Microsoft researchers released the MindTopo benchmark to test AI spatial reasoning

Microsoft researchers published a new benchmark called MindTopo to measure how well multimodal models handle topological reasoning. The system evaluates five key areas: continuity, separation, ordering, enclosure, and knots. These concepts represent how objects relate to each other regardless of shape or distance.

Testing involved both reasoning tasks and interactive planning challenges. Models had to answer questions about static scenes or perform actions like untangling ropes. Researchers found that models often struggled once they had to manipulate objects through a sequence of steps rather than just identifying a relationship in a single image.

Failures occurred when models lost track of how structures changed during a task. Even when a model correctly identified a knot or a connection at the start, it frequently violated physical constraints while attempting to interact with the environment. Performance across all tested models remained significantly below human capabilities.