This is a model card, not a transcript. I will summarize it as requested.

Abstract:

The document introduces DeepSeek-Math-V2, a large language model designed for advanced mathematical reasoning. The core idea is to move beyond simply rewarding correct final answers towards "self-verifiable mathematical reasoning." This involves training a verifier model to assess the rigor of step-by-step proofs and then using this verifier as a reward signal for a proof-generating model. The system is designed to create a feedback loop where the verifier is improved by labeling difficult proofs generated by the increasingly capable generator. The authors claim this approach results in strong performance, citing gold-level scores on IMO 2025 and CMO 2024 benchmarks, and a near-perfect score on the Putnam 2024 competition.

DeepSeekMath-V2: A Model Focused on Verifiable Reasoning

• Core Concept: Self-Verifiable Reasoning: The model's training process emphasizes correct, step-by-step reasoning over just achieving the correct final answer. This is aimed at tasks like theorem proving where the derivation is critical.
• Generator-Verifier Architecture: A verifier LLM is trained to evaluate the quality of proofs. This verifier then acts as a reward model to train a separate generator model, incentivizing it to produce rigorous and logically sound proofs.
• Claimed Performance: The model reportedly achieves high scores on several advanced math competitions:
  • Gold-level scores on IMO 2025 and CMO 2024.
  • A near-perfect 118/120 on Putnam 2024. As far as I can tell, these are self-reported claims, and this document doesn't link to independent verification.
• Technical Foundation: The model is built on DeepSeek-V3.2-Exp-Base and is licensed under Apache 2.0.