When Seeing Overrides Knowing: Disentangling Knowledge Conflicts in Vision-Language Models
This work addresses reliability issues in vision-language models for AI applications, but it is incremental as it builds on existing conflict analysis methods.
The paper tackled the problem of knowledge conflicts in vision-language models, where internal parametric knowledge clashes with external visual information, leading to hallucinations. The authors introduced a dataset of multimodal counterfactual queries, localized specific attention heads controlling the conflict, and demonstrated that modifying these heads can steer the model's responses, with their method outperforming gradient-based attribution in precision.
Vision-language models (VLMs) increasingly leverage diverse knowledge sources to address complex tasks, often encountering conflicts between their internal parametric knowledge and external information. Knowledge conflicts can result in hallucinations and unreliable responses, but the mechanisms governing such interactions remain unknown. To address this gap, we analyze the mechanisms that VLMs use to resolve cross-modal conflicts by introducing a dataset of multimodal counterfactual queries that deliberately contradict internal commonsense knowledge. We localize with logit inspection a small set of heads that control the conflict. Moreover, by modifying these heads, we can steer the model towards its internal knowledge or the visual inputs. Finally, we show that attention from such heads pinpoints localized image regions driving visual overrides, outperforming gradient-based attribution in precision.