1 paper · 1 filter
Weile Gong, Zijian Lu, Mingcai Chen +3
The paper investigates how vision-language models can become locked onto outdated textual priors, causing incorrect visual grounding, and identifies recurring latent directions—cal…