Long-context large language model (LLM) inference is increasingly constrained by the cost of maintaining an ever-growing KV-cache. Low-rank K/V projections p…
机构：杜克大学
来源：arXiv 2611.02583 | AI4Papers 论文推荐平台