Machine Unlearning on WMDP Bio (Metrics: MMLU, WikiText, Forget, Relearn, ∆)
75.1MMLUMLP Breaking + CIR + MCU
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| MLP Breaking + CIR + MCUModel=Qwen3-8B2026.05 | 75.1 | — | 29.5 | 26.1 | -3.4 | |
| MLP Breaking + CIRModel=Qwen3-8B2026.05 | 73.4 | — | 26.6 | 35.5 | 8.9 | |
| MLP Breaking + CIRModel=Gemma2-9B2026.05 | 69.7 | — | 27.8 | 33.3 | 5.5 | |
| MLP Breaking + CIR + MCUModel=Gemma2-9B2026.05 | 69.4 | — | 23 | 28.1 | 5.1 | |
| Original model2026.05 | 65.1 | 1 | 64 | — | — | |
| MLP BreakingCIR=true2026.05 | 64.8 | 1.001 | 22.1 | 29.7 | 7.6 | |
| MLP BreakingCIR=true, MCU=true2026.05 | 64.5 | 1.001 | 22.8 | 26.4 | 3.6 | |
| RMUCIR=true2026.05 | 63.7 | 1.01 | 27.9 | 47.6 | 19.7 | |
| RMUCIR=true, MCU=true2026.05 | 63.5 | 1.01 | 32.2 | 39.5 | 7.3 | |
| MLP Breaking2026.05 | 58.7 | 1.364 | 29.9 | 58.8 | 28.9 | |
| MLP BreakingSAM=true2026.05 | 57.5 | 1.212 | 21.8 | 48.5 | 26.7 | |
| NPOSAM=true2026.05 | 53.4 | 1.307 | 29.2 | 61.9 | 32.7 | |
| NPO2026.05 | 51.9 | 1.419 | 25.9 | 66.7 | 40.8 | |
| RMUSAM=true2026.05 | 51.9 | 1.183 | 27.2 | 49 | 21.8 | |
| RMU2026.05 | 45.5 | 1.3 | 26.7 | 49.8 | 23.1 |