When a vision-language model sees a picture that contradicts its accompanying text, it often resolves the conflict by favoring one modality. Earlier research described this as a fixed "modality bias," such as a blind faith in text. But that characterization leaves a puzzle: why do preferences shift across different models and different kinds of content?

A new arXiv paper proposes a more unified explanation. Rather than treating text and image preferences as arbitrary biases, the authors model the integration of conflicting evidence as a Bayesian inference process. In this view, the model behaves like an observer who starts with prior beliefs about the reliability of each modality and updates those beliefs as evidence arrives. The observed "bias" is the outcome of those priors, not a simple preference for one channel.

This framing could help researchers predict when a model will trust text over an image, or vice versa, based on the statistical structure of its training data. The abstract does not provide experimental details, but the theoretical shift is significant: it moves the discussion from descriptive labels to a mechanistic account of how multimodal models reason under uncertainty.