MindTopo finds agents lose topology in planning

MindTopo evaluates 14 multimodal LLMs across 11,030 topological-spatial instances and finds every model better at static reasoning than closed-loop planning; its best-performing model remains far below observed human performance. The MindTopo preprint covers continuity, separation, order, enclosure, and knots across 13 procedurally generated task types, with agents tested on both reasoning and environment action. Builders adding visual agents should treat plausible images as weak evidence of world-state understanding, the same evaluation discipline behind the archive’s warning about agent tests that cross scope.