Kimina-Prover-RL: Open-source RL training pipeline for Lean 4 theorem proving (Verl-based)

AI Impact Summary

Kimina-Prover-RL introduces an open-source RL training pipeline for Lean 4 formal proofs, built as a Verl fork and fully compatible with Verl. It ships two open-source models AI-MO/Kimina-Prover-RL-1.7B and AI-MO/Kimina-Prover-RL-0.6B, achieving state-of-the-art Pass@32 within their size class using a two-stage reasoning-then-generation workflow and GRPO reinforcement learning, with multiple proof rollouts per prompt verified by kimina-lean-server. The stack includes kimina-lean-server and kimina-client for scalable parallel verification, plus Kimina-Prover-Promptset and NuminaMath-LEAN datasets filtered to emphasize challenging problems. This enables teams to reproduce, adapt, and accelerate development of Lean 4 theorem-proving capabilities, potentially reducing manual proof effort and speeding internal formalization projects.