The benchmark combines visual perception with compositional mathematical reasoning. The initial study found a substantial gap between frontier multimodal models and human performance, while the public project provides dataset access, evaluation code and model-result tooling.
MathVista introduced 6,141 visual mathematical-reasoning examples drawn from 28 existing multimodal datasets and three newly created datasets, and used them to systematically evaluate 12 foundation models across mathematical and visual task types.
It broadened AI4Math evaluation beyond text-only word problems and formal statements, creating a reference point for later multimodal and Olympiad-level benchmarks that require models to reason over mathematical information embedded in images.
Significance rationale: MathVista made mathematical reasoning in visual contexts a dedicated evaluation target, consolidating diagrams, charts, geometry, scientific figures and other multimodal tasks into a large benchmark before multimodal mathematical evaluation became a standard capability axis.
—
The arXiv research record introduced MathVista and its initial evaluation of foundation models.
The public MathVista repository and project site expose benchmark data access and evaluation tooling. E3 records artifact-level inspectability rather than independent verification of every reported score.
No corrections recorded.
