Microsoft's MindTopo Benchmark Reveals VLMs' Severe Spatial Reasoning Flaws

Microsoft Research · rss · 2026-08-13

Microsoft Research introduced MindTopo, a new benchmark designed to evaluate the topological reasoning abilities of Vision-Language Models (VLMs), testing their understanding of structural relationships like connectivity, enclosure, order, separation, and knots.

A Massive Gap Between Static Recognition and Dynamic Planning

The study reveals that while current models perform adequately at static image recognition, they fail significantly at interactive planning tasks. Models can sometimes identify a path or a knot in a single frame, but their understanding breaks down when required to maintain or manipulate these relationships through a sequence of actions. Failures typically occur during the planning phase after the scene is understood, with models losing track of structural relationships or proposing physically invalid actions.

Generative Tools Fail to Solve Topology

The researchers also tested whether image and video generation tools could help models maintain topological understanding. Image generation occasionally helped for single frames, but video rollouts frequently altered topology or violated task dynamics.

The authors note that closing this gap is essential for robotics and interactive assistants, which may require models that carry an explicit topological state or world models that preserve topology by construction.

Original post →

More from Embodied

Embodied channel →