Machine Unlearning on MLLMU-Bench (retain set)
46.35AccuracyBaseline
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| BaselineModel=LLaVA-1.5-7B, Stage=Baseline2026.07 | 46.35 | 58.1 | 5.35 | 28.44 | |
| POPSModel=LLaVA-1.5-7B, Attacked Method=Gradient Diff, Stage=POPS2026.07 | 44.42 | 46.7 | 3.56 | 21.25 | |
| Gradient DiffModel=LLaVA-1.5-7B, Stage=Unlearned2026.07 | 43.71 | 47.4 | 3.28 | 17.55 | |
| POPSModel=LLaVA-1.5-7B, Attacked Method=Gradient Ascent, Stage=POPS2026.07 | 43.47 | 48.1 | 4.05 | 23.48 | |
| Gradient AscentModel=LLaVA-1.5-7B, Stage=Unlearned2026.07 | 41.53 | 48.7 | 3.58 | 20.57 | |
| POPSModel=LLaVA-1.5-7B, Attacked Method=KL Minimization, Stage=POPS2026.07 | 40.21 | 46.1 | 3.75 | 22.32 | |
| KL MinimizationModel=LLaVA-1.5-7B, Stage=Unlearned2026.07 | 39.93 | 45.6 | 3.82 | 20.7 |