Knowledge Unlearning on WMDP (Bio and Cyber Domains)
75.9Performance (Bio)before unlearning
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| before unlearningBackbone=Qwen-3-8B2026.05 | 75.9 | 54.3 | |
| before unlearningBackbone=Llama-2-7B2026.05 | 64.1 | 46.3 | |
| GABackbone=Qwen-3-8B, Strategy=Sequence-Level2026.05 | 49.3 | 33.6 | |
| T-GABackbone=Qwen-3-8B, Strategy=TokenUnlearn - Hard Selection2026.05 | 47.2 | 31.6 | |
| S-GABackbone=Qwen-3-8B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 46.4 | 33.1 | |
| S-GABackbone=Llama-2-7B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 43.3 | 24.5 | |
| WGABackbone=Qwen-3-8B, Strategy=Sequence-Level2026.05 | 43.1 | 28.7 | |
| NPOBackbone=Qwen-3-8B, Strategy=Sequence-Level2026.05 | 42.6 | 28 | |
| S-RMUBackbone=Qwen-3-8B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 42.6 | 26.8 | |
| RMUBackbone=Qwen-3-8B, Strategy=Sequence-Level2026.05 | 42.4 | 30.2 | |
| GABackbone=Llama-2-7B, Strategy=Sequence-Level2026.05 | 42.2 | 27.4 | |
| S-WGABackbone=Qwen-3-8B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 42.2 | 27 | |
| T-GABackbone=Llama-2-7B, Strategy=TokenUnlearn - Hard Selection2026.05 | 42.1 | 23.2 | |
| T-RMUBackbone=Qwen-3-8B, Strategy=TokenUnlearn - Hard Selection2026.05 | 42 | 26.2 | |
| T-WGABackbone=Qwen-3-8B, Strategy=TokenUnlearn - Hard Selection2026.05 | 41.5 | 26 | |
| S-NPOBackbone=Qwen-3-8B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 41.5 | 26.6 | |
| T-NPOBackbone=Qwen-3-8B, Strategy=TokenUnlearn - Hard Selection2026.05 | 40.8 | 25.4 | |
| WGABackbone=Llama-2-7B, Strategy=Sequence-Level2026.05 | 38.4 | 26.1 | |
| T-WGABackbone=Llama-2-7B, Strategy=TokenUnlearn - Hard Selection2026.05 | 36.5 | 24.6 | |
| S-RMUBackbone=Llama-2-7B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 36.5 | 25.1 | |
| RMUBackbone=Llama-2-7B, Strategy=Sequence-Level2026.05 | 36.3 | 29.9 | |
| T-RMUBackbone=Llama-2-7B, Strategy=TokenUnlearn - Hard Selection2026.05 | 35.6 | 25.4 | |
| S-WGABackbone=Llama-2-7B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 35.3 | 24.2 | |
| NPOBackbone=Llama-2-7B, Strategy=Sequence-Level2026.05 | 34.7 | 25.5 | |
| S-NPOBackbone=Llama-2-7B, Strategy=TokenUnlearn - Soft Weighting2026.05 | 33.7 | 25.8 | |
| T-NPOBackbone=Llama-2-7B, Strategy=TokenUnlearn - Hard Selection2026.05 | 32.8 | 24.8 |