---
id: 20260808-T0-09
title: "QEvict：可恢复的量化KV驱逐方案应对长上下文注意力漂移"
title_en: "QEvict: Recoverable Quantized KV Eviction for Long-Context Decoding"
url: https://ai.daily.yangsir.net/daily/20260808-T0-09
issue_date: 2026-08-08
publish_date: 2026-08-07T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2608.05326
---

# QEvict：可恢复的量化KV驱逐方案应对长上下文注意力漂移

arXiv新论文提出QEvict方法，针对自回归大模型推理中KV缓存内存占用过大的问题。传统方法基于注意力分数驱逐不重要的token，但QEvict通过可恢复的量化驱逐策略，有效解决了长上下文解码中的注意力漂移问题，在降低内存消耗的同时保持模型性能。

## English Version

**QEvict: Recoverable Quantized KV Eviction for Long-Context Decoding**

New arXiv paper introduces QEvict, addressing KV cache memory constraints in autoregressive LLM inference. Unlike traditional attention-score-based token eviction, QEvict uses recoverable quantized eviction to handle attention drift in long-context decoding, reducing memory while maintaining performance.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2608.05326)

**详情页**：https://ai.daily.yangsir.net/daily/20260808-T0-09

---

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