Multimodal MoE models fail visual reasoning via routing divergence

This paper highlights a fundamental architectural flaw in current multimodal MoE designs — perception overrides cognition instead of collaborating with it.

• –Routing divergence occurs in middle layers, preventing deeper cognitive processing
• –Models are "Seeing but Not Thinking" because perceptual signals hijack cognitive experts early
• –Findings suggest MoE architectures need explicit separation or staging between perception and reasoning layers
• –A critical read for AI researchers building next-gen multimodal foundation models