MathArena evaluates models on recurring competitions soon after problems are released and records whether a model predates each competition.
The MathArena paper introduced a continuously updated benchmark built from newly released mathematics competitions, reported evidence of contamination in AIME 2024, and added human-graded proof-writing evaluation alongside final-answer competitions.
MathArena evaluates models on recurring competitions soon after problems are released and records whether a model predates each competition. Its public pipeline supports final-answer parsing as well as proof-based competitions such as USAMO and IMO, where expert graders assess model-generated proofs.
As static math benchmarks became increasingly vulnerable to contamination and saturation, MathArena supplied a reusable live-evaluation pattern that better separates current reasoning capability from exposure to old benchmark material and makes proof quality part of the evaluation target.
Significance rationale: MathArena changed mathematical-reasoning evaluation from repeatedly scoring static historical datasets toward evaluating models on newly released competitions close to their publication date, while also incorporating expert-graded proof-writing tasks.
—
The arXiv research record introduced MathArena and reported the initial live-competition and proof-writing evaluation results.
The public MathArena repository exposes evaluation code, competition configurations and result-inspection tooling; the work also appeared in the NeurIPS 2025 Datasets and Benchmarks Track. E3 records inspectable infrastructure and does not imply independent verification of every model score or contamination finding.
No corrections recorded.
