Large Vision-Language Models (LVLMs) often process hundreds or thousands of visual tokens, resulting in substantial computational and memory overhead during …
机构:南大
来源:arXiv 2611.04537 | AI4Papers 论文推荐平台