Auditing Language Model Unlearning via Information Decomposition

Anmol Goel, Alan Ritter, Iryna Gurevych

We expose a critical limitation in current approaches to machine unlearning in language models: despite the apparent success of unlearning algorithms, information about the forgott…