可解释的大型语言模型的去学习通过推理实现

LLM unlearning is crucial for addressing safety, copyright, and privacy issues in pre-trained large language models (LLMs). Unlike preference alignment, unlearning provides a more explicit approach by removing undesirable knowledge characterized by specific unlearning datasets. Previous works utilizing gradient ascent (GA) and its variants have shown promise in implementing unlearning; however, their untargeted nature leads to unintended degradation of general capabilities, incomplete removal of knowledge, and the generation of incoherent responses. These issues arise from the lack of explicit guidance on what and how models should unlearn. To address this gap, a novel unlearning target is introduced, focusing on reasoning-based unlearning. This approach aims to provide clearer directions for unlearning, enhancing model performance and coherence.