---
id: 20261106-T0-08
title: "混合宽专家：让语言模型输出坐标本身可解释"
title_en: "Mixture of Wide Experts Makes Output Coordinates Interpretable"
url: https://ai.daily.yangsir.net/daily/20261106-T0-08
issue_date: 2026-11-06
publish_date: 2026-11-05T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.03655
---

# 混合宽专家：让语言模型输出坐标本身可解释

语言模型常把一个概念分散到大量内部维度，特征难以解读；稀疏自编码器（SAE）需在训练后才能还原较清晰的特征。该研究反过来提问：模型能否直接学出可解释的输出坐标？提出的Mixture of Wide Experts用稀疏路由在原生坐标上组织专家。摘要未给出量化结果。实际影响：若成立，模型内部表示的解释可以不依赖事后训练SAE，降低可解释性分析的成本。

## English Version

**Mixture of Wide Experts Makes Output Coordinates Interpretable**

Language models tend to spread a single concept across many internal dimensions, making features hard to interpret. Sparse autoencoders (SAEs) recover clearer features, but only after training. This work flips the question: can models learn interpretable output coordinates directly? Mixture of Wide Experts uses sparse routing over native coordinates to organize experts. The abstract reports no quantitative results. Practical impact: if it holds, interpretability would not require post-hoc SAE training, lowering the cost of analyzing what a model represents.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.03655)

**详情页**：https://ai.daily.yangsir.net/daily/20261106-T0-08

---

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