Reproduce “LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification” 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

Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen, James Shiffer, Saadia Gabriel, Pavel Izmailov

LeanPRM pairs natural-language proof steps with Lean steps, uses Lean to label them, and trains a verifier that reads prose. On a held-out benchmark, Qwen3-4B trained for 500 updates achieves 74.7% exact step-label accuracy, versus 60.3% for PRM800K. LeanPRM outperforms on four of five benchmarks, but PRM800K scores higher on ProcessBench.

Summary by alphaXiv.

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