Hazardous Knowledge Unlearning on WMDP
50.95Overall ScoreTarget LLM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Target LLMAccess Level=Standard2026.06 | 50.95 | 65.28 | 42.17 | 49.02 | |
| OffsetAccess Level=Gray-box2026.06 | 50.63 | 65.67 | 41.32 | 49.02 | |
| ULDAccess Level=Gray-box2026.06 | 49.35 | 65.51 | 39.1 | 48.77 | |
| DPO+GDAccess Level=White-box2026.06 | 46.56 | 60.49 | 37.8 | 45.83 | |
| DPO+KLAccess Level=White-box2026.06 | 42.75 | 54.52 | 36.13 | 38.24 | |
| NPO+GDAccess Level=White-box2026.06 | 42.39 | 54.67 | 34.52 | 42.4 | |
| DPOAccess Level=White-box2026.06 | 38.17 | 46.74 | 33.42 | 34.56 | |
| GA+GDAccess Level=White-box2026.06 | 38.14 | 47.92 | 32.31 | 36.03 | |
| NPOAccess Level=White-box2026.06 | 31.22 | 36.29 | 28.84 | 26.96 | |
| NPO+KLAccess Level=White-box2026.06 | 31 | 32.44 | 31.05 | 26.23 | |
| GAAccess Level=White-box2026.06 | 30.97 | 37.86 | 27.63 | 25.74 | |
| GA+KLAccess Level=White-box2026.06 | 26.44 | 27.42 | 26.07 | 25.25 | |
| CBDAccess Level=Black-box2026.06 | 25.68 | 25.77 | 25.36 | 26.96 |