MindCube, created by researchers at Northwestern, Stanford, NYU, and the University of Washington, measures whether vision-language models can build spatial mental models from limited multi-view images: inferring positions, orientations, and how a scene would look after movement, including parts that are occluded or unseen. It contains 21,154 multiple-choice questions over 3,268 images in three settings — Rotation, Among, and Around — with template-generated questions and targeted distractors. Humans score about 95% on a diagnostic subset, while frontier models remain far lower.

We source results from the public MindCube leaderboard. Our chart reports overall accuracy across all questions, and the per-setting Rotation, Among, and Around accuracies are kept in the data export. The leaderboard’s synthetic random-baseline rows are excluded.

Questions are answered from a small set of views of a real scene, so models must reason about spatial relations they cannot directly observe. The overall score is weighted by question counts, and the Among setting contributes most of the questions. The leaderboard mixes frontier proprietary models with open-weight and specialized spatial models, many of which are small research models; entries we cannot match to a tracked model version are stored but not charted.