Vision-language models (VLMs) are asked to report what they see: to count objects, box them, parse a screen or pick a click target (GUI grounding). We argue …
机构:UC Berkeley
来源:arXiv 2610.25503 | AI4Papers 论文推荐平台