Glossary · Term

MATH-500

← all terms

Definition

Plain language

A standard set of competition math problems used to test reasoning models.

As stated in the literature

A 500-problem subset of the MATH benchmark, drawn from competition mathematics across algebra, geometry, and number theory.

Also called: MATH

Why it matters: It's one of the standard benchmarks reasoning-focused models report on, so it shapes which capabilities researchers actively chase.

For example, a MATH-500 problem might ask for the smallest positive integer satisfying a particular modular equation, requiring multi-step algebraic reasoning.

Heard on the show

“The headline number is just over seventy-nine percent on MATH-500 — a standard competition-math benchmark — versus about fifty-four percent for the best comparable prior latent method.”
Episode 141 — How Two Tokens Reopened a Reasoning Method the Field Had Given Up On

Mentioned in 8 episodes

  1. 141
    How Two Tokens Reopened a Reasoning Method the Field Had Given Up On
  2. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  3. 074
    How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
  4. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
  5. 036
    Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.
  6. 026
    What RL Actually Does to Language Models, at the Token Level
  7. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means
  8. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't

Related concepts