---
id: 20260922-T0-10
title: "MoE路由只看当前token？Attention-Aware Routing把注意力信息喂给路由器"
title_en: "Attention-Aware Routing Feeds Attention Context into MoE Router"
url: https://ai.daily.yangsir.net/daily/20260922-T0-10
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/2609.20974
---

# MoE路由只看当前token？Attention-Aware Routing把注意力信息喂给路由器

在混合专家（MoE）语言模型中，路由器通常仅根据当前token的隐藏状态选择专家，利用的上下文信息有限。研究者提出Attention-Aware Routing（AAR），将时序和注意力信息注入路由器，使其在专家选择时参考更长范围的上下文。论文编号arXiv:2609.20974，目前未公布与标准路由器的 perplexity 或下游任务对比数据。如果有效，MoE模型在同等参数量下可能获得更好的专家利用率。

## English Version

**Attention-Aware Routing Feeds Attention Context into MoE Router**

In Mixture-of-Experts (MoE) language models, routers typically select experts based only on the current token's hidden state, using limited contextual information. Researchers propose Attention-Aware Routing (AAR), which augments the router with temporal and attention information so expert selection draws on longer-range context. Posted as arXiv:2609.20974, no perplexity or downstream task comparisons against standard routers are yet available. If effective, MoE models could achieve better expert utilization at equivalent parameter counts.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2609.20974)

**详情页**：https://ai.daily.yangsir.net/daily/20260922-T0-10

---

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