Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning
FreeLarge formal reasoning model for theorem proving in Lean 4
About Kimina-Prover Preview: Towards Large Formal Reasoning Models with Reinforcement Learning
Kimina-Prover Preview is a large formal reasoning model developed by MoonshotAI that proves mathematical theorems in Lean 4 with human-like reasoning. It achieves state-of-the-art performance, reaching an 80%+ pass rate on the miniF2F benchmark, outperforming prior works like BFS-Prover and Hunyuan-Prover. The model uses whole-proof generation enhanced by reinforcement learning, scaling to 72B parameters with a context window of 32K tokens, the longest in neural theorem proving. It introduces a 'Formal Reasoning Pattern' that bridges formal verification with informal mathematical intuition. The project is open source, releasing distilled models, an autoformalization model, a rectified miniF2F-test dataset, and the Kimina Lean Server.
Key Features
Pros & Cons
- Achieves 80%+ pass rate on miniF2F, surpassing all prior published results
- High sample efficiency with strong performance even at low sample budgets (e.g., pass@8 of 65.16%)
- Open source: releases distilled models, autoformalization model, proofs, and Lean server
- Uses simple RL without complex techniques like Monte Carlo tree search or process reward models
- Long context window of 32K tokens, longest in neural theorem proving
- Requires significant computational resources due to 72B model size
- Currently only supports theorem proving in Lean 4 language
- Limited to formal reasoning; may not generalize to informal mathematical problems