MathArena
evalMathArena is an uncontaminated, live evaluation of LLMs on recent competition mathematics (AIME, HMMT, USAMO, IMO and more): because it scores models on contests held after their training cutoff, it resists benchmark contamination and has become a standard reference in frontier-model math claims. The companion "Proof or Bluff? Evaluating LLMs on 2025 USAMO" (arXiv:2503.21934) grades full proofs, exposing large gaps between final-answer accuracy and rigorous reasoning.
From Martin Vechev's SRI Lab (minor joint→INSAIT); NeurIPS 2025 Datasets & Benchmarks.
Evaluation Details
Domains 2
Saturation low (rotates to fresh, post-cutoff contests)
Domains: competition mathematics, proof evaluation