MindTopo reveals VLMs’ spatial reasoning abilities
MindTopo finds multimodal models can often see topology in one frame, but lose it when they have to act over time.
The Microsoft Research benchmark tests continuity, separation, order, enclosure, and knots across static reasoning and interactive planning tasks. Models did better at recognizing connected paths, enclosed regions, or knots than at preserving those relationships through sequences of legal actions. The failures often appeared after perception, when models chose locally plausible moves, lost track over turns, or suggested actions that violated the simulated dynamics. Microsoft Research AI's note
The Microsoft Research benchmark tests continuity, separation, order, enclosure, and knots across static reasoning and interactive planning tasks. Models did better at recognizing connected paths, enclosed regions, or knots than at preserving those relationships through sequences of legal actions. The failures often appeared after perception, when models chose locally plausible moves, lost track over turns, or suggested actions that violated the simulated dynamics. Microsoft Research AI's note
score 5