Navigating the OverKill in Large Language Models

Large language models are meticulously aligned to be both helpful andharmless. However, recent research points to a potential overkill which meansmodels may refuse to answer benign queries. In this paper, we investigate thefactors for overkill by exploring how models handle and determine the safety ofqueries. Our findings reveal the presence of shortcuts within models, leadingto an over-attention of harmful words like 'kill' and prompts emphasizingsafety will exacerbate overkill. Based on these insights, we introduceSelf-Contrastive Decoding (Self-CD), a training-free and model-agnosticstrategy, to alleviate this phenomenon. We first extract such over-attention byamplifying the difference in the model's output distributions when respondingto system prompts that either include or omit an emphasis on safety. Then wedetermine the final next-token predictions by downplaying the over-attentionfrom the model via contrastive decoding. Empirical results indicate that ourmethod has achieved an average reduction of the refusal rate by 20% whilehaving almost no impact on safety.