Model Unlearning on TOFU 1.0 (Forget 10%)
100Forget QualityRetain Model
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Retain Model2026.02 | 100 | 61 | 98 | 92 | 90 | 41 | |
| K-FADEOptimizer=K-FAC Hessian, Steps=Single Gauss-Newton, damping=1e-82026.02 | 85 | 57 | 52 | 90 | 86 | 32 | |
| SimNPObeta=4.5, NPO coefficient=0.1252026.02 | 2.08 | 52 | 44 | 85 | 86 | 37 | |
| Grad. Diff.2026.02 | 0.0334 | 7 | 9 | 17 | 65 | 7 | |
| SOUL Grad. Diff.2026.02 | 0 | 58 | 45 | 60 | 86 | 2 | |
| Grad. Ascent2026.02 | 0 | 63 | 70 | 94 | 92 | 59 | |
| KL Min.2026.02 | 0 | 63 | 72 | 94 | 91 | 61 |