The paper introduces RULER, a set of representation-level verification metrics for machine unlearning, which aims to remove the influence of specific training records from a deployed model. Unlike current output-level evaluations, RULER detects residuals of forgotten records in intermediate representations, revealing that approximate unlearning methods may still encode forgotten information.
DualOptim+ is a novel optimization framework designed to improve machine unlearning in large language models by bridging shared and decoupled optimizer states. It uses base states for common representations and delta states for objective-specific residuals, also offering a quantized 8-bit variant to reduce memory overhead without compromising performance.
Machine unlearning typically focuses on single-task settings, but modern AI models often operate in multi-task environments with shared backbones, leading to unintended interference when data is removed. This paper introduces multi-task unlearning, proposing an interference-aware framework that uses task-aware gradient projection to address task-level and instance-level interference.
This paper introduces ManiF-SMC, a novel method for approximate machine unlearning that addresses limitations of existing approaches. It reformulates unlearning as moving erased samples' manifold representations toward retained data's semantic neighbors, aiming for equivalence with retraining.