Datasets:
key stringlengths 8 50 | title stringlengths 11 130 | venue stringlengths 1 111 | date stringlengths 4 7 | url stringlengths 28 159 | arxiv_id stringlengths 0 10 | code stringlengths 0 68 | citations int64 0 21.3k | github_stars int64 0 64k | collections listlengths 0 3 | sections listlengths 1 3 |
|---|---|---|---|---|---|---|---|---|---|---|
10778628 | JARVIS-1: Open-World Multi-Task Agents With Memory-Augmented Multimodal Language Models | IEEE Transactions on Pattern Analysis & Machine Intelligence | 2023-11 | https://doi.ieeecomputersociety.org/10.1109/TPAMI.2024.3511593 | 2311.05997 | https://github.com/CraftJarvis/JARVIS-1 | 206 | 410 | [
"harness-design"
] | [
"targets/harness"
] |
11334583 | LongCodeZip: Compress Long Context for Code Language Models | 2025 40th IEEE/ACM International Conference on Automated Software Engineering (ASE) | 2025-10 | https://doi.org/10.1109/ase63991.2025.00020 | 2510.00446 | https://github.com/YerbaPage/LongCodeZip | 42 | 164 | [
"harness-design"
] | [
"targets/harness"
] |
anokhin2025herobench | HeroBench: A Benchmark for Long-Horizon Planning and Structured Reasoning in Virtual Worlds | arXiv preprint arXiv:2508.12782 | 2025-08 | https://arxiv.org/abs/2508.12782 | 2508.12782 | https://github.com/stefanrer/HeroBench | 6 | 14 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
bfcl2025 | The Berkeley Function Calling Leaderboard (BFCL): From tool use to agentic evaluation of large language models | Proceedings of the 42nd International Conference on Machine Learning | 2025 | https://proceedings.mlr.press/v267/patil25a.html | https://github.com/ShishirPatil/gorilla | 435 | 13,007 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
bonatti2025windows | Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scale | International Conference on Machine Learning | 2024-09 | https://arxiv.org/abs/2409.08264 | 2409.08264 | https://github.com/microsoft/WindowsAgentArena | 194 | 889 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
buildbench2025 | BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software | arXiv | 2025-09 | https://arxiv.org/abs/2509.25248 | 2509.25248 | 1 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
cai2026mossselfevolutionsourcelevelrewriting | MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems | arXiv | 2026-05 | https://arxiv.org/abs/2605.22794 | 2605.22794 | https://github.com/hkgai-official/Moss | 5 | 21 | [
"harness-design"
] | [
"targets/harness"
] |
cemri2026multi | Why Do Multi-Agent LLM Systems Fail? | Advances in Neural Information Processing Systems | 2025-03 | https://proceedings.neurips.cc/paper_files/paper/2025/hash/b1041e52d3be19f0a9bc491657488e4a-Abstract-Datasets_and_Benchmarks_Track.html | 2503.13657 | https://github.com/multi-agent-systems-failure-taxonomy/MAST | 535 | 410 | [
"benchmarks",
"harness-design"
] | [
"evidence/benchmarks",
"targets/harness"
] |
chan2025mlebench | MLE-bench: Evaluating machine learning agents on machine learning engineering | ICLR 2025 | 2024-10 | https://arxiv.org/abs/2410.07095 | 2410.07095 | https://github.com/openai/mle-bench | 360 | 1,716 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
chen2024spin | SPIN: Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models | Proceedings of the 41st International Conference on Machine Learning | 2024 | https://proceedings.mlr.press/v235/chen24j.html | https://github.com/uclaml/SPIN | 621 | 1,254 | [] | [
"targets/weights"
] | |
chen2025iterresearch | IterResearch: Rethinking Long-Horizon Agents with Interaction Scaling | arXiv preprint arXiv:2511.07327 | 2025-11 | https://arxiv.org/abs/2511.07327 | 2511.07327 | https://github.com/Alibaba-NLP/DeepResearch | 17 | 19,873 | [
"model-design"
] | [
"targets/weights"
] |
chen2025loop | Reinforcement Learning for Long-Horizon Interactive LLM Agents | arXiv preprint arXiv:2502.01600 | 2025-02 | https://arxiv.org/abs/2502.01600 | 2502.01600 | 99 | 0 | [
"model-design"
] | [
"targets/weights"
] | |
