DeepSeek Launches DeepSeek-Prover-V2: Recursive Proof Search and New ProverBench Benchmark

DeepSeek AI has announced the release of DeepSeek-Prover-V2, an open-source large language model for formal theorem proving in the Lean 4 environment. The key innovation is a recursive proof search pipeline that uses DeepSeek-V3 to generate initial training data. The model was trained in two stages: first, a cold start where DeepSeek-V3 decomposes theorems into sub-goals and formalizes them, while a smaller 7B model checks the sub-tasks; then, a fine-tuning stage with reinforcement learning on synthetic data, where the signal is a binary correctness reward. The flagship version, DeepSeek-Prover-V2-671B (671 billion parameters), achieved 88.9% on MiniF2F-test and solved 49 of 658 problems from PutnamBench. Also introduced is the ProverBench benchmark, containing 325 problems of varying difficulty, including 15 formalized problems from AIME 2024 and 2025. Two model versions are available: 7B (based on DeepSeek-Prover-V1.5-Base, with a 32K token context) and 671B (based on DeepSeek-V3-Base).