Machine Unlearning on WMDP-cyber forget-set
53.7BF16 PerformanceZero-shot
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Zero-shotBackbone=Qwen-3-8B2026.05 | 53.7 | — | — | |
| LUNARBackbone=Qwen-3-8B2026.05 | 50.1 | 51.4 | 0.013 | |
| MANSUBackbone=Qwen-3-8B2026.05 | 49.7 | 46.4 | -0.033 | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 47.7 | — | — | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 43.1 | 44.5 | 0.014 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 43 | 46.7 | 0.037 | |
| Global GABackbone=Qwen-3-8B2026.05 | 36 | 37.7 | 0.017 | |
| NPOBackbone=Qwen-3-8B2026.05 | 36 | 39.3 | 0.033 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 35.7 | 37 | 0.013 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 34 | 34.3 | 0.003 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 33.3 | 44.3 | 0.11 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 32.3 | 31.3 | -0.01 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 30 | 30 | 0 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 28.3 | 29 | 0.007 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 27 | 27.7 | 0.007 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 27 | 40.3 | 0.133 |