chen2025mlrbench | MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research | arXiv preprint arXiv:2505.19955 | 2025-05 | https://arxiv.org/abs/2505.19955 | 2505.19955 | https://github.com/chchenhui/mlrbench | 49 | 34 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
chen2026agent | Agent^2 RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training? | arXiv preprint arXiv:2604.10547 | 2026-04 | https://arxiv.org/abs/2604.10547 | 2604.10547 | https://github.com/microsoft/RD-Agent | 3 | 14,332 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
chen2026aiscientist | Toward autonomous long-horizon engineering for ML research | arXiv | 2026-04 | https://arxiv.org/abs/2604.13018 | 2604.13018 | https://github.com/AweAI-Team/AiScientist | 8 | 145 | [
"benchmarks",
"harness-design"
] | [
"targets/substrate"
] |
chen2026harnessxcomposableadaptiveevolvable | HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry | arXiv | 2026-06 | https://arxiv.org/abs/2606.14249 | 2606.14249 | 15 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
chen2026knowu | KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation | arXiv preprint arXiv:2604.08455 | 2026-04 | https://arxiv.org/abs/2604.08455 | 2604.08455 | https://github.com/ZJU-REAL/KnowU-Bench | 16 | 75 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
chen2026pastprologueplugincontroller | The Past Is Prologue: A Plug-in Controller for Selective Updates in Sequentially Evolving LLM Memory | arXiv | 2026-06 | https://arxiv.org/abs/2606.31121 | 2606.31121 | 1 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
chen2026recursiveselfimprovementaibounded | Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops | — | 2026-07 | https://arxiv.org/abs/2607.07663 | 2607.07663 | 6 | 0 | [
"harness-design"
] | [
"analyses"
] | |
chhikara2025mem0 | Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory | European Conference on Artificial Intelligence (ECAI) | 2025-04 | https://doi.org/10.3233/faia251160 | 2504.19413 | https://github.com/mem0ai/mem0 | 567 | 64,001 | [
"harness-design"
] | [
"targets/harness"
] |
choudhury2025processrewardmodelsllm | Process Reward Models for LLM Agents: Practical Framework and Directions | arXiv | 2025-02 | https://arxiv.org/abs/2502.10325 | 2502.10325 | https://github.com/sanjibanc/agent_prm | 79 | 59 | [
"harness-design"
] | [
"targets/harness"
] |
cirepairbench2026 | CI-Repair-Bench: A Repository-Aware Benchmark for Automated Patch Validation via CI Workflows | arXiv | 2026-04 | https://arxiv.org/abs/2604.27148 | 2604.27148 | https://github.com/RabeyaMuna/CI-REPAIR-BENCH | 0 | 1 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
clitoolbench2026 | Evaluating LLM-Based 0-to-1 Software Generation in End-to-End CLI Tool Scenarios | arXiv | 2026-04 | https://arxiv.org/abs/2604.06742 | 2604.06742 | https://github.com/kinesiatricssxilm14/CLI-Tool-Bench | 1 | 2 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
dataenvgym2025 | DataEnvGym: Data Generation Agents in Teacher Environments with Student Feedback | International Conference on Learning Representations | 2024-10 | https://arxiv.org/abs/2410.06215 | 2410.06215 | https://github.com/codezakh/DataEnvGym | 20 | 34 | [] | [
"targets/data"
] |
deepswe2026 | DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks | arXiv preprint arXiv:2607.07946 | 2026-07 | https://arxiv.org/abs/2607.07946 | 2607.07946 | https://github.com/datacurve-ai/deep-swe | 13 | 1,488 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
deng2023mind2web | Mind2Web: Towards a Generalist Agent for the Web | Advances in Neural Information Processing Systems | 2023-06 | https://arxiv.org/abs/2306.06070 | 2306.06070 | https://github.com/OSU-NLP-Group/Mind2Web | 1,370 | 1,021 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
deng2026swemilestone | SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution | International Conference on Machine Learning | 2026-03 | https://arxiv.org/abs/2603.13428 | 2603.13428 | https://github.com/DeepCommit-ai/SWE-Milestone | 6 | 70 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
