Reproduce “MerLean-Prover: A Recursive Looping Harness for Lean 4 Theorem Proving” 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
Jinzheng Li, Zeru Zhu, Yuanjie Ren
A recursive looping harness, MERLEAN-PROVER, enables general-purpose large language models to function as an end-to-end Lean 4 theorem prover for complex formal mathematics. The system solves 10 out of 23 problems on FORMALQUALBENCH and all 12 problems on Putnam 2025, outperforming several existing open-source baselines by effectively mitigating LLM limitations without fine-tuning.
Key insights
- Effective harness design, even without complex fine-tuning or custom RL objectives, can significantly transform general-purpose LLMs into robust Lean provers for graduate-level formalization.
- Systematically addressing LLM weaknesses like context and constraint pressure through externalized global state and localized agent tasks improves long-horizon reasoning.
- The recursive looping architecture facilitates the decomposition of complex problems into smaller, verifiable sub-problems, enabling the system to manage intricate proof structures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.