{"id":"arxiv:2603.28052","title":"Meta-Harness: End-to-End Optimization of Model Harnesses","date":"2026-03-30","section":"Harness","quality":0.315536,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/","source_url":"https://arxiv.org/abs/2603.28052","text_available":true,"authors":["Yoonho Lee","Roshen Nair","Qizheng Zhang","Kangwook Lee"],"outgoing":[{"source":"arxiv:2603.28052","target":"arxiv:2303.17651","evidence":"title","id":"arxiv:2303.17651","title":"Self-Refine: Iterative Refinement with Self-Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/accef2157d87cf33be93cb3a9af246528b0a3742a9e2e700a194956fac8f5cf4/"},{"source":"arxiv:2603.28052","target":"arxiv:2408.08435","evidence":"title","id":"arxiv:2408.08435","title":"Automated Design of Agentic Systems","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c9890d5668d1c1b33e29569b74626fd586f456af6a97c657b8c99895c90f59a5/"},{"source":"arxiv:2603.28052","target":"arxiv:2507.19457","evidence":"id","id":"arxiv:2507.19457","title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/"},{"source":"arxiv:2603.28052","target":"arxiv:2512.18746","evidence":"id","id":"arxiv:2512.18746","title":"MemEvolve: Meta-Evolution of Agent Memory Systems","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2161237fe9d9768afaa2880a40019455779801af5fba05be5781c229c18ea86/"},{"source":"arxiv:2603.28052","target":"arxiv:2512.24601","evidence":"id","id":"arxiv:2512.24601","title":"Recursive Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ac41983c934a3432ce7b6531e367b9e639bddbed93d4c6513781e76407d9a8bd/"},{"source":"arxiv:2603.28052","target":"arxiv:2601.16175","evidence":"id","id":"arxiv:2601.16175","title":"Learning to Discover at Test Time","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5d0b45cbbc72945765e689cbfb7ac7e4e92dfbb764861b27682ac0744ef05b70/"},{"source":"arxiv:2603.28052","target":"arxiv:2601.21557","evidence":"id","id":"arxiv:2601.21557","title":"Meta Context Engineering via Agentic Skill Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/bedaeecadf58846f4728c50c03396fc96ba18429ecf14052b4c5e177394a246e/"},{"source":"arxiv:2603.28052","target":"arxiv:2602.07755","evidence":"title","id":"arxiv:2602.07755","title":"Learning to Continually Learn via Meta-learning Agentic Memory Designs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4d0a0ea160495ea50583e0e24027b6085e2f430a44195e3053ba69cda8cc82ea/"},{"source":"arxiv:2603.28052","target":"doi:10.1038/s41586-023-06924-6","evidence":"title","id":"doi:10.1038/s41586-023-06924-6","title":"Mathematical discoveries from program search with large language models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/70a6b427d42fd28c9d538648e148a97db2fb2614c7764f3231e53f4bebdce887/"},{"source":"arxiv:2603.28052","target":"url:openai.com/index/harness-engineering","evidence":"url","id":"url:openai.com/index/harness-engineering","title":"Harness Engineering (OpenAI)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a64858f28b8f0c5dd1c3a4da212807c69004781880993ef54e5c383d4c9efbe5/"}],"incoming":[{"source":"arxiv:2604.25850","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2604.25850","title":"Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2358d2261b7478833ca59529278f0f4872896ca5ae4434e8db5d420532e10bd/"},{"source":"arxiv:2606.01770","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2606.01770","title":"Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e168e03e04c685d889bb8d0c68d63fc7aaa73099093ade681bebd7b2d1069bac/"},{"source":"arxiv:2607.07847","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2607.07847","title":"When Does Continual Learning Require Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/86e88c6cf21a66879eab5c156913fc85d14eb69c394c78c33c0e81d144e528eb/"},{"source":"arxiv:2609.01437","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2609.01437","title":"HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/0e91a157b5aa6d15aa804a3849a31a55661a73d6c1470ff918d72163f5fe2076/"},{"source":"arxiv:2609.14858","target":"arxiv:2603.28052","evidence":"id","id":"arxiv:2609.14858","title":"Dream-RSI: Recursive Self-Improvement through Evolving Worlds","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4c9ce361df4d9a5f89f3dc0a70a71e795eb66f2c4e3c262f73f04b2d117c8ea1/"},{"source":"url:lilianweng.github.io/posts/2026-07-04-harness","target":"arxiv:2603.28052","evidence":"id","id":"url:lilianweng.github.io/posts/2026-07-04-harness","title":"Harnesses for models (Lilian Weng, 2026)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ea249334ba128a9fff468e25fa4f8d13d4cccd13661ee130bec020ebb6463b/"}],"text_parts":[{"start":0,"end":15910,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html"},{"start":15910,"end":31723,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html"},{"start":31723,"end":47263,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html"},{"start":47263,"end":57614,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html"}],"contents":[{"title":"Meta-Harness: End-to-End Optimization of Model Harnesses","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"Abstract","offset":60,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"1 Introduction","offset":1490,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"2 Related Work","offset":5557,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"3 Meta-Harness: A Harness for Optimizing Harnesses","offset":9043,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"4 Experiments","offset":15574,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-1.html#text"},{"title":"4.1 Online Text Classification","offset":16349,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"4.2 Harnesses for Retrieval-Augmented Reasoning","offset":22216,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"4.3 Evaluating Agentic Coding Harnesses on TerminalBench-2","offset":25261,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"5 Discussion","offset":28438,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"Acknowledgements","offset":30092,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"Appendix A Qualitative Proposer Behavior","offset":30436,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"A.1 File Access Statistics","offset":30626,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"A.2 Qualitative Behavior: Causal Reasoning Over Prior Failures","offset":31251,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-2.html#text"},{"title":"Appendix B Discovered Harnesses","offset":35503,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"B.1 Text Classification Harness","offset":36154,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Overview.","offset":36669,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Meta-Harness (Draft Verification).","offset":37193,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Meta-Harness (Label-Primed Query).","offset":38568,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"B.2 Math Retrieval Harness","offset":39888,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Overview.","offset":40375,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"B.3 TerminalBench-2 Harness","offset":42210,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Per-task analysis.","offset":43469,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"Appendix C Dataset Details","offset":44274,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"C.1 OOD Text Classification Datasets","offset":44305,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"C.2 Math Retrieval Corpus","offset":47232,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-3.html#text"},{"title":"C.3 Math IMO-level Test Set","offset":48751,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"Appendix D Practical Implementation Tips","offset":49423,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"Appendix E Extended Related Work","offset":53655,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"AlphaEvolve / OpenEvolve.","offset":54076,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"GEPA.","offset":55043,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"Prompt orchestration frameworks.","offset":55981,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"},{"title":"Instructions for reporting errors","offset":56650,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/part-4.html#text"}],"text_code_points":57614,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/source.md"}