dong2026longhorizon | Towards Long-Horizon Agents: A Survey | Preprints | 2026 | https://doi.org/10.20944/preprints202607.1328.v1 | 0 | 0 | [] | [
"analyses"
] | ||
du2025deepresearch | DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents | arXiv preprint arXiv:2506.11763 | 2025-06 | https://arxiv.org/abs/2506.11763 | 2506.11763 | https://github.com/Ayanami0730/deep_research_bench | 212 | 816 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
erdogan2025planandact | Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks | International Conference on Machine Learning | 2025-03 | https://arxiv.org/abs/2503.09572 | 2503.09572 | https://github.com/SqueezeAILab/plan-and-act | 197 | 45 | [
"model-design"
] | [
"targets/weights"
] |
fang-etal-2026-memp | Memp: Exploring Agent Procedural Memory | Findings of the Association for Computational Linguistics: ACL 2026 | 2025-08 | https://aclanthology.org/2026.findings-acl.866/ | 2508.06433 | https://github.com/zjunlp/MemP | 62 | 35 | [
"harness-design"
] | [
"targets/harness"
] |
featurebench2026 | FeatureBench: Benchmarking Agentic Coding for Complex Feature Development | arXiv | 2026-02 | https://arxiv.org/abs/2602.10975 | 2602.10975 | https://github.com/LiberCoders/FeatureBench | 26 | 87 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
feng2025gigpo | Group-in-Group Policy Optimization for LLM Agent Training | Advances in Neural Information Processing Systems | 2025-05 | https://arxiv.org/abs/2505.10978 | 2505.10978 | https://github.com/langfengQ/verl-agent | 375 | 2,250 | [
"model-design"
] | [
"targets/weights"
] |
frontierswe2026 | FrontierSWE | Proximal Blog | 2026 | https://frontierswe.com/blog | https://github.com/Proximal-Labs/frontier-swe | 0 | 219 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
gaosurvey | A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence | TMLR 2026 | 2025-07 | https://arxiv.org/abs/2507.21046 | 2507.21046 | https://github.com/CharlesQ9/Self-Evolving-Agents | 99 | 1,300 | [] | [
"analyses"
] |
gonzalezpumariega2025robotouille | Robotouille: An Asynchronous Planning Benchmark for LLM Agents | International Conference on Learning Representations | 2025-02 | https://arxiv.org/abs/2502.05227 | 2502.05227 | https://github.com/portal-cornell/robotouille | 37 | 46 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
gou2024critic | CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing | International Conference on Learning Representations | 2023-05 | https://arxiv.org/abs/2305.11738 | 2305.11738 | https://github.com/microsoft/ProphetNet | 863 | 746 | [
"harness-design"
] | [
"targets/harness"
] |
gou2025mind2web | Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge | arXiv preprint arXiv:2506.21506 | 2025-06 | https://arxiv.org/abs/2506.21506 | 2506.21506 | https://github.com/OSU-NLP-Group/Mind2Web-2 | 68 | 114 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
gulati2026askearlyasklate | Ask Early, Ask Late, Ask Right: When Does Clarification Timing Matter for Long-Horizon Agents? | arXiv | 2026-05 | https://arxiv.org/abs/2605.07937 | 2605.07937 | 2 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
guo2026questionansweringtaskcompletion | From Question Answering to Task Completion: A Survey on Agent System and Harness Design | arXiv | 2026-06 | https://arxiv.org/abs/2606.20683 | 2606.20683 | 3 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
gutierrez2024hipporag | HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models | Advances in Neural Information Processing Systems | 2024-05 | https://arxiv.org/abs/2405.14831 | 2405.14831 | https://github.com/OSU-NLP-Group/HippoRAG | 328 | 3,963 | [
"harness-design"
] | [
"targets/harness"
] |
han2026robocerebra | RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation | Advances in Neural Information Processing Systems | 2025-06 | https://arxiv.org/abs/2506.06677 | 2506.06677 | https://github.com/buaa-colalab/RoboCerebra | 30 | 75 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
