MAAT: multi-stage adapter-aware targeted machine unlearning

MAAT is the first machine unlearning method to achieve both high forgetting and high retention on Why-type causal knowledge. The method acts on LoRA adapter weights in three stages, combining gradient projection ascent, SVD rank-dimension pruning, task vector negation and KL hidden-state retention repair. The accompanying 5WBENCH benchmark contains 5,000 samples, with 1,000 entries for each of the 5W categories, and is the first to quantify causal unlearning failure; the difficulty of the Why category stems from multi-hop reasoning chains (accounting for 44% of its entries).