Large Language Models (LLMs) like OpenAI's GPT series, Anthropic's Claude, and Meta's LLaMa have shown remarkable capabilities in text generation. However, their susceptibility to toxic prompts presents significant security challenges. This paper investigates alignment techniques, including Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), to mitigate these risks. We conduct an empirical study on refusal patterns across nine LLMs, revealing that models with uniform refusal patterns, such as Claude3, exhibit higher security. Based on these findings, we propose self-distilling and cross-model distilling methods to enhance LLM security. Our results show that these methods significantly improve refusal rates and reduce unsafe content, with cross-model distilling achieving refusal rates close to Claude3's 94.51%. These findings underscore the potential of distillation-based alignment in securing LLMs against toxic prompts.
@article{arxiv.2406.11285,
title = {Self and Cross-Model Distillation for LLMs: Effective Methods for Refusal Pattern Alignment},
author = {Jie Li and Yi Liu and Chongyang Liu and Xiaoning Ren and Ling Shi and Weisong Sun and Yinxing Xue},
journal= {arXiv preprint arXiv:2406.11285},
year = {2024}
}
Comments
The method used in the paper has obvious problems and ambiguities. The security enhancement method we used cannot be considered distillation, but it is described as distillation in the paper, and the experiment lacks comparison and baseline, which has been criticized by many peers. In order to avoid further dissemination, we have decided to withdraw the paper