Kimina-Prover-RL: Open-source RL training pipeline for Lean 4 with AI-MO models achieving MiniF2F SOTA

AI Impact Summary

Kimina-Prover-RL releases an open-source training pipeline for Lean 4 theorem proving (kimina-prover-rl) built on Verl, featuring two new models that achieve state-of-the-art Pass@32 on MiniF2F for their sizes. It uses a two-stage reasoning-then-generation paradigm, GRPO reinforcement learning, and parallel Lean verification via kimina-lean-server, with kimina-client and a curated Kimina-Prover-Promptset drawn from NuminaMath-LEAN. This enables teams to reproduce experiments and tailor models, but deploying value requires GPU-heavy infra and careful integration with Verl/Lean verification workflows.