Reproduce “LeanTree: Accelerating White-Box Proof Search with Factorized States in Lean 4” 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
Matěj Kripner, Michal Šustr, Milan Straka
LeanTree introduces an open-source tool and dataset for white-box automated theorem proving in Lean 4, enabling the factorization of complex proof states and robust programmatic interaction. This framework facilitates improved proof search, achieving an 18.36% success rate on MiniF2F with Llemma-7B, outperforming black-box methods.
Key insights
- Factorizing complex proof states into independent goals simplifies evaluation, reduces LLM context, enables parallel search, and allows for efficient state reuse in ATP.
- Providing explicit intermediate proof state information to language models during proof search significantly improves their proof-finding capabilities compared to black-box generation methods.
- A refined incremental verification strategy, focusing on new assignments and preventing `sorry`-like tactics, is essential for ensuring the correctness and integrity of generated proofs in white-box systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.