Two-Stage Method for Auditable Machine Unlearning in LLMs

To address the lack of evidence in data deletion requests, this method introduces auditability as a key axis for machine unlearning. A two-stage approach using GapWiden hinge loss improves the statistical separation between requested forget sets and retained data.