MAAT: Multi-phase Adapter-Aware Targeted Unlearning

In three lines5WBENCH, a balanced 5,000-sample benchmark across 5W categories, reveals unlearning methods fail on causal (Why) questions. MAAT, a three-phase framework operating on LoRA weights, combines gradient-projected ascent, SVD rank pruning, and KL-hidden-state repair to simultaneously achieve high forgetting and retention on causal knowledge.

Summary generated by Claude — human-verified