1 paper
Xinran Nicole Han, Yiyang Lu, Katelyn Gan +1
Most vision-language models (VLMs) treat the vision encoder as a static feature extractor: it encodes an image once, independently of the question, and leaves subsequent reasoning…