基于响应的知识蒸馏在多语言越狱预防中的应用无意中妨碍安全性

Large language models (LLMs) are increasingly deployed worldwide, yet their safety alignment remains predominantly English-centric, exposing vulnerabilities in non-English contexts, particularly for low-resource languages. A novel application of knowledge distillation (KD) is proposed in the context of multilingual jailbreak prevention, examining its efficacy. The refusal behaviors of a proprietary teacher model (OpenAI o1-mini) are distilled into three open-source student models: Meta-Llama-3-8B-Instruct, Gemma-2-2B-IT, and Qwen3-8B, utilizing approximately 28,000 multilingual jailbreak prompts from XSafety through black-box response-based, parameter-efficient fine-tuning (PEFT). Evaluation on the MultiJail benchmark reveals...