Reproduce “RSI-Anything: Unlocking Recursive Self-Improvement with World Models” 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
Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao, Wanli Ouyang, Pheng Ann Heng and 3 more
RSI-Anything uses predicted outcomes to select actions for testing, then pairs predictions with observed results to guide policy updates and conditional world-model adaptation. With the coder fixed on CodeContests, hidden-test all-pass success rose from 26.3% with a static verifier to 29.2% with fixed-model selection and 31.2% with adaptation, at similar online test usage. This does not establish sustained joint improvement.
Key insights
- Prediction-guided selection can improve the value of a limited evaluation budget: in CodeContests, success reached 29.2% with a fixed world model versus 26.3% with a capacity-matched static verifier, and 31.2% when the predictor was also adapted.
- System-level quality is not captured by a single metric: on MiniWoB++, an adapted configuration used fewer actions but had more incorrect submissions than the final configuration.
- The framework’s applications use different update pathways and domain-specific models; the results do not yet establish sustained, joint policy-and-model improvement across multiple rounds in one setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.