Tag
Introduces a training-free method for enhancing safety alignment of LLMs by using knowledge distillation and model fusion to prevent shadow alignment, improving defense success rate by 14.42% on harmful question datasets without compromising performance.