interpretabilityfactneutralBoth CKA and SVCCA progressively decrease throughout Vision Transformer training, indicating increasing representational specializationMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02386v1