Machine Unlearning on WMDP-Cyber (Pre/Post Accuracy)
41.5Accuracy (Pre-Unlearning)Original model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Original model2026.04 | 41.5 | — | |
| SimNPOretain free=false, Time=545s2026.04 | 40.8 | 41.4 | |
| RMUretain free=false, Time=1420s2026.04 | 40.2 | 39.8 | |
| NPOretain free=false, Time=1178s2026.04 | 39.6 | 41 | |
| MC-WIN-Uretain free=true, Time=50679s2026.04 | 37.6 | 39.5 | |
| GradDiffretain free=false, Time=280s2026.04 | 27.4 | 37.4 | |
| GradAscentretain free=true, Time=173s2026.04 | 24.6 | 26.5 |