U-CAN：基于效用感知的对比衰减方法用于生成推荐中的高效遗忘

Generative Recommendation (GenRec) typically utilizes Large Language Models (LLMs) to redefine personalization as an instruction-driven sequence generation task. However, fine-tuning on user logs inadvertently encodes sensitive attributes into model parameters, raising significant privacy concerns. Existing Machine Unlearning (MU) techniques struggle with the Polysemy Dilemma, where neurons superimpose sensitive data with general reasoning patterns, resulting in catastrophic utility loss under traditional gradient or pruning methods. To address this issue, a precision unlearning framework called Utility-aware Contrastive Attenuation (U-CAN) is proposed, which operates on low-rank adapters. U-CAN quantifies risk by contrasting activations, aiming to achieve effective unlearning while minimizing the impact on model utility.