DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning (github.com)

DeepSeekMath-V2 is a new LLM system that refocuses progress in mathematical reasoning from optimizing final-answer accuracy to building self-verifiable, stepwise proofs. The team trains a dedicated LLM-based verifier to judge the correctness and rigor of proofs, then uses that verifier as a reward model to train a proof generator via reinforcement-style optimization. Crucially, the generator is encouraged to self-detect and resolve flaws in its own derivations before submission, and the pipeline actively scales verification compute to automatically label hard-to-verify proofs—creating data to further strengthen the verifier and close the generation–verification gap. The result is a practical, iterative approach to faithful theorem proving: with scaled test-time compute DeepSeekMath-V2 achieves gold-level performance on IMO 2025 and CMO 2024 and 118/120 on Putnam 2024, and is evaluated on IMO-ProofBench. This demonstrates that verifier-in-the-loop training and self-debugging incentives can produce both strong and more trustworthy mathematical reasoning—important for tasks (like formal proofs and open research problems) where final answers alone are insufficient. While the authors note more work is needed, the paper suggests scalable self-verification is a promising route toward LLMs that produce rigorous, inspectable mathematical arguments.