What's this about?
A benchmark arguing spatial intelligence should be evaluated by making models generate images — marking locations, regions, and paths in the scene — rather than describing space in text (35 upvotes).
What exactly is "spatial cognition"?
Understanding where things are relative to each other and how to move among them — reading a room, not just naming its contents. Text answers let models pattern-match; drawing the answer in pixels leaves nowhere to hide.
Why does it matter?
Agents that act in physical or on-screen worlds need real spatial grounding; a model that can say "the mug is left of the laptop" but can't point at it will misclick and mis-grasp.