Kimina-Prover-RL: Open-Source Lean 4 Theorem Proving Pipeline

AI Impact Summary

Kimina-Prover-RL introduces a novel training pipeline for large language models focused on formal theorem proving in Lean 4. This pipeline leverages a reasoning-then-generation approach inspired by DeepSeek-R1, incorporating structured outputs and a reinforcement learning framework with GRPO and DrGPO to improve accuracy and robustness. The system’s key innovations include format checking, error correction, and a curated dataset, demonstrating a significant advancement in open-source theorem proving models.