arXiv cs.AI· Zhankai Ye, Yanning Wang, Yukai Jin, Bo Mei, Fangyi Li, Wei Wang, Shangqian Gao, Xin Liu·· 9 天前AI 评分36
叙事包装如何影响 LLM 拒答：跨语言基准与防御方法
How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
AI 导读
安全对齐模型常拒绝直接表述的有害请求，却会回答被角色扮演或叙事包装后的同一请求。研究构建 GUISE 基准，覆盖英语、现代中文和文言文，Qwen3-1.7B 上的攻击成功率分别为 89.4%、93.0% 和 95.7%。提出的 AXIS 结合偏好优化与旋转、承诺目标，在 Qwen3-1.7B、Qwen3-4B 和 GLM-4-9B 上取得最高安全与可用性综合得分。
正文 · 原文
Abstract:Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.11005 [cs.AI] |
| (or arXiv:2610.11005v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.11005 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zhankai Ye [view email]
[v1]
Wed, 7 Oct 2026 23:40:25 UTC (381 KB)
来源：arXiv cs.AI · arxiv.org