infrastructurefactneutralLayers in LLMs contribute unequally to representation transformation, which can be characterized by cosine similarity between layer input and output hidden statesMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21291v1