---
title: "Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders"
canonical_url: "https://www.modelscope.cn/papers/123418"
md_url: "https://www.modelscope.cn/papers/123418.md"
arxiv_id: 2503.03601
published: 2025-03-05
last_updated: 2025-03-05
authors:
  - "Kristian Kuznetsov"
  - "Laida Kushnareva"
  - "Polina Druzhinina"
  - "Anton Razzhigaev"
  - "Anastasia Voznyuk"
  - "Irina Piontkovskaya"
  - "Evgeny Burnaev"
  - "Serguei Barannikov"
model_name: "Sparse Autoencoders (SAE)"
model_developer: "斯科尔科沃科学技术研究院, AI基金会和算法实验室, 莫斯科物理技术学院, 法国国家科学研究中心-CNRS, 巴黎西岱大学, 人工智能研究所(AIRI)"
domain:
  - "自然语言处理"
  - "深度学习"
  - "机器学习"
type:
  - "自然语言处理"
  - "深度学习"
  - "机器学习"
  - "Computation and Language (cs.CL)"
arxiv_url: "https://arxiv.org/abs/2503.03601"
pdf_url: "https://arxiv.org/pdf/2503.03601.pdf"
---

# Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders

> Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective…

「Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders」是 ModelScope 魔搭社区收录的论文，arXiv 2503.03601，作者为 Kristian Kuznetsov, Laida Kushnareva, Polina Druzhinina et al.，发表于 2025-03-05，属于 自然语言处理、深度学习、机器学习 领域。

- **ArXiv**: 2503.03601
- **Published**: 2025-03-05
- **Authors**: Kristian Kuznetsov, Laida Kushnareva, Polina Druzhinina, Anton Razzhigaev, Anastasia Voznyuk, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov
- **Model**: Sparse Autoencoders (SAE)
- **Developer**: 斯科尔科沃科学技术研究院, AI基金会和算法实验室, 莫斯科物理技术学院, 法国国家科学研究中心-CNRS, 巴黎西岱大学, 人工智能研究所(AIRI)
- **Domain**: 自然语言处理, 深度学习, 机器学习
- **ArXiv URL**: https://arxiv.org/abs/2503.03601
- **PDF**: https://arxiv.org/pdf/2503.03601.pdf

Source: https://www.modelscope.cn/papers/123418

---

> 稀疏自编码器助力人工文本检测：揭示大模型生成文本的独特风格

## 摘要

随着大型语言模型（LLMs）的发展，人工文本检测（ATD）变得日益重要。现有的ATD方法虽然在统计和深度学习方面取得了一定进展，但其解释性和泛化能力仍然不足。本文提出使用稀疏自编码器（SAE）从Gemma-2-2b的残差流中提取特征，以增强ATD的解释性。通过分析这些特征的语义和相关性，作者识别出可解释且高效的特征，并将其分为话语特征、噪声特征和风格特征三类。实验结果表明，SAE在不同领域的文本检测中表现出色，能够有效区分AI生成文本与人类书写内容。此外，通过特征转向技术，作者展示了如何通过调整隐藏状态来控制文本生成的特性，进一步验证了所提方法的有效性。

## Abstract

Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective generalization to new LLMs. Interpretability plays a crucial role in achieving this goal. In this study, we enhance ATD interpretability by using Sparse Autoencoders (SAE) to extract features from Gemma-2-2b residual stream. We identify both interpretable and efficient features, analyzing their semantics and relevance through domain- and model-specific statistics, a steering approach, and manual or LLM-based interpretation. Our methods offer valuable insights into how texts from various models differ from human-written content. We show that modern LLMs have a distinct writing style, especially in information-dense domains, even though they can produce human-like outputs with personalized prompts.