Current frontier LLMs struggle significantly with advanced mathematics—even the best model only solves 66% of doctoral-level problems—and are particularly weak at detecting errors in proofs, suggesting proof verification is a critical bottleneck for mathematical reasoning.
AdvancedMathBench is a new benchmark for evaluating how well AI models can generate and verify advanced mathematical proofs at the undergraduate and doctoral level. It includes 296 proof problems and an automated verification system trained on expert annotations to assess not just whether answers are correct, but whether the reasoning steps are valid.