he2024webvoyager | WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024-01 | https://arxiv.org/abs/2401.13919 | 2401.13919 | https://github.com/MinorJerry/WebVoyager | 434 | 1,122 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
hill2023mineplanner | MinePlanner: A Benchmark for Long-Horizon Planning in Large Minecraft Worlds | Proceedings of the 6th ICAPS Workshop on the International Planning Competition (WIPC) | 2023-12 | https://arxiv.org/abs/2312.12891 | 2312.12891 | https://github.com/IretonLiu/mine-pddl | 8 | 23 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
hou2026single | Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning | arXiv preprint arXiv:2607.07508 | 2026-07 | https://arxiv.org/abs/2607.07508 | 2607.07508 | 7 | 0 | [
"model-design"
] | [
"targets/weights"
] | |
hu2025adas | Automated design of agentic systems | ICLR 2025 | 2024-08 | https://arxiv.org/abs/2408.08435 | 2408.08435 | https://github.com/ShengranHu/ADAS | 282 | 1,631 | [
"harness-design"
] | [
"targets/harness"
] |
hu2025memory | Memory in the Age of AI Agents | arXiv preprint arXiv:2512.13564 | 2025-12 | https://arxiv.org/abs/2512.13564 | 2512.13564 | 245 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
hu2025step | Step-DeepResearch Technical Report | arXiv preprint arXiv:2512.20491 | 2025-12 | https://arxiv.org/abs/2512.20491 | 2512.20491 | https://github.com/stepfun-ai/StepDeepResearch | 12 | 571 | [
"harness-design"
] | [
"targets/harness"
] |
huang2024selfcorrect | Large Language Models Cannot Self-Correct Reasoning Yet | International Conference on Learning Representations | 2023-10 | https://arxiv.org/abs/2310.01798 | 2310.01798 | 1,135 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
huang2026rawexperienceskillconsumption | From Raw Experience to Skill Consumption: A Systematic Study of Model-Generated Agent Skills | arXiv | 2026-05 | https://arxiv.org/abs/2605.23899 | 2605.23899 | 13 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
imajuku2025alebench | ALE-Bench: A Benchmark for Long-Horizon Objective-Driven Algorithm Engineering | Advances in Neural Information Processing Systems | 2025-06 | https://arxiv.org/abs/2506.09050 | 2506.09050 | https://github.com/SakanaAI/ALE-Bench | 28 | 213 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
jansen2025codescientist | CodeScientist: End-to-End Semi-Automated Scientific Discovery with Code-based Experimentation | Findings of the Association for Computational Linguistics: ACL 2025 | 2025 | https://aclanthology.org/2025.findings-acl.692/ | https://github.com/allenai/codescientist | 55 | 348 | [] | [
"targets/substrate"
] | |
jiang2025aide | AIDE: AI-driven exploration in the space of code | arXiv | 2025-02 | https://arxiv.org/abs/2502.13138 | 2502.13138 | https://github.com/WecoAI/aideml | 174 | 1,491 | [] | [
"targets/data"
] |
jiang2026darwindynamicagenticallyrewriting | DARWIN: Dynamic Agentically Rewriting Self-Improving Network | arXiv | 2026-02 | https://arxiv.org/abs/2602.05848 | 2602.05848 | https://github.com/henryyjiang/DARWIN | 1 | 0 | [
"harness-design"
] | [
"targets/harness"
] |
jimenez2024swe | SWE-bench: Can Language Models Resolve Real-World GitHub Issues? | ICLR 2024 | 2023-10 | https://arxiv.org/abs/2310.06770 | 2310.06770 | https://github.com/SWE-bench/SWE-bench | 3,453 | 5,705 | [
"benchmarks"
] | [
"foundations/long-horizon"
] |
jin2026chainswe | ChainSWE: Benchmarking Coding Agents on Multi-Bug Software Maintenance | arXiv preprint arXiv:2607.02606 | 2026-07 | https://arxiv.org/abs/2607.02606 | 2607.02606 | 1 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
jin2026reveal | ReVeal: Self-Evolving Code Agents via Reliable Self-Verification | The Fourteenth International Conference on Learning Representations | 2025-06 | https://arxiv.org/abs/2506.11442 | 2506.11442 | 12 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
