Evaluation
MATH500
2023ActiveUpdated: 22 September 2026Published
Key
innovation
A standard, compact (500-problem) math benchmark for fast evaluation.
Category
Evaluation
Abstraction level
Building block
Operation level
Evaluation (runtime)
Use cases
Math reasoning evaluationComparing reasoning modelsAssessing post-training methods
How it works
The model solves 500 math problems; the score is the fraction of final answers matching the reference solution.
Problem solved
The full MATH set is large and costly to evaluate; a smaller, representative subset was needed.