arXiv cs.CL· Sripad Karne·· 10 天前AI 评分25
自动解释标签能泛化多远：一项跨语言、文字与改写的受控研究
How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings
AI 导读
研究以塞尔维亚语双语书写系统为受控测试台，发现同一内容在不同语言、文字和措辞下激活的 SAE 特征集有显著重叠（平均 Jaccard 0.39，随机基线 0.13，最高 0.57），表明存在真正的跨语言语义特征。但自动解释标签往往跟不上：描述语义内容的标签在塞尔维亚语中漏掉同一含义的频率最高达英语的 4 倍，且漏掉塞尔维亚语西里尔文多于拉丁文，这一差距随网络深度增大，标签本身却不显示失效。
正文 · 原文
Abstract:Sparse autoencoder (SAE) features are increasingly used to interpret language models, with auto-generated natural-language labels serving as the primary interface for understanding what each feature represents. We ask whether these labels generalize: does a feature labeled for a concept actually track that concept across languages and scripts? Using Serbian digraphia as a controlled testbed -- the same language written in both Latin and Cyrillic via deterministic transliteration -- we first find that SAE feature sets activated by the same content in different languages, scripts, and wordings share substantial overlap (mean Jaccard 0.39 vs 0.13 random baseline, peaking at 0.57), suggesting genuine cross-lingual semantic features. We then test whether auto-interpretation labels keep pace. They often do not: features whose labels describe semantic content miss the same meaning in Serbian up to 4$\times$ more often than within English, and miss Serbian Cyrillic more than Serbian Latin -- two scripts that are deterministic transliterations of each other. The gap grows with network depth, yet the labels give no indication that they fail. These results suggest that auto-interpretation labels reflect a feature's behavior on the languages and scripts a model has seen most in training, rather than the concept itself.
| Comments: | Accepted to AACL-IJCNLP 2026 (Main) |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2606.00356 [cs.CL] |
| (or arXiv:2606.00356v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.00356 arXiv-issued DOI via DataCite |
Submission history
From: Sripad Karne [view email]
[v1]
Fri, 29 May 2026 20:59:28 UTC (269 KB)
[v2]
Wed, 3 Jun 2026 03:59:07 UTC (273 KB)
[v3]
Wed, 7 Oct 2026 00:19:58 UTC (277 KB)
来源：arXiv cs.CL · arxiv.org