Researchers reveal a blind spot in AI's understanding of how objects stay connected and tangled

Microsoft researchers have unveiled a gap in how artificial intelligence grasps a seemingly simple but crucial aspect of spatial reasoning: whether objects remain connected, enclosed, or knotted as they move and shift.

The team created MindTopo, a new test suite designed to probe whether multimodal AI models truly understand topology—the branch of geometry concerned with structural relationships that hold their shape even when objects bend or stretch. Questions like whether two points remain linked in a maze, whether an animal stays inside a fence, or whether a rope is genuinely knotted test this capability.

Reasoning versus acting: where AI falters#

The research uncovered a striking disparity in how current models handle topology. When presented with a static image and asked to identify a connected path, enclosed boundary, or interlocked rope, these systems often succeed. But the moment a model must plan a series of actions—rotating pipes, rearranging blocks, or untangling ropes—performance plummets dramatically.

The benchmark evaluates five core topological concepts:

• Continuity: whether a path or object remains unbroken
• Separation: whether nearby elements form a single structure or multiple distinct ones
• Order: how elements are arranged along a path through a transformation
• Enclosure: whether a boundary creates a clear inside and outside
• Knots: whether ropes are truly knotted rather than just appearing tangled

Each concept is tested both in recognition mode, where a model views a scene and answers a question, and in planning mode, where it must interact with a simulated environment to achieve a goal.

Why perception isn't enough#

Analysis of model failures reveals the core problem: mistakes during static reasoning typically stem from visual perception—a missed wall or opening—while errors in planning happen after the scene is understood. Models tend to make locally sensible moves without tracking downstream consequences, or they propose actions that violate the environment's physical rules.

When researchers tested whether image or video generation could help models maintain topological understanding across sequences of changes, results proved mixed. Video rollouts frequently corrupted the topology they were meant to preserve, suggesting that mere visual prediction isn't sufficient to sustain structural awareness over time.