Reproduce “When LLMs Modify Themselves: Convergence, Values, and Stopping in Iterative 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

Oleg Stroganov

The authors test whether models can revise their own system prompts and decide when to stop without an external measure of improvement. Across six models and modes, stopping varied more by model than by mode. Claude Opus 4.5 declared convergence in all nine runs, while GPT-4o did so in one; this result is limited by small samples and provider defaults.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.