kamoi2024can | When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs | Transactions of the Association for Computational Linguistics | 2024-06 | https://arxiv.org/abs/2406.01297 | 2406.01297 | 327 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
kang2025acon | ACON: Optimizing Context Compression for Long-horizon LLM Agents | arXiv preprint arXiv:2510.00615 | 2025-10 | https://arxiv.org/abs/2510.00615 | 2510.00615 | https://github.com/microsoft/acon | 84 | 106 | [
"harness-design"
] | [
"targets/harness"
] |
kapoor2024omniact | OmniACT: A Dataset and Benchmark for Enabling Multimodal Generalist Autonomous Agents for Desktop and Web | Computer Vision -- ECCV 2024 | 2024-02 | https://arxiv.org/abs/2402.17553 | 2402.17553 | 167 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
kapoor2025holistic | Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation | arXiv preprint arXiv:2510.11977 | 2025-10 | https://arxiv.org/abs/2510.11977 | 2510.11977 | https://github.com/princeton-pli/hal-harness | 56 | 311 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
khalifa2026process | Process Reward Models That Think | Transactions on Machine Learning Research | 2025-04 | https://arxiv.org/abs/2504.16828 | 2504.16828 | https://github.com/mukhal/ThinkPRM | 103 | 91 | [
"harness-design"
] | [
"targets/harness"
] |
khanal2026beyond | Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents | arXiv preprint arXiv:2603.29231 | 2026-03 | https://arxiv.org/abs/2603.29231 | 2603.29231 | 5 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
kim2026sciencescalingagentsystems | Towards a Science of Scaling Agent Systems | arXiv | 2025-12 | https://arxiv.org/abs/2512.08296 | 2512.08296 | https://github.com/ybkim95/agent-scaling | 115 | 42 | [
"benchmarks",
"harness-design"
] | [
"evidence/benchmarks",
"targets/harness"
] |
kimi2026k3 | Kimi K3: Open Frontier Intelligence | arXiv | 2026-07 | https://arxiv.org/abs/2607.24653 | 2607.24653 | https://github.com/MoonshotAI/Kimi-K3 | 8 | 8,617 | [
"benchmarks",
"model-design",
"harness-design"
] | [
"evidence/benchmarks",
"targets/weights",
"targets/harness"
] |
kirgis2026shadow | Can AI Agents Conduct Open-Ended AI Research? Early Evidence from Two Case Studies | arXiv | 2026-07 | https://arxiv.org/abs/2607.27191 | 2607.27191 | 3 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
koh2024visualwebarena | VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks | Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2024-01 | https://arxiv.org/abs/2401.13649 | 2401.13649 | https://github.com/web-arena-x/visualwebarena | 0 | 485 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
kovacs2026squeez | Squeez: Task-Conditioned Tool-Output Pruning for Coding Agents | arXiv preprint arXiv:2604.04979 | 2026-04 | https://arxiv.org/abs/2604.04979 | 2604.04979 | https://github.com/KRLabsOrg/squeez | 2 | 23 | [
"harness-design"
] | [
"targets/harness"
] |
kulikov2026autodata | Autodata: An Agentic Data Scientist to Create High Quality Synthetic Data | arXiv | 2026-06 | https://arxiv.org/abs/2606.25996 | 2606.25996 | 6 | 0 | [
"model-design"
] | [
"targets/data"
] | |
kwa2026measuring | Measuring AI Ability to Complete Long Software Tasks | NeurIPS 2025 | 2025-03 | https://arxiv.org/abs/2503.14499 | 2503.14499 | https://github.com/METR/eval-analysis-public | 134 | 313 | [] | [
"evidence/measurement"
] |
laban2026llms | LLMs Get Lost In Multi-Turn Conversation | International Conference on Learning Representations | 2025-05 | https://arxiv.org/abs/2505.06120 | 2505.06120 | https://github.com/microsoft/lost_in_conversation | 398 | 296 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lange2025shinkaevolve | ShinkaEvolve: Towards open-ended and sample-efficient program evolution | arXiv | 2025-09 | https://openreview.net/forum?id=lKEdGCoDNC | 2509.19349 | https://github.com/SakanaAI/ShinkaEvolve | 139 | 1,354 | [] | [
"targets/substrate"
] |
