AI and the World's Hard Problems · Mathematics · arXiv · 2026-10-19
LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
Step-by-step verification of long AI-written proofs needs per-step correctness labels, which are costly to get from humans. LeanPRM uses the Lean proof assistant to label each step of natural-language proofs, with no human annotation. A Qwen3-4B verifier trained on it raised step-label accuracy on LeanPRM-Bench from 48.2 to 74.7, above the 60.3 obtained when training on PRM800K.