LeanPRM Trains Math Proof Verifiers on Lean-Generated Synthetic Data

Oxford researchers' LeanPRM uses the Lean proof assistant to auto-label 152,000 proof steps without human annotation, training a 4B verifier that finds errors in natural-language math proofs, lifting accuracy from 48.2% to 74.7%.

2026-11-07 ~ 2026-11-07 · 3 related posts

• LeanPRM: Lean-grounded step labels train a 4B verifier lifting accuracy 48.2→74.7 — Pavel_Izmailov · 2026-11-07
• New paper trains Lean-grounded verifiers to spot errors in natural-language math proofs — Pavel_Izmailov · 2026-11-07
• LeanPRM Trains Natural-Language Math Proof Verifiers on Lean-Generated Synthetic Data — Pavel_Izmailov · 2026-11-07