Vision-language models (VLMs) fail at some tasks that are simple for humans---but why? We hypothesize that many such failures stem from VLMs' difficulties tr…
机构:牛津
来源:arXiv 2610.25252 | AI4Papers 论文推荐平台