Visual Spatial Reasoning

Visual spatial reasoning asks whether models can understand where things are, how they relate, and how scenes change across viewpoints or actions. Common approaches add explicit geometry such as depth, coordinates, bounding boxes, scene graphs, or maps, and combine these with structured reasoning, visual memory, tools, or imagined views. Benchmarks test static image relations and transformations, multi-view and video understanding, and embodied tasks such as navigation and manipulation. Models have improved with targeted data and training, but often rely on language priors rather than visual evidence and remain unreliable on cross-view, long-horizon, and action-grounded tasks.