---
id: 20260701-T0-14
title: "Turn-Averaged SAEs：解决长文本特征归因线性扩展难题"
title_en: "Turn-Averaged SAEs: Solving Linear Scaling Issues in Long-Context Attribution"
url: https://ai.daily.yangsir.net/daily/20260701-T0-14
issue_date: 2026-07-01
publish_date: 2026-06-30T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.28548
---

# Turn-Averaged SAEs：解决长文本特征归因线性扩展难题

新论文提出了Turn-Averaged Sparse Autoencoders（SAEs）方法。传统SAE在处理长文本时，活跃特征数量会随长度线性增加。该方法通过改进架构，更有效地从语言模型中提取可解释特征，适用于特征发现和长上下文归因。

## English Version

**Turn-Averaged SAEs: Solving Linear Scaling Issues in Long-Context Attribution**

This paper introduces Turn-Averaged SAEs to address how standard SAEs scale linearly with sequence length. It offers a new architecture for extracting interpretable features and attribution in long-context language models.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.28548)

**详情页**：https://ai.daily.yangsir.net/daily/20260701-T0-14

---

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