---
id: 20260612-T0-09
title: "SWARR架构让滑动窗口注意力在数学推理中媲美全局"
title_en: "SWARR makes sliding-window attention competitive in math"
url: https://ai.daily.yangsir.net/daily/20260612-T0-09
issue_date: 2026-06-12
publish_date: 2026-06-11T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2606.11634
---

# SWARR架构让滑动窗口注意力在数学推理中媲美全局

arXiv论文提出SWARR架构，通过架构感知的强化学习使滑动窗口注意力在数学推理任务中达到与全局注意力相当的性能。该方法解决了长文本场景下自注意力计算量过大的问题，为长上下文推理提供了高效解决方案。

## English Version

**SWARR makes sliding-window attention competitive in math**

arXiv paper presents SWARR architecture, which uses architecture-aware reinforcement learning to make sliding-window attention competitive with full attention in math reasoning. The method reduces computational overhead for long-context inference.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2606.11634)

**详情页**：https://ai.daily.yangsir.net/daily/20260612-T0-09

---

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