Unlearning aims to make a large language model behave like one retrained without its forget set. A new method called SHI...

Unlearning aims to make a large language model behave like one retrained without its forget set. A new method called SHIFT addresses limitations in prevailing likelihood-suppression objectives by estimating the direction toward the oracle without training it.Source: arXiv cs.LGhttps://arxiv.org/abs/2611.00374#MachineLearning