MindCube

MindCube asks a question that is easy for people and hard for machines: shown a scene from only a couple of angles, can a model work out where everything is — including the things currently out of frame — and how the layout would look if it moved? Built by Manling Li’s lab at Northwestern with collaborators at Stanford, NYU and the University of Washington, the benchmark pairs 3,268 images into 976 multi-view sets and asks more than 21,000 multiple-choice questions across three settings — Rotation, Around and Among — each demanding a different mental transformation of the scene.

The design deliberately withholds the convenient view. A model cannot answer “what is to the left of the white jar, from the viewpoint in image 4?” by reading a single picture; it has to fuse the available views into an internal spatial model and then query that model from a perspective it was never shown. That is what the authors mean by a “spatial mental model”, and it is a capability distinct from recognising objects or captioning an image.

MindCube is one of the vision evaluations aggregated on the Center for AI Safety’s AI Dashboard, which runs it as a standardized third-party test. On that dashboard the climb has been steep — from the high 40s for GPT-4o to the mid-80s by mid-2026 — but the frontier is now closely bunched, with Claude Opus 5 and Gemini 3.1 Pro separated by two-tenths of a point. Whether the remaining gap to the top reflects genuinely harder spatial reasoning or the benchmark approaching its ceiling is not yet settled.