LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
LeanPRM uses Lean to auto-label 152K proof steps, lifting a Qwen3-4B verifier's step accuracy from 48.2 to 74.7.
LeanPRM generates step-level correctness labels for natural-language mathematical proofs by pairing them with Lean formal proofs and filtering syntax-only failures. The released set contains 10,634 proofs and 152,000 labeled steps without human annotation. A Qwen3-4B verifier trained on it reaches 74.7 exact step-label accuracy on LeanPRM-Bench, up from 48.2 and above 60.3 for the same base model trained on PRM800K. Gains appear across 0.6B–8B verifiers and on best-of-N IMO selection plus three of four external benchmarks.
- LeanPRM pairs natural-language proofs with Lean formal proofs to label steps automatically.
- Dataset has 10,634 proofs and 152K labeled steps with no human annotation.
- Qwen3-4B step accuracy rises from 48.2 to 74.7, versus 60.3 on PRM800K.
- Gains hold from 0.6B to 8B and on Hard2Verify and IMO-GradingBench.
From the original article
1 min read at huggingface.co
AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck. A learned verifier could check proofs cheaply, but training one requires a correctness label on every step, and such labels are hard to obtain for open-ended proofs.
By Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.