Sparse cross-modal units that connect language concepts to specific scales of 3D geometry, routing semantic information to the most relevant level of detail.