AIME

AIME (American Invitational Mathematics Examination) is a challenging mathematics competition for top-performing high school students in the United States, serving as a qualifying step toward the International Mathematical Olympiad. Each exam consists of 15 problems requiring integer answers between 000 and 999, covering algebra, geometry, number theory, and combinatorics. The problems demand creative mathematical reasoning and precise calculation rather than rote knowledge.

The AI research community adopted AIME as a benchmark because its problems test genuine mathematical reasoning rather than pattern matching or memorization. The AIME 2025 benchmark uses 30 problems from the 2025 competition (15 per exam session) and has become a standard evaluation for measuring how well models handle multi-step mathematical problem-solving. Frontier models now routinely exceed human performance, with top models achieving above 85% accuracy compared to the median human competitor solving roughly 27-40% of problems.

A key concern with AIME as an AI benchmark is data contamination. Since problems and solutions are publicly available after each competition, models may have been exposed to them during pretraining. Researchers have noted that models tend to perform better on older questions compared to newer ones, suggesting some degree of memorization. Each year's new AIME problems provide a fresh, uncontaminated test set, which is why the benchmark is updated annually.