Benchmark
MathArena Apex
Test mathematical reasoning on a small set of competition problems selected for difficulty.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Jasper Dekoninck, Nikola Jovanović, Ivo Petrov and Martin Vechev introduced 12 problems from 2025 competitions, including six adapted from proof questions.
Methodology
Candidates were filtered using four attempts from each of four models; the original evaluation then gave nine models 16 attempts per problem and averaged success. Selection favors models not used in filtering. The small public set can become contaminated. MathArena now marks Apex deprecated; Apex Shortlist is a separate selection. No questions or scores are reproduced.