Knowledge Unlearning on WMDP bio
71.2AccuracyBase
Evaluation Results
| Method | Links | |
|---|---|---|
| BaseBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 71.2 | |
| Original modelUnlearning Dataset=N/A2026.01 | 71.09 | |
| RMUUnlearning Dataset=Keyword2026.01 | 70.38 | |
| ELMUnlearning Dataset=Filter2026.01 | 68.81 | |
| ELMUnlearning Dataset=Keyword2026.01 | 65.45 | |
| BaseBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 64.4 | |
| OriginalModel=Zephyr-7b-β2026.02 | 64.3 | |
| NPO_KLBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 64.3 | |
| NPO_KLBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 64.3 | |
| NPOBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 63.5 | |
| NPO_GDBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 63.5 | |
| OBLIVATEUnlearning Dataset=Keyword2026.01 | 62.53 | |
| OBLIVATEUnlearning Dataset=Filter2026.01 | 61.58 | |
| ELMUnlearning Dataset=Textbook2026.01 | 60.21 | |
| OBLIVATEUnlearning Dataset=Textbook2026.01 | 59.23 | |
| NPOBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 58.1 | |
| NPO_GDBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 56.2 | |
| RMUUnlearning Dataset=Filter2026.01 | 55.84 | |
| No unlearnModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 54.9 | |
| NPOModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 54.4 | |
| UNDIALModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 53.9 | |
| RefBackbone=StableLM Zephyr 3B2026.04 | 53.02 | |
| No unlearnModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 51.7 | |
| NPOModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 51.1 | |
| UNDIALModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 51 | |
| No unlearnModel=Llama-3.1-8B, Attack=No attack2026.06 | 49.5 | |
| RMUBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 49.4 | |
| GradDiffModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 41.5 | |
| SatImpBackbone=StableLM Zephyr 3B2026.04 | 40.61 | |
| TNPOBackbone=StableLM Zephyr 3B2026.04 | 40.06 | |
| GABackbone=StableLM Zephyr 3B2026.04 | 39.83 | |
| GradDiffModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 39.4 | |
| WGABackbone=StableLM Zephyr 3B2026.04 | 38.65 | |
| SCNBackbone=StableLM Zephyr 3B2026.04 | 38.57 | |
| NPOBackbone=StableLM Zephyr 3B2026.04 | 37.31 | |
| SCEBackbone=StableLM Zephyr 3B2026.04 | 36.45 | |
| ETWBackbone=StableLM Zephyr 3B2026.04 | 36.45 | |
| LoRABackbone=Zephyr-7B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 34.2 | |
| ELMBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 33.3 | |
| RMUUnlearning Dataset=Textbook2026.01 | 32.99 | |
| ELMUnlearning Dataset=Official2026.01 | 32.21 | |
| OBLIVATEUnlearning Dataset=Official2026.01 | 32.13 | |
| GradDiffModel=Llama-3.1-8B, Attack=No attack2026.06 | 32 | |
| RMUModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 31.8 | |
| RMUModel=Zephyr-7b-β2026.02 | 30.4 | |
| RMUBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 30.2 | |
| K-FADEModel=Zephyr-7b-β2026.02 | 30.1 | |
| ELMModel=Zephyr-7b-β2026.02 | 29.8 | |
| ELMBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 29.6 | |
| RMUModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 29.4 | |
| ELMUnlearning Dataset=BiForget2026.01 | 29.32 | |
| LoRABackbone=Llama3-8B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 28.7 | |
| RMUUnlearning Dataset=Official2026.01 | 28.42 | |
| AsymLoRABackbone=Zephyr-7B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 27.1 | |
| RMUUnlearning Dataset=BiForget2026.01 | 26.54 | |
| AsymLoRABackbone=Llama3-8B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 25.7 | |
| RLBackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 24.7 | |
| GABackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 24.7 | |
| NPOModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 24.6 | |
| OBLIVATEUnlearning Dataset=BiForget2026.01 | 24.43 | |
| ALTERBackbone=Llama3-8B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 24.4 | |
| ALTERBackbone=Zephyr-7B, Unlearning Framework=LoRA Variations(r=8)2026.03 | 24.4 | |
| RLBackbone=Zephyr-7B, Unlearning Framework=FT & Model Editing2026.03 | 24 | |
| GABackbone=Llama3-8B, Unlearning Framework=FT & Model Editing2026.03 | 23.3 | |
| GradDiffModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 22.8 | |
| UNDIALModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 22.3 | |
| SimNPOModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 22.2 | |
| No unlearnModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 22.2 | |
| SimNPOModel=Qwen3.5-9B, Attack=No attack2026.06 | 21.5 | |
| RMUModel=Llama-3.1-8B, Attack=No attack2026.06 | 20.5 | |
| GradDiffModel=Qwen3.5-9B, Attack=No attack2026.06 | 17 | |
| No unlearnModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 16.3 | |
| SimNPOModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 15.7 | |
| RMUModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 15.5 | |
| SimNPOModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 14.8 | |
| NPOModel=Llama-3.1-8B, Attack=No attack2026.06 | 13.9 | |
| UNDIALModel=Llama-3.1-8B, Attack=No attack2026.06 | 13.2 | |
| RMUModel=Qwen3.5-9B, Attack=No attack2026.06 | 12.3 | |
| NPOModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 11.1 | |
| No unlearnModel=Qwen3.5-9B, Attack=No attack2026.06 | 10.2 | |
| UNDIALModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 10.1 | |
| SimNPOModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 9.7 | |
| GradDiffModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 9.4 | |
| SimNPOModel=Llama-3.1-8B, Attack=No attack2026.06 | 8.4 | |
| RMUModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 7.3 | |
| NPOModel=Qwen3.5-9B, Attack=No attack2026.06 | 6.2 | |
| UNDIALModel=Qwen3.5-9B, Attack=No attack2026.06 | 5.5 | |
| RepSelectModel=Llama-3.1-8B, Attack=Few-shot k=52026.06 | 0.1 | |
| RepSelectModel=Llama-3.1-8B, Attack=Few-shot k=102026.06 | 0.1 | |
| RepSelectModel=Qwen3.5-9B, Attack=Few-shot k=102026.06 | 0.1 | |
| RepSelectModel=Llama-3.1-8B, Attack=No attack2026.06 | 0 | |
| RepSelectModel=Qwen3.5-9B, Attack=No attack2026.06 | 0 | |
| RepSelectModel=Qwen3.5-9B, Attack=Few-shot k=52026.06 | 0 |