DeepSeek Unveils DeepSeek-Prover-V2 for Advanced Neural Theorem Proving
DeepSeek AI has announced the release of DeepSeek-Prover-V2, an open-source large language model designed for formal theorem proving in the Lean 4 environment. This new model introduces a recursive proof search pipeline and utilizes DeepSeek-V3 to generate high-quality synthetic training data through a cold-start procedure. By decomposing complex theorems into subgoals and combining informal reasoning with formal proofs, the model undergoes reinforcement learning to enhance its mathematical capabilities. The flagship 671-billion parameter version achieved state-of-the-art performance, securing an 88.9% pass ratio on the MiniF2F-test benchmark and solving 49 problems from PutnamBench. Alongside the model, DeepSeek introduced ProverBench, a new evaluation dataset comprising 325 problems from AIME competitions and educational textbooks. Available in both 7B and 671B parameter sizes, DeepSeek-Prover-V2 aims to bridge the gap between intuitive mathematical reasoning and rigorous formal verification, marking a significant advancement in automated theorem proving technology.
Wire timeline
DeepSeek Unveils DeepSeek-Prover-V2 for Advanced Neural Theorem Proving
DeepSeek AI has announced the release of DeepSeek-Prover-V2, an open-source large language model designed for formal theorem proving in the Lean 4 environment. This new model introduces a recursive proof search pipeline and utilizes DeepSeek-V3 to generate high-quality synthetic training data through a cold-start procedure. By decomposing complex theorems into subgoals and combining informal reasoning with formal proofs, the model undergoes reinforcement learning to enhance its mathematical capabilities. The flagship 671-billion parameter version achieved state-of-the-art performance, securing an 88.9% pass ratio on the MiniF2F-test benchmark and solving 49 problems from PutnamBench. Alongside the model, DeepSeek introduced ProverBench, a new evaluation dataset comprising 325 problems from AIME competitions and educational textbooks. Available in both 7B and 671B parameter sizes, DeepSeek-Prover-V2 aims to bridge the gap between intuitive mathematical reasoning and rigorous formal verification, marking a significant advancement in automated theorem proving technology.
Synced