Mixture of Layers: Dynamic Layer Routing for Visual Reasoning.

arXiv:2610.09440v2 Announce Type: replace-cross Abstract: Pre-trained vision encoders contain layer-wise visual representations that differ in spatial granularity, semantic abstraction, and sensitivity to local details. However, most Multimodal Large Language