OpenAttention 方法在 1% 读取率下仍保持 99% 的准确率

作者 PulseAugur 编辑部 · [1 个来源] · 2026-10-23 04:00

研究人员推出了一种新颖的大型语言模型稀疏注意力方法 OpenAttention，该方法在显著降低计算成本的同时保持了高准确率。该技术利用“尾部代理”来表示键值缓存中未读取的部分，改进了之前丢弃这些信息或效率低下地对其进行采样的旧方法。在 RULER-32K HARD 基准测试上，使用 Llama-3.1-8B 时，OpenAttention 在仅 1% 的读取率下达到了密集注意力准确率的 99%，并且其自定义注意力内核在解码注意力阶段提供了十倍的速度提升。 AI

影响降低了大型语言模型的计算成本，可能以更少的硬件实现更快的推理和训练。

排序理由该条目是一篇 arXiv 预印本论文，详细介绍了一种用于大型语言模型的稀疏注意力新方法。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 4 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该条目是一篇 arXiv 预印本论文，详细介绍了一种用于大型语言模型的稀疏注意力新方法。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, infra; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Same-day; Cluster formed today. Ranking reflects the current source set at time of score.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv cs.LG TIER_1 English(EN) · Hyunwoo Yu, Jongbeom Lee, Hyunbin Kim, Sungjun Park, Jaemin Choi, Jincheol Yang, Seunghun Moon, Bongjoon Hyun, Kyomin Sohn, Kyeongbo Kong, Suk-Ju Kang · 2026-10-23 04:00

OpenAttention: One Percent Is Enough with Tail Proxies

arXiv:2610.30951v1 Announce Type: new Abstract: Sparse attention reads only a small subset of the key-value cache, and how to treat the unread tail remains a key challenge. Discarding the tail biases the output, and sampling it as previous approaches do fails at reads of 5% and b…

本文如何被排名

报道来源 [1]

OpenAttention: One Percent Is Enough with Tail Proxies

相关实体

相关话题