Kimina-Prover introduces Test-Time RL Search for Lean proofs with 92.2% miniF2F pass

AI Impact Summary

Kimina-Prover introduces a Test-Time Reinforcement Learning (TTRL) search that lets a Lean 4 prover recursively select and compose intermediate lemmas, plus an error-fixing mechanism that interprets Lean error messages to propose fixes. This yields significantly improved proof search efficiency and a new state-of-the-art pass rate (92.2% on miniF2F with Kimina-Prover-72B), indicating strong potential to automate formal proof workloads. For deployment, teams can leverage Kimina-Prover-72B or distilled variants (8B, 1.7B) to balance accuracy and cost, but expect substantial GPU inference budgets and careful integration with Lean pipelines and lemma sources. The approach reduces manual proof effort and accelerates formalization, but long-horizon proofs may require more sophisticated search strategies and validation to ensure reliability across diverse problem sets.