---
id: 20260501-T0-07
title: "KV缓存驱逐新方法提升长文本生成效率"
title_en: "New KV Cache Eviction Method Boosts Long-Context Generation"
url: https://ai.daily.yangsir.net/daily/20260501-T0-07
issue_date: 2026-05-01
publish_date: 2026-04-30T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2604.25975
---

# KV缓存驱逐新方法提升长文本生成效率

研究提出了一种基于信息论的统一目标函数来优化KV缓存驱逐策略。传统方法依赖经验启发式规则，而新方法通过信息理论分析缓存数据的价值，有效降低了长文本生成时的内存开销，提升了大模型推理效率。

## English Version

**New KV Cache Eviction Method Boosts Long-Context Generation**

Researchers introduced a unified information-theoretic objective for KV cache eviction policies. Unlike traditional heuristic-based methods, this approach analyzes cache data value through information theory, reducing memory overhead during long-context generation.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2604.25975)

**详情页**：https://ai.daily.yangsir.net/daily/20260501-T0-07

---

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