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

Key points

• AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck.1
• A model writes a natural-language proof together with a paired formal proof, without feedback from Lean.1
• A Qwen3-4B verifier trained on LeanPRM raises exact step-label accuracy on a held-out set of 200 proofs (LeanPRM-Bench) from 48.2 to 74.7, as opposed to 60.3 when the same base model is trained on PRM800K.1
• Given the labels need no human annotation, LeanPRM can grow with the supply of formal statements, offering a path to step-level feedback on AI-generated proofs that are hard for humans to check.1

Sources (2)

• [1]LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 04:23 AM

  We introduce LeanPRM, a framework that uses the Lean proof assistant to create step-level labels for natural-language proofs.

  AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck.

• [2]LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
  Hugging Face Daily Papers · Oct 13, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.