Reproduce “EvoBelief: Falsifiable World Models for Self-Improving LLM Agents” 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

Janet Wang, Jiang Guo, Yi Xiang, Jiaqing Yuan, Mingwen Dong, Hideo Kobayashi, Shuaichen Chang, Chandan K. Reddy and 1 more

EvoBelief lets agents learn and revise testable claims about environment dynamics without changing model parameters. It gathers evidence before allowing validated rules to affect predictions and plans, and can narrow or retire rules after counterevidence. On 134 unseen ALFWorld tasks with GPT-5.4 and no actor-side transition memory, it achieves 76.12% success versus Frozen WM’s 44.78%; this compares complete systems.

Summary by alphaXiv.

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