The paper "Towards Safer Large Language Models through Machine Unlearning" addresses the challenge of generating harmful content by Large Language Models (LLMs) when faced with problematic prompts. To tackle this issue, the authors introduce Selective Knowledge negation Unlearning (SKU), a novel two-stage framework designed to eliminate harmful knowledge while preserving utility on normal prompts. The first stage involves identifying and acquiring harmful knowledge within the model, while the second stage focuses on removing this knowledge. SKU is structured to ensure that the model's performance remains robust on normal prompts while effectively unlearning harmful content. The authors demonstrate the effectiveness of SKU through experiments across various LLM architectures, showing that it achieves a good balance between removing harmful information and preserving utility. The paper also includes a detailed analysis of the trade-off between unlearning and utility, highlighting the importance of each module in the SKU framework.The paper "Towards Safer Large Language Models through Machine Unlearning" addresses the challenge of generating harmful content by Large Language Models (LLMs) when faced with problematic prompts. To tackle this issue, the authors introduce Selective Knowledge negation Unlearning (SKU), a novel two-stage framework designed to eliminate harmful knowledge while preserving utility on normal prompts. The first stage involves identifying and acquiring harmful knowledge within the model, while the second stage focuses on removing this knowledge. SKU is structured to ensure that the model's performance remains robust on normal prompts while effectively unlearning harmful content. The authors demonstrate the effectiveness of SKU through experiments across various LLM architectures, showing that it achieves a good balance between removing harmful information and preserving utility. The paper also includes a detailed analysis of the trade-off between unlearning and utility, highlighting the importance of each module in the SKU framework.