New Method Enhances LLM Safety Alignment Against Jailbreaks.

Researchers have developed a novel fine-tuning technique called HARC (Harmfulness-And-Refusal Coupling) to significantly bolster the safety alignment of large language models. The method is based on the insight that LLMs encode harmfulness and refusal as distinct, separable directions within their internal processing streams. By understanding how jailbreaks manipulate these internal representations at the prompt encoding stage, HARC intervenes by pairing these two directions across both prompt and response generation. This targeted intervention ensures that the model recognizes and refuses harmful content more consistently, even if the initial prompt-side detection failed. Crucially, HARC operates within a specific "harmfulness-refusal subspace," preserving the model's general capabilities and preventing excessive refusal of benign queries. Extensive testing across various model families and scales demonstrates HARC's superior trade-off between robustness, capability, and usability compared to existing safety methods.