DeepSeek releases DeepSeek-Prover-V2 with recursive proof search and new ProverBench benchmark

DeepSeek AI announced DeepSeek-Prover-V2, an open-source large language model for formal theorem proving in Lean 4. The model employs a recursive theorem-proving pipeline that uses DeepSeek-V3 to generate high-quality initialization data by decomposing complex theorems into subgoals and formalizing them. A 7B parameter model then proves these subgoals, and the complete proofs are paired with DeepSeek-V3's chain-of-thought reasoning for training. Reinforcement learning with binary feedback further refines the model. The 671B parameter version achieves 88.9% on MiniF2F-test and solves 49 out of 658 PutnamBench problems. DeepSeek also introduced ProverBench, a dataset of 325 problems from AIME competitions and textbooks for evaluating mathematical reasoning. The model is available in two sizes: 7B and 671B parameters.