Glossary · Term

DeepSeekMath

← all terms

Definition

Plain language

DeepSeek's family of math-focused language models, including a version that grades full mathematical proofs.

As stated in the literature

DeepSeek's math-specialized model series, including DeepSeekMath-V2, used as a proof-grading judge in olympiad-level RL training.

Also called: DeepSeekMath-V2

Why it matters: Reliable proof graders are the bottleneck for training models on olympiad-level math, since otherwise the RL signal is too noisy to learn from.

For example, DeepSeekMath-V2 reads a multi-step proof and outputs a structured grade that an RL trainer uses as the reward signal.

Heard on the show

“Now, the method they build this on is GRPO — that's the critic-free approach from the DeepSeekMath folks.”
Episode 160 — Training an AI to Take Its Own Notes, So Its Future Self Works Better

Mentioned in 4 episodes

  1. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
  2. 099
    How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes
  3. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
  4. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't

Related terms