What Keeps Vision-Language Models Looking at the Image?

arXiv:2607.12815v3 Announce Type: replace Abstract: When do vision-language models need direct access to the image while generating an answer? We study image dependence during answer generation by examining how the visual information needed for the current que