arXiv cs.CL· Pardis Sadat Zahraei, Gokhan Tur, Dilek Hakkani-T\"ur, Ehsaneddin Asgari·· 6 天前AI 评分48
对齐否决：安全训练如何压制 LLM 中的文化知识
The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs
AI 导读
研究覆盖 16 个 MENA 国家、26 个模型和 153 万份人类调查问卷，发现对齐训练与文化价值冲突时发生的是压制而非抹除：模型在拒绝回答时，其内部 logit 分布与人类调查数据的相关性反而强于自由生成的答案，作者称之为"对齐否决"。
正文 · 原文
Abstract:What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we show the answer is suppression, not erasure: at the moment of refusal, a model's internal logit distribution correlates with human survey data more strongly than its freely generated answers. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. The gate is inequitable: the safety tax reaches 37.6%, with a 19.8% alignment-quality gap between best- and worst-served nations, and native-language prompting widens rather than closes it. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2510.13154 [cs.CL] |
| (or arXiv:2510.13154v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2510.13154 arXiv-issued DOI via DataCite |
|
| Journal reference: | EMNLP 2026 Main Oral |
Submission history
From: Pardis Sadat Zahraei [view email]
[v1]
Wed, 15 Oct 2025 05:10:57 UTC (32,655 KB)
[v2]
Thu, 18 Jun 2026 05:54:33 UTC (5,205 KB)
[v3]
Fri, 9 Oct 2026 02:47:49 UTC (4,007 KB)
来源：arXiv cs.CL · arxiv.org