MAAT: Multi-phase Adapter-Aware Targeted Unlearning

Researchers introduced 5WBENCH, a balanced 5W benchmark with 1,000 samples each for who, what, when, where, and why questions.

They found Why-type causal facts are underrepresented in prior datasets and are much harder to forget because they often require longer, multi-hop reasoning chains.

They also proposed MAAT, a three-stage LoRA-only unlearning pipeline that uses gradient-projected ascent, rank-dimension pruning with task-vector negation, and retain repair.

On 5WBENCH, MAAT improved the forget-retain tradeoff over baselines and was the only reported method to exceed 60% on both forgetting and retention across all five categories.