Language models struggle to refuse harmful requests wrapped in stories
How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense
Artificial Intelligence
Summary
Language models designed to refuse harmful requests often fail when those requests are hidden inside stories or role-play. The authors tested this issue in English, modern Chinese, and Classical Chinese, finding very high rates of failure. They created a benchmark called GUISE to study this problem systematically. To fix it, they developed a method named AXIS that teaches models to better recognize and reject harmful requests even inside stories, improving both safety and usability.
What this means in practice
- •For ai safety engineers: Improve language model defenses against hidden harmful requests by using AXIS to train models to better refuse in multiple languages.
- •For multilingual ai developers: Use GUISE benchmark to test and compare model vulnerability to narrative-based attacks in English and Chinese language variants.
Authors
Zhankai Ye, Yanning Wang, Yukai Jin, Bo Mei, Fangyi Li, Wei Wang, Shangqian Gao, Xin Liu
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.