lee2024benchmarking | Benchmarking Mobile Device Control Agents across Diverse Configurations | arXiv preprint arXiv:2404.16660 | 2024-04 | https://arxiv.org/abs/2404.16660 | 2404.16660 | https://github.com/jylee425/b-moca | 47 | 33 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lee2025learning | Learning to Contextualize Web Pages for Enhanced Decision Making by LLM Agents | The Thirteenth International Conference on Learning Representations | 2025-03 | https://arxiv.org/abs/2503.10689 | 2503.10689 | https://github.com/dgjun32/lcow_iclr2025 | 21 | 6 | [
"harness-design"
] | [
"targets/harness"
] |
lee2026metaharness | Meta-Harness: End-to-End Optimization of Model Harnesses | arXiv | 2026-03 | https://arxiv.org/abs/2603.28052 | 2603.28052 | https://github.com/stanford-iris-lab/meta-harness-tbench2-artifact | 141 | 1,183 | [
"harness-design"
] | [
"targets/harness"
] |
lee2026rhi | Recursive harness self-improvement | arXiv | 2026-07 | https://arxiv.org/abs/2607.15524 | 2607.15524 | 6 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
li2023behavior | BEHAVIOR-1K: A benchmark for embodied AI with 1,000 everyday activities and realistic simulation | Proceedings of The 6th Conference on Robot Learning | 2023 | https://proceedings.mlr.press/v205/li23a.html | https://github.com/StanfordVL/BEHAVIOR-1K | 382 | 1,657 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
li2025salt | SALT: Step-level Advantage Assignment for Long-horizon Agents via Trajectory Graph | Findings of the Association for Computational Linguistics: EACL 2026 | 2025-10 | https://arxiv.org/abs/2510.20022 | 2510.20022 | 16 | 0 | [
"model-design"
] | [
"targets/weights"
] | |
li2025sorl | Stabilizing Off-Policy Training for Long-Horizon LLM Agent via Turn-Level Importance Sampling and Clipping-Triggered Normalization | arXiv preprint arXiv:2511.20718 | 2025-11 | https://arxiv.org/abs/2511.20718 | 2511.20718 | https://github.com/Cloud0723/SORL | 4 | 0 | [
"model-design"
] | [
"targets/weights"
] |
li2025webweaver | WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research | arXiv preprint arXiv:2509.13312 | 2025-09 | https://arxiv.org/abs/2509.13312 | 2509.13312 | https://github.com/Alibaba-NLP/DeepResearch | 41 | 19,873 | [
"harness-design"
] | [
"targets/harness"
] |
li2026acmagenticcontextmanagement | ACM: Agentic Context Management for Long Horizon Tasks | arXiv | 2026-07 | https://arxiv.org/abs/2607.23809 | 2607.23809 | https://github.com/lixiaochuan2020/agentic-context-management | 0 | 31 | [
"harness-design"
] | [
"targets/harness"
] |
li2026autosota | AutoSOTA: An End-to-End Automated Research System for State-of-the-Art AI Model Discovery | arXiv | 2026-04 | https://arxiv.org/abs/2604.05550 | 2604.05550 | https://github.com/tsinghua-fib-lab/AutoSOTA | 15 | 662 | [
"benchmarks"
] | [
"targets/substrate"
] |
li2026compactionrlreinforcementlearningcontext | CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents | arXiv | 2026-07 | https://arxiv.org/abs/2607.05378 | 2607.05378 | 1 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
li2026harness | Agent Harness Engineering: A Survey | — | 2026 | https://picrew.github.io/LLM-Harness/ | https://github.com/Picrew/LLM-Harness | 0 | 2 | [] | [
"analyses"
] | |
li2026weavebench | WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces | arXiv preprint arXiv:2606.09426 | 2026-06 | https://arxiv.org/abs/2606.09426 | 2606.09426 | 4 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
liao2026kernelevolve | KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Meta | arXiv | 2025-12 | https://doi.org/10.1109/ISCA66397.2026.00063 | 2512.23236 | 2 | 0 | [] | [
"targets/substrate"
] | |
lindenbauer2025complexity | The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management | arXiv preprint arXiv:2508.21433 | 2025-08 | https://arxiv.org/abs/2508.21433 | 2508.21433 | 23 | 0 | [
"harness-design"
] | [
"targets/harness"
] | |
