruitao Lu",
"Lizhe Li",
"Junwei Han",
"Dingwen Zhang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.13045v1 | VERIFIED_LIVE | https://github.com/YiShi99/P2Fusion | [
"https://github.com/YiShi99/P2Fusion"
] | 3 | 0 | 2026-07-16 | MIT | 51.8 | Infrared-visible image fusion (IVIF) is pivotal for multimodal perception, yet reconciling the inherent information disparity between thermal and textural features remains a fundamental challenge. Existing prior-guided methods often rely on static constraints that induce optimization conflicts or utilize extrinsic sema... | [
0.005991999991238117,
-0.10612799972295761,
0.10586600005626678,
0.013425000011920929,
0.03565799817442894,
-0.044468000531196594,
-0.008605999872088432,
-0.0648059993982315,
0.014519999735057354,
-0.008956000208854675,
0.029380999505519867,
0.018231000751256943,
-0.019504999741911888,
-0.... | [
0.027966000139713287,
-0.12815900146961212,
0.08182600140571594,
0.05521100014448166,
0.085207998752594,
-0.0632689967751503,
0.002856000093743205,
-0.05811699852347374,
-0.004505000077188015,
-0.016246000304818153,
-0.018640000373125076,
-0.05207499861717224,
0.033218998461961746,
0.02159... | Multimodal AI, Vision-Language Models & Video Generation | [
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/YiShi99/P2Fusion && cd P2Fusion && (pip install -e . || pip install -r requirements.txt) | Infrared-visible image fusion (IVIF) is pivotal for multimodal perception, yet reconciling the inherent information disparity between thermal and textural features remains a fundamental challenge. | Existing prior-guided methods often rely on static constraints that induce optimization conflicts or utilize extrinsic semantic priors from large-scale foundation models (e.g., CLIP/DINO), which frequently fail to exploit the intrinsic modality characteristics essential for high-fidelity fusion. | Notably, our framework demonstrates consistent performance advantages in fusion quality, achieving state-of-the-art results in 14 out of 20 key evaluation metrics across 5 benchmarks. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:21.916164 |
2608.12904v1 | HounsWorld: A Multimodal World Model for Hidden Patient-State Readout, Reconstruction, and Simulation | 2026-08-13T07:41:14Z | [
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Yunhao Bai | 8 | [
"Yunhao Bai",
"Zhongwei Qiu",
"Guangyu Guo",
"Yiming Huang",
"Tony C. W. Mok",
"Qinji Yu",
"Ling Zhang",
"Yan Wang"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.12904v1 | VERIFIED_LIVE | https://github.com/byhwhite/HounsWorld.git | [
"https://github.com/byhwhite/HounsWorld.git"
] | 3 | 0 | 2026-08-19 | Unspecified | 51.8 | Clinical intelligence requires estimating a patient's underlying condition from incomplete observations rather than learning isolated mappings from scans to answers. Volumetric medical images provide dense observations of anatomy, attenuation, and lesions, whereas clinical language provides sparse but complementary sem... | [
-0.021386999636888504,
0.015883000567555428,
-0.004484000150114298,
-0.041776999831199646,
-0.00965999998152256,
-0.007480999920517206,
-0.022561000660061836,
-0.0197759997099638,
0.05565100163221359,
-0.031221000477671623,
0.029834000393748283,
0.027688000351190567,
-0.0013089999556541443,
... | [
0.011680000461637974,
-0.042413000017404556,
0.03418299928307533,
-0.020431999117136,
0.024452000856399536,
-0.029069000855088234,
-0.05488700047135353,
-0.007433000020682812,
-0.00803999975323677,
-0.0117790000513196,
-0.04934699833393097,
-0.04216400161385536,
0.015432000160217285,
0.070... | Multimodal AI, Vision-Language Models & Video Generation | [
"3D / Spatial Scene & World State Simulation"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Local Open-Weights Multimodal VLM | 2B - 4B (Edge / Mobile Vision VLM - Qwen2-VL-2B / Moondream) | 9 | 3.2 | Edge / Laptop GPU (RTX 3060 / Apple M-series) | [
"vLLM Multimodal (v0.6+)",
"SGLang (Fast Vision Batching)",
"Ollama Vision",
"TGI"
] | [
"3D Spatial Scene & World Dynamics Evaluation"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/byhwhite/HounsWorld.git && cd HounsWorld.git && (pip install -e . || pip install -r requirements.txt) | Clinical intelligence requires estimating a patient's underlying condition from incomplete observations rather than learning isolated mappings from scans to answers. | To operationalize this view, we introduce HounsBench, a computed tomography (CT) centric patient-state benchmark that unifies these three task families with patient-disjoint splits and per-family metrics, and HounsWorld, a 3B multimodal world model that treats volumetric scans and language as observations of the shared... | Our project is available at https://github.com/byhwhite/HounsWorld.git | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:24.257760 |
2608.06270v1 | The Illusion of Visual Tool-Use: A Causal Audit of Thinking with Images | 2026-08-06T17:01:08Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 2 | Level 2: Ready Codebase (Full Repository + Dependency Spec) | 0 | 1 | Zhiheng Wang | 4 | [
"Zhiheng Wang",
"Bo Peng",
"Lai Wei",
"Chaochao Lu"
] | [
"Multimodal Foundation AI Research Institute"
] | http://arxiv.org/abs/2608.06270v1 | VERIFIED_LIVE | https://github.com/OpenCausaLab/CauAudit | [
"https://github.com/OpenCausaLab/CauAudit"
] | 3 | 0 | 2026-08-19 | Unspecified | 51.52 | The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. They may also repeatedly crop irrelevant regions and fail on ques... | [
0.03545000031590462,
0.016915999352931976,
0.011687999591231346,
-0.018567999824881554,
0.07326000183820724,
-0.022888999432325363,
0.07732100039720535,
0.024699000641703606,
0.10286600142717361,
0.06148799881339073,
0.03359299898147583,
0.009752999991178513,
0.029707999899983406,
0.030674... | [
0.05398999899625778,
-0.06494899839162827,
0.06981699913740158,
0.0037720000836998224,
0.10152299702167511,
-0.013729999773204327,
-0.0518219992518425,
0.039788998663425446,
0.06543800234794617,
0.045336998999118805,
0.006930999923497438,
-0.01998100057244301,
0.03267800062894821,
0.041636... | Multimodal AI, Vision-Language Models & Video Generation | [
"Video Temporal Reasoning & Action Understanding"
] | Standard Vision Transformer (ViT) & Multi-Modal Projection | Multimodal Vision Architecture & World Model Engine | Model-Agnostic / Vision Architecture | 0 | 0 | CPU Server / Model-Agnostic Vision Host | [
"Custom PyTorch Vision Pipeline",
"HuggingFace Accelerate",
"ONNX Runtime"
] | [
"Multimodal Comprehensive Visual Question Answering Benchmark"
] | [
"Quantitative Multimodal Visual & Reasoning Evaluation"
] | [
"Requires High-Compute Multi-GPU Cluster & High-Resolution Fine-Tuning"
] | git clone https://github.com/OpenCausaLab/CauAudit && cd CauAudit && (pip install -e . || pip install -r requirements.txt) | The "thinking-with-images" paradigm equips multimodal LLMs with active visual operations such as crop-and-zoom. | However, models using these operations often achieve only marginal or negative gains over direct inference at substantially higher token cost. | We term this discrepancy the illusion of visual tool-use: despite aggregate accuracy gains, visual tool-use is not causally effective across a broad range of rollouts. | Explosive (>50/mo) | 1,442 | 2026-08-19T18:55:48.127793 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.