arXiv cs.CL· Liam Dugan, Callum McDougall, Matthew Denton, Philip Pham, Christine Kaeser-Chen, Chris Callison-Burch·· 13 小时前AI 评分42
用稀疏自编码器构建可审计的 AI 文本检测器
Building Auditable AI Text Detectors with Sparse Autoencoders
AI 导读
研究者提出一条流水线，将微调后的神经 AI 文本检测器转换为基于人类可读概念的分类器：直接在检测器残差流上训练稀疏自编码器（SAE），为每个 SAE 特征分配自动解释标签，仅保留描述通过全局保真度阈值的特征，再在其上训练轻量分类器。
正文 · 原文
Abstract:AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose decisions are opaque, while transparent alternatives (stylometric and linguistic-feature classifiers) underperform them by wide margins. We present a pipeline that converts a fine-tuned neural detector into a classifier that operates on human-readable concepts. We accomplish this by training sparse autoencoders (SAEs) directly on a detector's residual stream, assigning auto-interpretation labels to each SAE feature, and retaining only features whose generated descriptions pass a global fidelity threshold. We then train a lightweight classifier on just the interpretable features. We find that our SAEs recover latent concepts that correspond to highly specific and predictive linguistic "tells" of AI-generated text. Classifiers built on these extracted features are not only highly interpretable but exhibit stronger performance on out-of-distribution texts than either neural or stylometric alternatives.
| Comments: | EMNLP 2026 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| ACM classes: | I.2.7 |
| Cite as: | arXiv:2610.30006 [cs.CL] |
| (or arXiv:2610.30006v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.30006 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Liam Dugan [view email]
[v1]
Wed, 21 Oct 2026 16:41:14 UTC (1,556 KB)
来源：arXiv cs.CL · arxiv.org