DeepSeek Launches Prover-V2: Open-Source LLM for Math Proofs

DeepSeek has introduced DeepSeek-Prover-V2, an open-source large language model designed for formal theorem proving using Lean 4, which extends the capabilities of DeepSeek-V3 with a recursive theorem proving pipeline. Available as a 7B model and a more advanced 671B model utilizing the mixture-of-experts architecture, it can handle up to 32K tokens for managing complex proofs. The model's recursive approach allows for decomposing complex theorems into subgoals for efficient solving, achieving an 88.9% pass rate on the MiniF2F-test benchmark and solving several problems from the PutnamBench and AIME competitions. DeepSeek has also developed ProverBench, a new benchmark with 325 formalized problems to assess theorem proving models, aiming to connect informal reasoning with formal proof construction. Despite its successes, concerns about potential misformalizations have been raised, necessitating rigorous testing and validation, while DeepSeek plans future model releases to advance mathematical reasoning further.