{"id":"arxiv:2507.19457","title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","date":"2025-07-25","section":"Post-training","quality":0.450458,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/","source_url":"https://arxiv.org/abs/2507.19457","text_available":true,"authors":["Lakshya A Agrawal","Shangyin Tan","Dilara Soylu","Noah Ziems"],"outgoing":[{"source":"arxiv:2507.19457","target":"arxiv:2212.14024","evidence":"id","id":"arxiv:2212.14024","title":"Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/09c4c9bae4e016c9ce276cf84a46742b5b21168f4181202b35c207baf887f71e/"},{"source":"arxiv:2507.19457","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:2507.19457","target":"arxiv:2309.16797","evidence":"title","id":"arxiv:2309.16797","title":"Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/daac3bcfcd7f39084d995905e9ef96ae01c5aab7bfe8e565bd7622b4e44088f7/"},{"source":"arxiv:2507.19457","target":"arxiv:2402.03300","evidence":"id","id":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/636fb871904bc06f742cfdea069796f9e9665f95bd58dc63260eaff5154ef985/"},{"source":"arxiv:2507.19457","target":"arxiv:2402.16822","evidence":"title","id":"arxiv:2402.16822","title":"Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d4803e891a4b9c929eca007e3c3f003c330d2fc741109975c6ddd7d2c32828fc/"},{"source":"arxiv:2507.19457","target":"arxiv:2504.16084","evidence":"id","id":"arxiv:2504.16084","title":"TTRL: Test-Time Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ad54dd58d24afdfefc8aa5087d07c4aab2143e90ef06f6f7db8d3f3333452c48/"},{"source":"arxiv:2507.19457","target":"url:rlhfbook.com","evidence":"title","id":"url:rlhfbook.com","title":"Reinforcement Learning from Human Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/20e354559fad09ad4d6338d104c9c3f822a48e7fbf86cd8afa1ddf951b223658/"}],"incoming":[{"source":"arxiv:2601.21557","target":"arxiv:2507.19457","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:2507.19457","evidence":"id","id":"arxiv:2603.28052","title":"Meta-Harness: End-to-End Optimization of Model Harnesses","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8bb0b87384dfdd42c419cf26cc7e52f1f0d89f63afa5d5968e246b73a44c99cb/"},{"source":"arxiv:2604.19341","target":"arxiv:2507.19457","evidence":"id","id":"arxiv:2604.19341","title":"Structured Scaling of AI Discovery Across Diverse Scientific Domains","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/33ad3375247f86197210840a19064548b997e8617475e606d7a7ac2cda64740e/"},{"source":"arxiv:2604.25850","target":"arxiv:2507.19457","evidence":"title","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:2605.13821","target":"arxiv:2507.19457","evidence":"id","id":"arxiv:2605.13821","title":"Harnessing Agentic Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1f1060a507e2bf439d2934d610827d1b3177d85c90c1a2cab483608a56de5138/"},{"source":"arxiv:2606.01770","target":"arxiv:2507.19457","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:2507.19457","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:2607.13104","target":"arxiv:2507.19457","evidence":"title","id":"arxiv:2607.13104","title":"Self-Improvements in Modern Agentic Systems: A Survey","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1452caf7b16ffdb8c2efd5093f46dc4df0f89c3d45a5566cd5fdafd69385b2a7/"},{"source":"url:lilianweng.github.io/posts/2026-07-04-harness","target":"arxiv:2507.19457","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":15394,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html"},{"start":15394,"end":31335,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-2.html"},{"start":31335,"end":46955,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html"},{"start":46955,"end":56073,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html"}],"contents":[{"title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"Abstract","offset":75,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"1 Introduction","offset":1447,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"2 Problem Statement","offset":5690,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"Sample-Efficient Optimization.","offset":8610,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"3 GEPA: Reflective Prompt Evolution","offset":9747,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"3.1 Pareto-based candidate selection","offset":15352,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-1.html#text"},{"title":"4 Evaluation","offset":16741,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-2.html#text"},{"title":"5 Extended Applications of GEPA","offset":27607,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-2.html#text"},{"title":"5.1 GEPA For Inference-Time Search (Contd.)","offset":27643,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-2.html#text"},{"title":"5.2 GEPA for Adversarial Prompt Search (Contd.)","offset":31282,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-2.html#text"},{"title":"6 Related Work","offset":34101,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"7 Conclusion","offset":36950,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix A Appendix Outline","offset":37727,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix B Usage of Large Language Models","offset":38399,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix C GEPA’s Reflection and Prompt Update Meta Prompt","offset":38928,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix D GEPA Algorithm and Methodology Details","offset":39185,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"D.1 Merge: System-aware crossover strategy for Compound AI optimization","offset":39346,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix E Evaluation Setup (Contd.)","offset":40048,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"E.1 Benchmarks, Reference compound AI systems, and Feedback Functions","offset":40089,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"E.2 Models and Inference Parameters","offset":44994,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"E.3 Costs","offset":45855,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"E.4 Optimizers","offset":46069,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-3.html#text"},{"title":"Appendix F Results and Analysis (Contd.)","offset":52426,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix G Performance vs. Budget (Rollouts) Curves","offset":52584,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix H Generalization Gap","offset":52757,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix I Cost vs. Performance Analysis for optimized systems","offset":52872,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix J GEPA Search Trees","offset":53327,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix K Visualizing the Iterative Refinement achieved by GEPA","offset":53522,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"K.1 Prompts at intermediate stages for PUPA","offset":53778,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix L Examples of best prompts for every benchmark","offset":53827,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.1 HotpotQA, GPT-4.1 Mini","offset":54417,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.2 HotpotQA, Qwen3 8B","offset":54449,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.3 IFBench, GPT-4.1 Mini","offset":54477,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.4 IFBench, Qwen3 8B","offset":54508,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.5 HoVer, GPT-4.1 Mini","offset":54535,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.6 HoVer, Qwen3 8B","offset":54564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.7 PUPA, GPT-4.1 Mini","offset":54589,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"L.8 PUPA, Qwen3 8B","offset":54617,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix M GEPA generated prompts for kernel generation","offset":54641,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"M.1 NPUEval: Kernel Code Generation for new hardware architecture","offset":54701,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"M.2 KernelBench: CUDA Kernel Code Generation for NVIDIA GPUs","offset":54965,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Appendix N Number of reflection LM calls made by GEPA during optimization","offset":55031,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"},{"title":"Instructions for reporting errors","offset":55109,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/part-4.html#text"}],"text_code_points":56073,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c1ef72d96fa3292590a6a3d112ab509840081b59aaa771cf42b8b0e8ef62ba08/source.md"}