interpretabilityfactneutralEdit-related cross-modal responses in MLLMs concentrate in deeper semantic layersComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01978v1