liu2024agentbench | AgentBench: Evaluating LLMs as Agents | International Conference on Learning Representations | 2023-08 | https://arxiv.org/abs/2308.03688 | 2308.03688 | https://github.com/THUDM/AgentBench | 1,185 | 3,691 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
liu2025mlmaster | ML-Master: Towards AI-for-AI via integration of exploration and reasoning | arXiv | 2025-06 | https://arxiv.org/abs/2506.16499 | 2506.16499 | https://github.com/sjtu-sai-agents/ML-Master | 54 | 447 | [
"benchmarks"
] | [
"targets/data"
] |
liu2026diveclaudecodedesign | Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems | arXiv | 2026-04 | https://arxiv.org/abs/2604.14228 | 2604.14228 | https://github.com/VILA-Lab/Dive-into-Claude-Code | 28 | 2,078 | [
"harness-design"
] | [
"targets/harness"
] |
liu2026escherloopmutualevolutionclosedloop | Escher-Loop: Mutual Evolution by Closed-Loop Self-Referential Optimization | arXiv | 2026-04 | https://arxiv.org/abs/2604.23472 | 2604.23472 | https://github.com/scaling-group/escher-loop | 5 | 7 | [
"harness-design"
] | [
"targets/harness"
] |
liu2026llms | Do LLMs Catch Their Own Mistakes? A Comprehensive Benchmark for Reflective Tool Use LLMs | Findings of the Association for Computational Linguistics: ACL 2026 | 2026 | https://aclanthology.org/2026.findings-acl.86/ | 0 | 0 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | ||
longclibench2026 | LongCLI-Bench: A Preliminary Benchmark and Study for Long-horizon Agentic Programming in Command-Line Interfaces | arXiv | 2026-02 | https://arxiv.org/abs/2602.14337 | 2602.14337 | https://github.com/finyorko/longcli-bench | 21 | 46 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lu-etal-2025-runaway | Runaway is Ashamed, But Helpful: On the Early-Exit Behavior of Large Language Model-based Agents in Embodied Environments | Findings of the Association for Computational Linguistics: EMNLP 2025 | 2025-05 | https://aclanthology.org/2025.findings-emnlp.1304/ | 2505.17616 | https://github.com/Coldmist-Lu/AgentExit | 7 | 2 | [
"harness-design"
] | [
"targets/harness"
] |
lu2024aiscientist | The AI Scientist: Towards fully automated open-ended scientific discovery | arXiv | 2024-08 | https://arxiv.org/abs/2408.06292 | 2408.06292 | https://github.com/SakanaAI/AI-Scientist | 1,059 | 14,442 | [] | [
"targets/research"
] |
lu2026endtoendautomation | Towards End-to-End Automation of AI Research | Nature 2026 | 2026-03 | https://doi.org/10.1038/s41586-026-10265-5 | https://github.com/SakanaAI/AI-Scientist-v2 | 211 | 7,048 | [
"benchmarks"
] | [
"evidence/benchmarks"
] | |
lu2026meta | The Meta-Agent Challenge: Are Current Agents Capable of Autonomous Agent Development? | arXiv preprint arXiv:2606.04455 | 2026-06 | https://arxiv.org/abs/2606.04455 | 2606.04455 | https://github.com/ant-research/meta-agent-challenge | 2 | 20 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lu2402weblinx | WebLINX: Real-World Website Navigation with Multi-Turn Dialogue | International Conference on Machine Learning | 2024-02 | https://arxiv.org/abs/2402.05930 | 2402.05930 | https://github.com/McGill-NLP/weblinx | 182 | 163 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lu2408toolsandbox | ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities | Findings of the Association for Computational Linguistics: NAACL 2025 | 2024-08 | https://aclanthology.org/2025.findings-naacl.65/ | 2408.04682 | https://github.com/apple/ToolSandbox | 221 | 279 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
lupidi2026airsbench | AIRS-Bench: A Suite of Tasks for Frontier AI Research Science Agents | arXiv | 2026-02 | https://arxiv.org/abs/2602.06855 | 2602.06855 | https://github.com/facebookresearch/airs-bench | 19 | 111 | [
"benchmarks"
] | [
"evidence/benchmarks"
] |
Awesome AI4AI — the catalog behind the survey
The structured catalog accompanying "AI4AI Survey: From Long-Horizon Agents to Recursive Self-Improvement — Definitions, Reliable Horizons, and Open Problems", by 23 authors across TJU, SJTU, UC Berkeley, UCAS, NUS, NTU, and Simple Agent Lab.
