SpatialViz-Bench measures spatial visualization in multimodal LLMs: the ability to mentally imagine and manipulate objects to infer unseen relationships. Its 1,180 multiple-choice questions are programmatically generated, which limits training contamination, and cover four sub-abilities — mental rotation, mental folding, visual penetration, and mental animation — each with three task types and controlled difficulty levels. Humans score about 82% overall, while current models remain far lower.

We source results from Table 2 of the SpatialViz-Bench paper, using the chain-of-thought prompting setting, which covers every evaluated model. Our chart reports overall accuracy, and the four sub-ability averages are kept in the data export.

Because there is no live leaderboard with the paper’s current numbers — the project page’s embedded leaderboard lags the paper — results are transcribed from the paper and updated when it is revised. The paper documents a “CoT paradox” in which chain-of-thought prompting helps most closed-source models but degrades many open-source ones, so with-CoT scores for some open models are lower than their direct-answer scores. The paper’s human, random, and text-only reference rows are excluded, and models we cannot match to a tracked model version are stored but not charted.