Hugging Face has released Kimina-Prover-72B, a new state-of-the-art theorem proving model based on Qwen2.5-72B and trained with the Kimi k1.5 RL pipeline. The model introduces a Test-Time Reinforcement Learning Search framework that allows it to recursively discover and combine lemmas for complex proofs, and an error-fixing capability that interprets Lean error messages to propose corrections. On the miniF2F benchmark, Kimina-Prover achieved a 92.2% pass rate, surpassing previous methods. AI
IMPACT Sets a new state-of-the-art for formal reasoning benchmarks, potentially accelerating AI's role in mathematical and scientific discovery.
RANK_REASON New model release by a major AI lab (Hugging Face) with novel capabilities and benchmark results. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- AlphaProof
- DeepSeek-Prover-V2
- DSP+
- Google DeepMind
- Hugging Face
- International Mathematical Olympiad
- Kimi k1.5
- Kimina-Prover-72B
- Kimina-Prover-Distill-8B
- Lean 4 Programming Language
- miniF2F
- Qwen2.5-72B
- Qwen3 1.7B
- Qwen3-8B
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →