SpatialViz

SpatialViz-Bench isolates a specific human ability that image recognition does not require: mentally picturing a shape and transforming it in the mind’s eye. Built by researchers at the Chinese Academy of Sciences’ Institute of Automation with several university and industry collaborators, it splits spatial visualization into four sub-abilities — mental rotation, mental folding, visual penetration and mental animation — and generates 1,180 image questions across twelve tasks designed to test each.

The questions are produced programmatically rather than hand-collected, a deliberate choice: procedurally generated items are far less likely to have appeared in a model’s training data, so a high score is harder to attribute to memorisation. A typical task shows a stack of cubes and four candidate rotations and asks which one could not have been produced by rotating the original — a question a person answers by turning the shape over mentally, and one that a model cannot shortcut by pattern-matching against a caption.

SpatialViz is one of the vision evaluations aggregated on the Center for AI Safety’s AI Dashboard, which runs it as an independent third-party test. There the trajectory mirrors the other spatial benchmarks — from around 30% for GPT-4o, near chance on the harder sub-abilities, up to roughly 80% for GPT-5.6 Sol by mid-2026 — but the leading score still leaves the hardest mental-transformation tasks unsolved, and the authors frame the benchmark as a diagnostic of which specific spatial skills a model lacks rather than a single pass/fail number.