Visual Anchoring Bias Found in Vision-Language Models
A recent investigation indicates that the presence of numeric anchors within images significantly skews quality assessments in Vision-Language Models. This study, available on arXiv, evaluated six VLMs across five different architectural families and discovered that the influence of anchors is 2.5 times greater than that of severe image quality deterioration, demonstrating that the bias cannot be attributed solely to visual alterations. Through layer-wise probing, a clear dissociation was observed: layers where anchor classification reaches saturation (L12-L34) are less effective for quality prediction, while optimal layers are situated deeper (R^2 = 0.69-0.91). Fusion analysis reveals integration varies by architecture, with two models showing immediate fusion at L1-L2, while three others exhibit partial or no fusion. These findings provide a causal explanation for visual anchoring bias, connecting behavioral vulnerability to representation dynamics.
Key facts
- Embedded numeric anchors on images bias Vision-Language Model quality judgments.
- Six VLMs from five architectural families were tested.
- ANOVA eta^2 = 0.18-0.77, all p < 0.001.
- Anchor effects are 2.5x larger than severe image quality degradation.
- Layer-wise probing reveals dissociation between anchor classification and quality prediction.
- Anchor classification saturates at layers L12-L34.
- Optimal layers for quality prediction are deeper (R^2 = 0.69-0.91).
- Fusion analysis shows architecture-dependent integration: instant fusion at L1-L2 in two models, partial or no fusion in three others.
Entities
Institutions
- arXiv