Reproduce “Reap: Uncovering Monte Carlo proof search with a lightweight model” 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
Yutong Wang, Xuanzhi Ren, Tianyi Xu, Ruitong Liu, Jingda Xu, Peihao Wu, Bryan Dai, Bin Dong
Reap puts Lean proof search inside Lean, using model-guided Monte Carlo tree search over native proof states and replaying candidate proofs for verification. The shared search loop supports inference and reinforcement-learning rollouts while recording model calls and tactic execution. Reaper-1, with a 1.7-billion-parameter policy–value model and separate premise selector, achieves 79.9% pass@32 on miniF2F-test; comparisons are not compute-matched.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.