VSI-Bench

"Thinking in Space": a video visual-spatial-intelligence benchmark (5K+ QA) from NYU with Stanford (Fei-Fei Li) and Yale. VSI-Bench probes whether video MLLMs can build and reason over spatial mental models — distances, sizes, layouts, and egocentric perspective — and seeded the spatial-supersensing agenda behind Cambrian-S. It is now widely used to evaluate video MLLMs.