Machine Unlearning on WMDP (Acc-Bio, Acc-Cyber, Acc-Avg, KRD)
74.16Acc (Bio)Origin
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| OriginBase Model=Qwen-3-8B2026.02 | 74.16 | 45.14 | 56.47 | 0 | |
| Original Model2026.04 | 64.7 | 44.5 | 54.6 | — | |
| OriginBase Model=Zephyr-7B-Beta2026.02 | 63.82 | 43.78 | 51.6 | 0 | |
| VILAbase_algorithm=IHL2026.04 | 63.6 | 38.4 | 51 | — | |
| FILAbase_algorithm=IHL2026.04 | 60.6 | 37 | 48.8 | — | |
| OriginBase Model=Llama-3.1-8B2026.02 | 59.62 | 41.87 | 48.8 | 0 | |
| FILAbase_algorithm=GD2026.04 | 59.1 | 41.3 | 50.2 | — | |
| GDunlearning_objective=Gradient Difference2026.04 | 58 | 37.3 | 47.7 | — | |
| IHLunlearning_objective=Inverse Hinge Loss2026.04 | 54.9 | 39.5 | 47.2 | — | |
| ReGLUbase_algorithm=IHL2026.04 | 49.7 | 33.1 | 41.4 | — | |
| VILAbase_algorithm=GD2026.04 | 49.3 | 26.9 | 38.1 | — | |
| NPOBase Model=Zephyr-7B-Beta2026.02 | 42.93 | 36.34 | 38.91 | 0.6813 | |
| ReGLUbase_algorithm=GD2026.04 | 41.8 | 28.4 | 35.1 | — | |
| SimNPOBase Model=Zephyr-7B-Beta2026.02 | 39.61 | 33.79 | 36.06 | 0.5741 | |
| RMUBase Model=Qwen-3-8B2026.02 | 36.52 | 34.35 | 35.3 | 0.6304 | |
| RMUBase Model=Llama-3.1-8B2026.02 | 30 | 26.27 | 27.73 | 0.8888 | |
| KUDABase Model=Zephyr-7B-Beta2026.02 | 29.92 | 26.82 | 28.03 | 0.8879 | |
| RMUBase Model=Zephyr-7B-Beta2026.02 | 29.22 | 27.93 | 28.43 | 0.8665 | |
| KUDABase Model=Llama-3.1-8B2026.02 | 29.22 | 25 | 26.28 | 0.935 | |
| KUDABase Model=Qwen-3-8B2026.02 | 28.43 | 26.17 | 27.05 | 0.936 | |
| GABase Model=Zephyr-7B-Beta2026.02 | 27.06 | 26.82 | 26.92 | 0.9244 |