MindTopo reveals VLMs’ spatial reasoning abilities

Microsoft Research has put out MindTopo, a benchmark built to test a kind of spatial reasoning many multimodal models still don’t seem to have in hand. It looks at whether systems can deal with connectivity, enclosure, order, separation, and knots — not just in a still image, but while planning actions that change a scene.

That distinction matters. A model can sometimes point to a connected path or a rope that looks knotted. But MindTopo asks a harder question: can it keep that structure intact, or deliberately change it, over a sequence of moves? The benchmark separates those two skills on purpose, because the failures don’t always come from bad perception. Often, the model understands the scene at first and then loses the plot as the environment shifts.

The tasks are grouped into five categories: continuity, separation, order, enclosure, and knots. In one mode, a model answers questions about rendered scenes, such as whether two points in a maze are connected or whether sheep are inside a fence. In the other, it has to act inside a simulated environment, choosing moves that create, preserve, or remove a topological relation. The setup includes legal-action limits, so no cheating by, say, threading one rope through another in a way the environment would not allow.

Across a wide mix of proprietary and open-weight models, the results were better on static reasoning than on interactive planning, and both were still well below human performance. The failure modes were revealing. Static errors often started with perception: a missed wall, opening, or crossing. Planning errors came later, after the model had already seen what was going on, when it picked a move that looked reasonable in the moment but broke the structure down the line.

Microsoft also checked whether generation tools could help. Image generation sometimes helped when the key relation fit in one frame, but it wasn’t dependable once crossings or multiple moves entered the picture. Video rollouts were even shakier, often changing the topology or ignoring the task’s physical rules. So the real lesson here is not that models can sort of see space. It’s that they still struggle to keep space coherent while they act on it.