Reproduce “TreeThink: A Modular Tree Search Library for Mathematical Reasoning with LLMs” 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
Burak S. Akbudak, Zeynel A. Uluşan, Can S. Erer, Gözde Gül Şahin
TREETHINK is an open-source Python framework that standardizes LLM-driven mathematical reasoning by separating tree search, language-model policies, node evaluation, and formal-environment interaction. Across miniF2F, it improves proof success over single-pass generation in Lean, Isabelle, and Rocq; asynchronous execution reaches an 8.08× speedup at concurrency 16, while rollout-free MCTS raises MATH500 accuracy from 23.0% to 40.0%.
Key insights
- Tree search improves over single-pass inference across all three tested proof assistants, but the best search strategy depends on the language: rollout-free MCTS leads in Lean and Rocq, whereas best-first search performs better in Isabelle.
- Asynchronous branch processing substantially increases throughput, producing up to an 8.08× speedup in the Isabelle timing study, although the highest concurrency causes a modest reduction in average successful solutions.
- The same modular architecture transfers to informal problem solving: rollout-free MCTS with a generative judge outperforms both single-pass inference and majority voting on MATH500.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.