---
id: 20260818-T0-07
title: "线性注意力新方法：让模型更懂'忘记'什么，提升长上下文检索"
title_en: "New Erase Direction for Linear Attention Improves Long-Context Retrieval"
url: https://ai.daily.yangsir.net/daily/20260818-T0-07
issue_date: 2026-08-18
publish_date: 2026-08-17T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2608.13668
---

# 线性注意力新方法：让模型更懂'忘记'什么，提升长上下文检索

一篇新的arXiv论文提出了一种针对线性注意力的新方法。线性注意力模型（如Gated DeltaNet-2）在长上下文处理中，由于固定大小的状态存储过多信息，导致检索性能下降。该研究通过改进'擦除向量'的生成方式，解决信息干扰问题。实验结果表明，该方法能有效提升模型在长上下文任务中的表现。

## English Version

**New Erase Direction for Linear Attention Improves Long-Context Retrieval**

A new arXiv paper introduces a method to derive a second, more effective erase direction for linear attention models like Gated DeltaNet-2. By addressing inter-item interference in the fixed-size state, the proposed approach significantly improves retrieval performance in long-context tasks. This addresses a key weakness of these models.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2608.13668)

**详情页**：https://ai.daily.yangsir.net/daily/20260818-T0-07

---

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