面向智能体的缓存工程:AKV 直接管理 KV 缓存
原标题:Paper page - Cache Engineering for Agents
AI 摘要
该论文提出「缓存工程」(cache engineering)新方案,让 LLM 智能体直接管理自身的 KV 缓存,通过 drop 和 repos 操作在不修改轨迹、不重新预填充的前提下控制活跃缓存增长。作者指出相同 token 序列在不同缓存操作序列下会产生不同缓存状态,因此共同设计了 AKV 服务引擎,提供无状态接口并内部管理有状态缓存复用、驱逐与恢复。实验显示,工具响应滚动丢弃策略相比传统阈值触发摘要平均基准分数提升 1.7 分、平均成本降低 27.7%,输出吞吐最高提升 1.2 倍,P95 延迟加速 1.5 倍,递归自我改进缓存策略再降成本 7.1%。
正文节选
Cache Engineering for Agents Abstract AKV introduces cache engineering to manage LLM agent key-value caches without altering trajectories, improving efficiency and accuracy via drop and repos operations with a co-designed serving engine. Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's m