MindTopo: VLMs see topology fine and lose it the moment they act.

Microsoft Research tested continuity, separation, order, enclosure, and knots in both static analysis and interactive planning. Static performance was consistently better than interactive, both well below human. The failure modes differ in a way that matters: static errors are perception failures, planning errors appear after the scene is understood correctly, as models lose track of relationships across multiple actions. Image and video generation tools helped little and often altered topology outright.