Mathematical Reasoning Benchmarks

These benchmarks measure AI problem solving from arithmetic word problems and Olympiad questions to graduate-level proofs, formal theorem proving, and diagram-based geometry. Evaluation approaches include answer accuracy, problem perturbations, human or model-based proof grading, and symbolic or Lean verification, sometimes with tools or multi-step agent workflows. A recurring finding is that strong scores on familiar problems may not transfer to equivalent reformulations, new problem distributions, other languages, or complete proof construction. The field still faces tensions around benchmark contamination, reliable grading, and how to compare systems that use different amounts of test-time computation.