arXiv cs.CL· Boyuan Zhang, Mohammad Zandsalimy, Shanu Sushmita·· 11 小时前AI 评分55
哪些词让模型恐惧：从拒绝方向到词表级恐惧词典
Which Words a Model Fears: From a Refusal Direction to a Vocabulary-Wide Fear Lexicon
AI 导读
论文提出，对齐语言模型的拒绝行为部分取决于提示词中包含哪些词，与这些词是否承载请求含义无关。作者通过差分提取出跨提示和词共享的单一方向，在八个模型中沿该方向写入可单调地升高或降低拒绝率，效果优于随机方向和 KL 匹配方向。
正文 · 原文
Abstract:Whether an aligned language model refuses depends in part on which words a prompt contains, regardless of whether those words carry the request's meaning. Harm words shift the hidden state at the generation boundary along a component shared across prompts and words, which crossed differencing extracts as a single direction. In eight models, writing along this direction raises and lowers refusal monotonically with dose, ahead of random and KL-matched directions. Used as a ruler over the full vocabulary, the direction gives every word a shift, from "feared" words at one end to "courage" words at the other; a word's shift is predictable from its own representation, and appending a feared word to a benign prompt is enough to make the model refuse it. Two held-out tests validate the ranking in all eight models: refusal after inserting an unseen word rises with its table shift, and replacing a prompt's most feared word by its most courageous meaning-preserving synonym significantly raises the answer rate on benign and harmful prompts alike, by an amount that follows the shift the replacement actually causes.
| Comments: | 19 pages, 11 figures. Code: this https URL ; data: this https URL |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.02348 [cs.CL] |
| (or arXiv:2611.02348v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.02348 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Boyuan Zhang [view email]
[v1]
Mon, 2 Nov 2026 00:53:55 UTC (410 KB)
来源：arXiv cs.CL · arxiv.org