New Benchmark for Diagrammatic Reasoning in Olympiad Geometry
August 20, 2026
A new open-source benchmark consisting of 954 olympiad geometry problems isolates the ability to construct accurate diagrams from pure mathematical reasoning. The dataset includes a 297-problem hard subset paired with high-fidelity, human-authored diagrams to measure geometric construction capabilities.
HOW THIS AFFECTS YOU
●
researcherUse this benchmark to distinguish between a model's ability to solve math problems versus its ability to perform spatial and geometric construction.