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

Researchers introduced LeanPRM, a Lean-grounded process supervision technique designed for natural language proof verification. The system uses Lean to generate synthetic data and trains verifiers that read only natural language to identify mistakes in math proofs. According to the announcement, the verifiers outperformed models trained on human-labeled data across four out of five benchmarks. Open article →