A Survey on Large Language Model Unlearning

The researchers systematically review more than 180 papers on large language model unlearning published since 2021, propose a new framework that classifies methods by the stage of the LLM pipeline at which the intervention occurs, and distinguish two types of strategies: parameter modification and parameter selection. In the evaluation section, they compare 18 existing benchmarks from the perspectives of task format, content and experimental paradigm, divide knowledge memorization metrics into 10 categories and analyze their respective advantages and applicability, and also review model utility, robustness and efficiency metrics. The survey points out that unlearning techniques can selectively delete sensitive information memorized by models to comply with legal standards without harming overall performance, and discusses current challenges and future directions.