{"id":"arxiv:2512.18552","title":"Toward Training Superintelligent Software Agents through Self-Play SWE-RL","date":"2025-12-21","section":"Agents, open-endedness, AGI","quality":-0.239509,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/","source_url":"https://arxiv.org/abs/2512.18552","text_available":true,"authors":["Yuxiang Wei","Zhiqing Sun","Emily McMilin","Jonas Gehring"],"outgoing":[{"source":"arxiv:2512.18552","target":"arxiv:2505.22954","evidence":"id","id":"arxiv:2505.22954","title":"Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c29f716871f80314868b5647692de387f26e12166b93bd362915fd272b6e081c/"},{"source":"arxiv:2512.18552","target":"arxiv:2506.13585","evidence":"id","id":"arxiv:2506.13585","title":"MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5beb2ec6c7e6362f9d957219d7a205dd1adcc53f739acc0c7fd5f1f2bded4d83/"},{"source":"arxiv:2512.18552","target":"arxiv:2510.13786","evidence":"title","id":"arxiv:2510.13786","title":"The Art of Scaling Reinforcement Learning Compute for LLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/72580bcba3d9b32b2f3e83d074f626234b26dc18126edd22a5eacef927cf87da/"}],"incoming":[{"source":"arxiv:2603.19461","target":"arxiv:2512.18552","evidence":"id","id":"arxiv:2603.19461","title":"Hyperagents","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/17eec8a15debc212b5530f117dd8ba8bad339590729cbbc4f75a47755ebcb544/"},{"source":"arxiv:2608.19197","target":"arxiv:2512.18552","evidence":"id","id":"arxiv:2608.19197","title":"SPADE: Self-Play in Adaptive Synthetic Executable Environments","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c524376d6034b5d79661d6f1a444591404f4a250bcb0790d802c05e02f8249df/"}],"text_parts":[{"start":0,"end":15980,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html"},{"start":15980,"end":30795,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html"},{"start":30795,"end":46654,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html"},{"start":46654,"end":56747,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html"}],"contents":[{"title":"Toward Training Superintelligent Software Agents through Self-Play SWE-RL","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Abstract","offset":77,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Keywords:","offset":1445,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"1 Introduction","offset":1519,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"2 Self-Play SWE-RL","offset":10167,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"2.1 Overview","offset":10190,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"2.2 Input Assumptions","offset":11640,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"2.3 Agentic Bug Injection","offset":12434,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Bug artifact.","offset":12465,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Complex bug generation.","offset":13975,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Consistency validation.","offset":15774,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-1.html#text"},{"title":"Higher-order bugs.","offset":17883,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Reward design.","offset":19077,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"2.4 Agentic Bug Repair","offset":20684,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Construct the buggy codebase.","offset":20712,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Initial prompt.","offset":21575,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Bug-repair process.","offset":22524,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Reward design.","offset":23890,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3 Evaluation","offset":24192,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3.1 Experimental Setup","offset":24209,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Base model.","offset":24237,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Evaluation setup.","offset":24753,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3.2 Baseline Comparison","offset":25289,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3.3 Ablation of Self-Play Components","offset":26602,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3.4 Ablation of Bug Injection Methods","offset":27858,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"3.5 Reward Ablation","offset":29340,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"4 Related Work","offset":30726,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Agent scaffolding for software engineering.","offset":30745,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-2.html#text"},{"title":"Training software agents via reinforcement learning.","offset":32078,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Data and environments for training software agents.","offset":33308,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"5 Conclusion","offset":35077,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Acknowledgement","offset":36027,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Impact Statement","offset":36633,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Appendix A Analysis of the Challenger-Solver Game","offset":37484,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"A.1 The Challenger Has an Optimal Target Solve Rate","offset":38447,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"A.2 The Challenger Strategy","offset":40608,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"The dominant challenger.","offset":40831,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"The tunnel-vision challenger.","offset":42756,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Improving natural language communication with humans.","offset":43914,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Mitigations.","offset":46027,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-3.html#text"},{"title":"Appendix B Opposing Incentives in the Reward Design","offset":47168,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Appendix C Experimental Details","offset":48249,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Training details.","offset":48285,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Evaluation noise.","offset":49575,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Appendix D Prompt Templates","offset":49881,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"D.1 Bug Injection","offset":49913,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"D.2 Bug Solving","offset":49936,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Appendix E Discussion","offset":50221,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"E.1 Theoretical Analysis","offset":50247,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"E.2 Limitations","offset":50768,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"E.3 Unsuccessful Attempts","offset":51995,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"E.4 Future Work","offset":53793,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"},{"title":"Instructions for reporting errors","offset":55783,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/part-4.html#text"}],"text_code_points":56747,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/da99b397e01b725186243c38a2dc6072d3a5c81bd8f759e898f04ad09aa0f4b1/source.md"}