Metric & Benchmark
MATH
12,500-problem competition-math benchmark spanning algebra, geometry and number theory.
Definition
MATH (Hendrycks et al. 2021) draws from US math competitions (AMC/AIME). It is much harder than GSM8K and the standard reasoning-model benchmark — frontier reasoning systems now achieve 90%+, prompting moves to FrontierMath and Olympiad-level evaluations.
Common use cases
- Reasoning eval
- Reasoning-model comparison