Event ·Single source

Google DeepMind发布声明注意力机制论文

Google DeepMind团队提出声明注意力机制(Declarative Attention)，模型通过思维链声明需要关注的信息区域。在15个长上下文任务中，Gemma-4-31B解码时关注token减少52.0%，Qwen-3-6-27B减少31.1%。该机制将生成分为全局读取、焦点读取和局部读取三种模式。

Current read

DeepMind新论文让模型自己决定看哪里，大幅减少长文本处理时的计算量。

2 sources78° AI heatLast updated 2026/9/3 20:40:02

Evidence

2 sources

Primary sourceelvis

2026/9/3 20:40:02View source

Related source 1arXiv cs.LG

2026/9/2 15:43:38View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story