DeepSeekMath V2

DeepSeek-Math-V2 targets self-verifiable mathematical reasoning instead of only chasing final answer accuracy. The authors first train a rigorous LLM-based proof verifier, then use it as the reward model in reinforcement learning so the proof generator is pushed to detect and repair issues in its own derivations before finalizing them. They further scale verification compute to label new hard-to-verify proofs and iteratively strengthen the verifier. This loop yields a model with strong theorem-proving performance on IMO-ProofBench and recent competitions such as IMO 2025, CMO 2024, and Putnam 2024, suggesting that scalable self-checking is a viable path toward more reliable deep mathematical reasoning systems.