Given photos from several viewpoints, people can say roughly where things sit in space. Large , which read images and text together, cannot. Existing tests ask single questions, like which object is nearer, never checking whether the answers form one scene.
This gives a model several photos of one scene, indoor, tabletop or outdoor, names an object, and asks for a box giving its position, size and orientation in one shared frame. Separate scores cover layout and true scale.
Layout proved far stronger than scale: the best model scored 81.4 percent on structure against 58.7 percent on metric accuracy. Supplying correct geometry instead of the model's own map sharply raised navigation success, so failures came from misperceived space, not poor planning.