Do two rooms remain connected after a wall is added? Is an animal inside a fence? What separates a true knot from a tangled loop? Can several ropes be rearranged without letting them pass through one another?

These questions concern 3D topology: a form of spatial understanding based not on exact distances, angles or shapes, but on structural relationships that persist as objects bend, stretch or deform.

An unmeasured layer

Connectivity, enclosure, ordering and knottedness are examples of such topological properties. In cognitive science they are treated as a foundational layer of human spatial understanding — yet they remain largely absent from how multimodal AI systems are evaluated.

Most spatial evaluations for multimodal models focus on Euclidean properties: distance, direction, size and relative position. Microsoft Research's new benchmark, MindTopo, draws on Piaget and the wider cognitive literature's classification of topological ability to measure a different layer.

Two separate skills

The central design idea is to separate two things:

  • Recognition: can the model see a topological relationship in a static image — a connected path, an enclosed region, a knot?
  • Preservation and manipulation: can it maintain and alter those relationships through a sequence of actions?

The finding is clear: current multimodal models perform much better on static recognition than on interactive tasks. That suggests they struggle to maintain a consistent understanding of topology over time.

Where the failure happens

The most telling detail is which stage breaks. Failures emerge mostly in planning rather than perception.

Models lose track of structural relationships as scenes change, or propose actions that violate physical constraints. The model identifies the knot correctly at the start, then behaves while planning the steps to untie it as though it had forgotten what the knot was.

As the researchers put it, models can sometimes identify a connected path, an enclosed region or a knot in a single scene, but that understanding often breaks down once the model must manipulate the scene through a sequence of actions.

Why it matters

The finding points squarely at robotics and interactive environments. Understanding what stays connected, enclosed, ordered or knotted is a precondition for reliable decision-making there.

A robot arm tidying a cable without wrapping it around itself; a navigation system knowing which rooms remain reachable after a door closes — both require topological reasoning, and neither is solved by reading a static image correctly.

Why the benchmark itself is new

MindTopo's real contribution is not the weakness it finds but the design that makes that weakness visible. Asking about the same topological relationship both statically and interactively surfaces a gap that a single accuracy figure hides.

This is a methodological choice that keeps recurring: measuring not what a model does not know, but when it loses what it does know. The same distinction appeared in Google Research's knowledge-profiling work, where models held facts they could not recall.

The benchmark's limit is equally clear: MindTopo measures an ability without saying how to instil it. But knowing that the break happens during planning also says where to look for the fix.