DeepSeekMath-V2 aims to enhance mathematical reasoning in large language models (LLMs) by introducing a self-verification mechanism. Traditional approaches have improved LLMs' performance in quantitative reasoning, allowing them to excel in competitions like AIME and HMMT. However, achieving a correct final answer doesn't ensure the reasoning process is sound. Many mathematical tasks, particularly theorem proving, require detailed step-by-step reasoning rather than just numerical outcomes. Recognizing this gap, the authors propose a method to train a verifier that assesses the quality of mathematical proofs, thereby pushing the boundaries of deep reasoning. The model operates by using the verifier as a reward system for a proof generator, encouraging it to identify and rectify issues in its proofs before submission. To keep the verifier challenged as the generator improves, the authors suggest scaling the verification process to label difficult proofs, which then serve as training data for the verifier. In practice, DeepSeekMath-V2 has shown impressive results, achieving gold-level scores in IMO 2025 and CMO 2024 competitions, along with a near-perfect score of 118 out of 120 in the Putnam 2024 exam. These outcomes indicate that self-verifiable mathematical reasoning is a promising direction for developing more capable AI systems in mathematics.