---
id: 20260922-T0-09
title: "长文本预填充太慢？RBS-Attention用半径约束稀疏块加速"
title_en: "RBS-Attention Speeds Up Long-Context Prefill with Radius-Bounded Sparse Blocks"
url: https://ai.daily.yangsir.net/daily/20260922-T0-09
issue_date: 2026-09-22
publish_date: 2026-09-21T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2609.20971
---

# 长文本预填充太慢？RBS-Attention用半径约束稀疏块加速

长文本LLM推理的瓶颈正从解码转向预填充阶段，密集自注意力需要先处理整个提示词。研究者提出RBS-Attention（半径约束稀疏预填充），通过稀疏块选择降低预填充成本。论文指出，现有块选择方法存在缺陷：块质心可能掩盖块内高度相关的信息，导致选择偏差。RBS-Attention用半径约束来修正这一问题。论文编号arXiv:2609.20971，尚未公布与密集预填充的速度和精度对比数据。

## English Version

**RBS-Attention Speeds Up Long-Context Prefill with Radius-Bounded Sparse Blocks**

Long-context LLM inference is increasingly bottlenecked by prefill, where dense self-attention must process the entire prompt before generation. Researchers propose RBS-Attention (Radius-Bounded Sparse Prefill), which reduces prefill cost through sparse block selection. The paper notes a flaw in existing block-selection methods: a block centroid can mask highly relevant information within the block, causing selection bias. RBS-Attention applies radius bounds to correct this. Posted as arXiv:2609.20971, no speed or accuracy comparisons against dense prefill are yet available.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2609.20971)

**详情页**：https://ai.daily.yangsir.net/daily/20260922-T0-09

---

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