Mathematics
MathVista MINI
MathVista MINI tests whether vision language models can answer mathematical questions about an accompanying image, some multiple choice and some free response, with an answer counted correct when it matches the ground truth.
1,000items
263subjects
CC-BY-SA-4.0license
mathematicsdomain
textmodality
imagemodality
item-level responses released
Saturation status: No
Response matrix
Rasch analysis p = σ(θ − z)
253,207 responses, 80/20 split over cells · 263 subjects · 1,000 items
AUC train
0.853
AUC test
0.849
Loading session strips…
Correct (1)Incorrect (0)Unobserved
Scale: 1 = correct · 0 = incorrect