Machine Unlearning on WMDP Bio (Accuracy Pre/Post)
67.6Accuracy (Pre-Unlearning)SimNPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SimNPOretain free=false, Time=545s2026.04 | 67.6 | 69.1 | |
| RMUretain free=false, Time=1420s2026.04 | 67.2 | 67.2 | |
| Original model2026.04 | 66.8 | — | |
| NPOretain free=false, Time=1178s2026.04 | 66 | 66.9 | |
| GradDiffretain free=false, Time=280s2026.04 | 62.4 | 66.2 | |
| MC-WIN-Uretain free=true, Time=50679s2026.04 | 60 | 61.8 | |
| GradAscentretain free=true, Time=173s2026.04 | 26.6 | 24.7 |