{"id":"arxiv:2505.22954","title":"Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents","date":"2025-05-29","section":"Agents, open-endedness, AGI","quality":0.058379,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/","source_url":"https://arxiv.org/abs/2505.22954","text_available":true,"authors":["Jenny Zhang","Shengran Hu","Cong Lu"],"outgoing":[{"source":"arxiv:2505.22954","target":"arxiv:1901.01753","evidence":"id","id":"arxiv:1901.01753","title":"Paired Open-Ended Trailblazer (POET): Endlessly Generating Increasingly Complex and Diverse Learning Environments and Their Solutions","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a9790a1ede488fbbb2d7b4cdb97b884a5a1e21357b72c5e3419a83ea629d5a4b/"},{"source":"arxiv:2505.22954","target":"arxiv:1905.10985","evidence":"id","id":"arxiv:1905.10985","title":"AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/089142f7297bed414d01b82c78dc888e4018fecff7efbe5cae66442e03f87c4b/"},{"source":"arxiv:2505.22954","target":"arxiv:2004.12919","evidence":"title","id":"arxiv:2004.12919","title":"First return, then explore","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3cd790c501eaa633193918533947c58e30d1939bc08577606f1d88c56d368285/"},{"source":"arxiv:2505.22954","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:2505.22954","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:2505.22954","target":"arxiv:2312.00276","evidence":"title","id":"arxiv:2312.00276","title":"Metalearning Continual Learning Algorithms","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5d7361f404c0455d9f759fc6a8dd7bd1da75146e72717b7a2fbcdb8bac80412a/"},{"source":"arxiv:2505.22954","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:2505.22954","target":"arxiv:2402.16823","evidence":"title","id":"arxiv:2402.16823","title":"Language Agents as Optimizable Graphs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/750bd3807bf9c9a9b5cd2d60cc14387431aec3fe8a3f942916ca5d8922b2a5b0/"},{"source":"arxiv:2505.22954","target":"arxiv:2406.04268","evidence":"id","id":"arxiv:2406.04268","title":"Open-Endedness is Essential for Artificial Superhuman Intelligence","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/13964b2db9a9f793d1fcae07ca83fbe09f65454214269cac56777cb8be423a4e/"},{"source":"arxiv:2505.22954","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:2505.22954","target":"arxiv:2410.04444","evidence":"id","id":"arxiv:2410.04444","title":"Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d04ae5d37f10547be79c728efed48a629e631ab5384155564e52558e12ff4362/"},{"source":"arxiv:2505.22954","target":"arxiv:2501.19393","evidence":"id","id":"arxiv:2501.19393","title":"s1: Simple test-time scaling","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2a88818f7605b71cd0160831faebd51b3274c24ffd29626b4d519390b4bffffe/"},{"source":"arxiv:2505.22954","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/"}],"incoming":[{"source":"arxiv:2507.18074","target":"arxiv:2505.22954","evidence":"title","id":"arxiv:2507.18074","title":"AlphaGo Moment for Model Architecture Discovery","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1c22da54910a26f1f81b276f5cae3a81e22f7db024f7d1a5fa89ccd4f291ffa8/"},{"source":"arxiv:2512.18552","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2512.18552","title":"Toward Training Superintelligent Software Agents through Self-Play SWE-RL","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/"},{"source":"arxiv:2512.18746","target":"arxiv:2505.22954","evidence":"title","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:2601.21557","target":"arxiv:2505.22954","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:2602.07755","target":"arxiv:2505.22954","evidence":"id","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:2603.19461","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2603.19461","title":"Hyperagents","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/17eec8a15debc212b5530f117dd8ba8bad339590729cbbc4f75a47755ebcb544/"},{"source":"arxiv:2605.13821","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2605.13821","title":"Harnessing Agentic Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1f1060a507e2bf439d2934d610827d1b3177d85c90c1a2cab483608a56de5138/"},{"source":"arxiv:2607.07847","target":"arxiv:2505.22954","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:2505.22954","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":"arxiv:2608.08311","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2608.08311","title":"Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/4b8f53ea59f810bf1717d78f1907b7596b91df2bec6751aba12accc8598c78e9/"},{"source":"arxiv:2608.23875","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2608.23875","title":"AI Finds A Way","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/afe34dd5ee58c2e9739f9651349c4c9180d95dabff51994a13339f48dfd0bb27/"},{"source":"arxiv:2609.14858","target":"arxiv:2505.22954","evidence":"title","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:2505.22954","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":15251,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html"},{"start":15251,"end":30921,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html"},{"start":30921,"end":46835,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html"},{"start":46835,"end":62835,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html"},{"start":62835,"end":78230,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html"},{"start":78230,"end":80220,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-6.html"}],"contents":[{"title":"Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html#text"},{"title":"Abstract","offset":71,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html#text"},{"title":"1 Introduction","offset":2450,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html#text"},{"title":"2 Related Work","offset":8390,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html#text"},{"title":"3 Darwin Gödel Machine","offset":13990,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-1.html#text"},{"title":"4 Experiments","offset":18694,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html#text"},{"title":"4.1 Experiment Setup","offset":19441,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html#text"},{"title":"4.2 Benchmarks","offset":20694,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html#text"},{"title":"4.3 Baselines","offset":24727,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html#text"},{"title":"4.4 Results","offset":26827,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-2.html#text"},{"title":"5 Safety Discussion","offset":32588,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"6 Conclusion and Limitations","offset":36605,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Ethics Statement","offset":40036,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Reproducibility Statement","offset":43508,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Acknowledgments","offset":44371,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Appendix","offset":45121,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Table of Contents","offset":45134,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"Appendix A Additional Results","offset":45156,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"A.1 Baselines on SWE-bench","offset":45190,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"A.2 Generality across models on Polyglot","offset":45222,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"A.3 Ablation of Parent Selection","offset":45974,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-3.html#text"},{"title":"A.4 Additional Statistics of DGM runs","offset":46977,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"Appendix B Additional Related Work","offset":48093,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"Appendix C Algorithmic Details","offset":52414,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"C.1 Initial Coding Agent","offset":52449,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"C.2 Parent Selection","offset":52702,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"C.3 Self-Improve Prompts","offset":55014,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"C.4 Pseudocode","offset":56782,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"Appendix D Experiment Details","offset":57078,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"D.1 Hyperparameters for Foundation Models","offset":57112,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"Appendix E Benchmark Details","offset":57577,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"E.1 Cost Estimate","offset":57610,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"E.2 SWE-bench Tasks","offset":59654,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-4.html#text"},{"title":"E.3 Polyglot Tasks","offset":64827,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"E.4 SWE-bench State-of-The-Art","offset":66207,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"E.5 Polyglot Representative Agent","offset":66721,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Appendix F Best-Discovered Agents","offset":67613,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"F.1 DGM on SWE-bench","offset":67651,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"F.2 DGM on Polyglot","offset":67760,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Appendix G Similar Target Functionality, Different Implementations","offset":67867,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Appendix H Case Study: Solving Hallucination","offset":68956,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Appendix I Additional Safety Discussion","offset":73765,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Appendix J Additional Future Work Directions","offset":74905,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-5.html#text"},{"title":"Instructions for reporting errors","offset":79256,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/part-6.html#text"}],"text_code_points":80220,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/source.md"}