- 📄 Paper: https://www.preprints.org/manuscript/202608.2108/v1
- 🔗 DOI: https://doi.org/10.5281/zenodo.22198847
- 📥 PDF, original layout: https://raw.githubusercontent.com/KaiWU5/Awesome-AI4AI/main/assets/AI4AI-Survey.pdf
- 🌐 Project site: https://kaiwu5.github.io/Awesome-AI4AI/
- ⭐ Repo (updated weekly): https://github.com/KaiWU5/Awesome-AI4AI
- 🔧 Harness: https://github.com/simple-agent-lab/RSIHub
What this is
"Can AI improve AI" is asserted constantly and measured rarely. The literature is fragmented: long-horizon agents, AI4AI, self-improvement, and recursive self-improvement are four largely separate conversations that don't reliably cite each other, so it is genuinely hard to tell how much progress is real.
We read 223 papers and organized them around one question:
How far can an AI system reliably carry an improvement process from idea to verified result?
That reframing forces you to say what is being improved — data, weights, the harness, the evaluator, or the research process itself — and, critically, who supplies each part of the loop.
Finding 1: AI does the work; humans still set the bar
Today's systems are increasingly excellent at the work of improvement: planning, writing code, running experiments, optimizing, repairing. But humans still overwhelmingly determine the goals, the evaluation criteria, and what counts as progress. The part that got automated and the part that decides whether the automation was worth running are not the same part.
Finding 2: The composition gap
Strong performance on individual components rarely translates into reliable end-to-end improvement. A system can beat every component benchmark and still fail the full idea→verified-result loop. A lot of "self-improving AI" claims are really component claims being extrapolated across a gap the evidence doesn't cover.
On evidence standards
We separate demonstrated capability from extrapolated autonomy. For an intervention to count as extending the reliable horizon, the evaluation has to actually attribute a boundary shift — matched evaluation, not a number rising on a benchmark whose contamination status is unclear. By that standard, evidence for reliable research judgment, causal experimentation, persistent gains, and compounding improvement is thinner than the discourse suggests.
Contents
| Collection | Papers |
|---|---|
| Benchmarks | 111 |
| Harness design | 98 |
| Model design | 26 |
| Total public papers | 223 |
225 have arXiv IDs; 166 link to code. Citations, GitHub stars, and rankings are refreshed weekly in the source repository.
Schema
One JSON object per line in papers.jsonl:
| field | type | description |
|---|---|---|
key |
string | citation key from the survey bibliography |
title |
string | paper title |
venue |
string | publication venue |
date |
string | YYYY-MM |
url |
string | canonical link (DOI or publisher) |
arxiv_id |
string | arXiv ID where one exists |
code |
string | code repository where one exists |
citations |
int | citation count at last refresh |
github_stars |
int | stars at last refresh |
collections |
list[string] | benchmarks, harness-design, model-design |
sections |
list[string] | survey section(s) citing the paper |
Usage
from datasets import load_dataset
ds = load_dataset("awesome-ai4ai", split="train")
# every benchmark paper that ships code
benches = ds.filter(lambda r: "benchmarks" in r["collections"] and r["code"])
print(len(benches))
Citation
@techreport{{ai4ai2026survey,
title = {{{{AI4AI}} Survey: From Long-Horizon Agents to Recursive Self-Improvement---Definitions, Reliable Horizons, and Open Problems}},
author = {{Wu, Kai and Lyu, Hao and Luo, Zhen and others}},
year = {{2026}},
howpublished = {{Preprints.org}},
doi = {{10.5281/zenodo.22198847}},
url = {{https://doi.org/10.5281/zenodo.22198847}}
}}
License
Catalog metadata: MIT. The survey itself is CC BY 4.0. Individual papers remain under their own licenses; this dataset holds metadata and links only.
- Downloads last month
- -