irrelevant-context

This paper studies how irrelevant text context biases predictions in multimodal large language models, showing that context-induced decision margins follow an affine transformation of context-free margins, offering insights into model sensitivity.