Intuition vectors.

arXiv:2609.32619v1 Announce Type: cross Abstract: Large self-supervised vision models learn representations that support scene segmentation and the semantic decomposition of physical objects. We ask whether their representational geometry supports transfer to