Researchers introduce AdvancedMathBench, a benchmark suite designed to evaluate the advanced mathematical reasoning capabilities of large language models, addressing gaps in disciplinary scope and evaluation granularity found in existing benchmarks.
The suite includes ProverBench, containing 296 problems at undergraduate and doctoral qualifying-exam levels, and VerifierBench, which consists of 888 model-generated proof trajectories paired with expert ground truth to assess verification capabilities. The authors developed a dedicated automatic verification pipeline trained on large-scale expert annotations to provide correctness verdicts and fine-grained error assessments.
Experiments indicate that AdvancedMathBench remains challenging for frontier models; the best-performing model, GPT-5.5-xhigh, achieved only 75.8 and 66.1 on ProverBench splits, while the best verification model attained a Balanced F1 of only 65.1.