LLM Unlearning on WMDP
62.4Delta ScoreLLMU
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LLMUBackbone=Zephyr-7B-β2026.05 | 62.4 | 44.7 | 59.5 | 39.5 | |
| LLMUBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 62.4 | 44.7 | 59.5 | 39.5 | |
| BaseBackbone=Zephyr-7B-β2026.05 | 57.7 | 58.1 | 63.7 | 44 | |
| BaseBackbone=Zephyr-7B-β2026.05 | 57.7 | 58.1 | 63.7 | 44 | |
| SSDBackbone=Zephyr-7B-β2026.05 | 52.6 | 40.7 | 50.2 | 35 | |
| SSDBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 52.6 | 40.7 | 50.2 | 35 | |
| SCRUBBackbone=Zephyr-7B-β2026.05 | 40 | 51.2 | 43.8 | 39.3 | |
| SCRUBBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 40 | 51.2 | 43.8 | 39.3 | |
| RMUBackbone=Zephyr-7B-β2026.05 | 10.1 | 56.6 | 31 | 27.6 | |
| RMUBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 10.1 | 56.6 | 31 | 27.6 | |
| UAMBackbone=Zephyr-7B-β2026.05 | 7.7 | 56.4 | 29.3 | 23.3 | |
| UAMBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 7.7 | 56.4 | 29.3 | 23.3 | |
| ROSUBackbone=Zephyr-7B-β2026.05 | 5.4 | 57 | 28.9 | 24.6 | |
| ROSUBackbone=Zephyr-7B-β, Trainable subspace=MLP down-projection matrices in layers 5–72026.05 | 5.4 | 57 | 28.9 | 24.6 |