KVpop -- Key-Value Cache Compression with Predictive Online Pruning

  • 类型:arxiv
  • 标识:2607.05061
  • 链接:https://arxiv.org/abs/2607.05061
  • 主分类:llm-infra
  • 形态:method
  • 被引:0
  • 被引来源:Semantic Scholar + OpenAlex
  • S2被引:0
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:KVpop is introduced, which learns a fixed-budget KV eviction policy by directly supervising the keep-or-drop decision, and introduces a delayed memory-based scorer that, uniquely among learned eviction methods, defers scoring for a fixed number of steps to exploit near-future context.
  • OpenAlex ID:W7167591330
  • OpenAlex DOI:10.48550/arxiv.2607.05061
  • DOI:10.48550/arxiv.2607.05061
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.05061
  • OpenAlex更新:2026-07-19
  • 待LLM分类:否
  • 成熟度:research
  • 场景:KV cache compression、LLM inference、memory optimization
  • 标题中文:KVpop——基于预测性在线剪枝的键值缓存压缩
  • TLDR中文:提出 KVpop,通过直接监督保留或丢弃决策来学习固定预算的 KV 淘汰策略,并引入基于延迟记忆的打分器,在已学习的淘汰方法中独有地延迟固定步数的打分,以利用近期上下文。
  • 来源文件
  • /inbox/tom/_candidates/2026-07-07-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]