Reproduce “daVinci-Prover: Agent-Native Mid-Training for Formal 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

Wenhong Zhu, Xuefeng Li, Jiashuo Zhang, Yu Chen, Tiantian Mi, Qipeng Guo, Pengfei Liu

daVinci-Prover trains models to prove theorems in Lean by combining mathematical skill-prior examples with trajectories that include tool use and verifier feedback. This mid-training teaches proof decomposition, tactic search, error repair, and interaction with Lean before supervised fine-tuning. Against a matched Qwen3-8B SFT-only baseline, it raises average pass@1 by 13.7 points across MiniF2F, FATE-M, and ProofNet.

Summary by alphaXiv.

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