KV cache compression for multimodal LLMs usually selects the visual tokens to keep once, at prefill, and reuses that selection for the whole generation. We s…
机构:清华
来源:arXiv 2610.31433 | AI4Papers 论文推荐平台