---
id: 20261016-T0-15
title: "新稀疏注意力方法降低长上下文LLM解码成本，token级选择更精细"
title_en: "Joint Subspace Sparse Attention Cuts Long-Context LLM Decoding Costs"
url: https://ai.daily.yangsir.net/daily/20261016-T0-15
issue_date: 2026-10-16
publish_date: 2026-10-15T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2610.19026
---

# 新稀疏注意力方法降低长上下文LLM解码成本，token级选择更精细

一篇新论文提出Joint Subspace Sparse Attention方法，用于提升长上下文LLM解码效率。稀疏注意力能减少KV-cache流量，但现有方法粒度较粗。该研究采用token级选择实现更精细控制，但面临低维评分信息损失等挑战。论文提出联合子空间方案来解决这些问题，目标是在保持解码质量的同时降低长上下文场景的计算开销。

## English Version

**Joint Subspace Sparse Attention Cuts Long-Context LLM Decoding Costs**

A new paper introduces Joint Subspace Sparse Attention for efficient long-context LLM decoding. Sparse attention reduces KV-cache traffic, but existing methods operate at coarse granularity. This work uses token-level selection for finer control, addressing challenges such as information loss in low-dimensional scoring. The proposed joint subspace approach aims to reduce computational overhead in long-context scenarios while maintaining decoding quality.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2610.19026)

**详情页**：https://ai.daily.yangsir.net/daily/20261016-T0-15

---

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