Reasoning vision-language models (VLMs) impose substantial KV cache demands from both large multimodal contexts and growing reasoning traces. Low-rank KV com…
机构:杜克大学
来源:arXiv 2611.15569 | AI4Papers 论文推荐平台