AI BriefingTheorem-Proving SOTA Model Kimina-Prover Released
Key point
The Numina and Kimi teams released Kimina-Prover, which applies test-time RL search to achieve SOTA on the miniF2F benchmark.
Details
The Numina and Kimi teams released a new series of theorem-proving models including Kimina-Prover-72B. Kimina-Prover-72B is based on Qwen2.5-72B, and distilled models in 8B and 1.7B sizes were also released together.
The key technical innovations are as follows:
- Test-Time Reinforcement Learning (TTRL) Search: Introduced an agent-based framework that allows the model to autonomously discover and combine multiple intermediate lemmas for complex proofs.
- Error-Correction Ability: By interpreting Lean's error messages and proposing targeted fixes, it efficiently corrects errors instead of regenerating the proof from scratch.
Through these techniques, Kimina-Prover recorded a 92.2% pass rate on the miniF2F-test benchmark, surpassing DeepSeek-Prover-V2 and DSP+ to achieve new SOTA (State-of-the-Art) performance.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.