Event ·Single source

语言模型可自主控制注意力机制

研究团队提出声明式注意力(DA)协议，让模型在思维链中声明需要关注的内容区域。该协议将生成分为全局、聚焦和局部三种模式，在15个长上下文任务中，Gemma-4-31B和Qwen-3.6-27B模型解码时关注token分别减少52.0%和31.1%，准确率仅下降1.27pp和2.75pp。

Current read

新DA协议让模型自己决定关注哪部分内容，大幅减少计算量，Gemma和Qwen模型效果显著。

2 sources78° AI heatLast updated 2026/9/2 15:43:38

Evidence

2 sources

Primary sourcearXiv cs.LG

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

Related source 1elvis

2026/9/3 20:40:02View 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