Particle.news

DeepSeek Releases DeepSeekMath‑V2 With Self‑Verifying Math Reasoning

Built-in LLM verification audits proofs, generating harder training samples to strengthen reasoning.

Overview

  • DeepSeek announced the math-focused model on November 27 as a training framework that emphasizes proof correctness over answer-only accuracy.
  • The team reports gold‑medal performance on IMO 2025 and CMO 2024 benchmarks and a 118/120 score on Putnam 2024.
  • An LLM-based verifier automatically reviews generated proofs and leverages expanded verification compute to produce more challenging samples.
  • The system is built on the DeepSeek‑V3.2‑Exp‑Base model.
  • Model weights and code are publicly available via Hugging Face and GitHub, with the company noting substantial work remains.