Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models
For researchers and developers of multimodal AI systems, this provides a mechanistic understanding of how VLMs handle perception-knowledge conflicts, revealing an asymmetric causal structure that could inform model debugging and reliability improvements.
The paper identifies a sparse causal circuit in vision-language models that resolves conflicts between visual evidence and memorized world knowledge, finding that visual grounding is the default while prior grounding depends on a small set of attention heads (2.5-4.8%) in the second half of the network. Ablating these heads flips predictions from knowledge-grounded to visually grounded in 68-96% of cases under prior-knowledge prompts.
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 combine activation patching across three granularities (residual stream, attention heads, and MLP sublayers) with model-component ablation studies and mechanistic analysis. Across three VLM families, we find that visual grounding emerges by default, whereas prior grounding depends on a small set of causally necessary attention heads (2.5-4.8%) concentrated in the second half of the network. These heads enable answers from stored world knowledge (e.g., "red" for a strawberry) despite conflicting visual input. Ablating them flips predictions from knowledge-grounded to visually grounded answers in 68-96% of cases under prior-knowledge prompts, but changes only 0.8-7.5% of visually grounded predictions, establishing an asymmetric causal structure. The identified heads decompose into routing heads, which modulate information flow, and writing heads, which directly project answer tokens into the residual stream. This structure is consistent across model families and scales, revealing a sparse causal circuit underlying perception-knowledge conflict in VLMs.