Reproduce “RSIGym: A Flexible Environment for Recursive Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fanqing Meng, Lingxiao Du, Haocheng Lu, Qiguang Chen, Ziqi Zhao, Zijian Wu, Jiayuan Zhuo, Mengkang Hu and 1 more
RSIGym provides reusable services for training, inference, rollouts, evaluation, and sandbox computation, while permissions let agents change data, training settings, harnesses, or combinations. In one budgeted run, six agents improved 29 of 30 settings across five benchmarks. Opus 5 achieved an RSI-Index of 0.4809 and raised SWE-bench Verified from 17.67% to 50.33%. The results do not establish progress across later cycles.
Key insights
- Harness changes can yield substantial gains even when model weights are fixed: one case improved Terminal-Bench from 30.34% to 40.82%, while increasing tasks solved on all three attempts from 15 to 26.
- Improvement is uneven across domains: the leading Joint-track agent raised SWE-bench Verified from 17.67% to 50.33% and AIME from 31.67% to 97.78%, but gains on Terminal-Bench and SkillsBench were more limited.
- More training or spending does not guarantee better performance; most of the ten reported within-run comparisons favored the less-trained reference, and doubling the budget produced mixed results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.