Not All Visual Tokens Are Equally Safe to Remove:Consequence-Sensitive Visual Token Compression

作者：Jingbo Wen, Liang He, Mingyu Cao, Haoyu Wang, Minxuan Hu, Kangning Cui, Xilu Wang · 单位：The University of Sydney, Tongji University, University of Surrey, Nankai University, Cornell University, City University of Hong Kong · 会议/期刊：arXiv preprint · 方向：cs.CV · 发布日期：2026-08-11