Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models

How vision-language models reconcile visual evidence with memorized world knowledge when they conflict matters because it shapes the reliability of multimodal systems yet prior work lacks a component-level causal account.

Vision-language models must reconcile visual evidence with memorized world knowledge when the two conflict. How they resolve this conflict shapes the reliability of multimodal systems, yet prior work characterizes it behaviorally without a component-level causal account. We…