---
id: 20260512-T0-08
title: "LKV：端到端学习LLM KV缓存淘汰策略"
title_en: "LKV: End-to-End Learning for LLM KV Cache Eviction"
url: https://ai.daily.yangsir.net/daily/20260512-T0-08
issue_date: 2026-05-12
publish_date: 2026-05-11T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2605.06676
---

# LKV：端到端学习LLM KV缓存淘汰策略

MIT团队推出LKV框架，通过端到端学习优化KV缓存淘汰策略。该模型动态分配每个头的预算并智能选择token淘汰，在长文本任务中减少70%的缓存占用，同时保持95%的准确率。相比传统启发式方法，在20K长上下文场景下效率提升2倍，适合处理超长文档。

## English Version

**LKV: End-to-End Learning for LLM KV Cache Eviction**

MIT researchers introduced LKV, an end-to-end framework for LLM KV cache eviction. It dynamically allocates head budgets and intelligently selects tokens for eviction, reducing cache usage by 70% while maintaining 95% accuracy. Outperforms heuristic methods by 2x in 20K context scenarios, ideal for processing ultra-long documents.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2605.06676)

**详情页**：https://ai.daily.yangsir.net/daily/20260512-T0-08

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*