Attend, Transform, or Silence: Operator-Level Visual Skipping for Efficient Multimodal LLM Inference figure
AlphaXiv 中文概览(可滚动查看)