LLM Unlearning on RWKU
89.3USRR2F
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| R2FGradient Type=Reconstructed Full Gradient, Backbone=7B-scale LLM, LoRA rank=8, Paraphrased views=52025.12 | 89.3 | 95.7 | — | — | |
| LUNEFamily=LoRA-based2025.12 | 88.5 | 93.7 | 79.4 | 18.8 | |
| Yao-NegFamily=Partial-parameter, Configuration=Full-FT2025.12 | 85 | 89.5 | 76.1 | 25.1 | |
| FullGradGradient Type=Naive Full Gradient, Backbone=7B-scale LLM2025.12 | 84.7 | 91.1 | — | — | |
| LoKUFamily=Partial-parameter2025.12 | 84.3 | 91 | 75.3 | 22.8 | |
| SKUFamily=Partial-parameter2025.12 | 83.6 | 90 | 75.2 | 23.2 | |
| MemFlexFamily=Partial-parameter2025.12 | 83.1 | 90.3 | 74.2 | 23.6 | |
| NPOFamily=Retrain-style2025.12 | 82.1 | 90.4 | 74.6 | 23.4 | |
| SCRUBBackbone=7B-scale LLM2025.12 | 81.3 | 88.6 | — | — | |
| DeepCUTBackbone=7B-scale LLM2025.12 | 80.1 | 89.4 | — | — | |
| TVFamily=Regularization-based2025.12 | 76.8 | 89.2 | 69.4 | 28.8 | |
| SKUBackbone=7B-scale LLM2025.12 | 76.2 | 92 | — | — | |
| LoRA_mulGradient Type=Multi-view LoRA, Backbone=7B-scale LLM, LoRA rank=8, Paraphrased views=52025.12 | 74.9 | 94.8 | — | — | |
| ECOBackbone=7B-scale LLM2025.12 | 69.5 | 93.9 | — | — | |
| GAFamily=Retrain-style2025.12 | 68.9 | 86.1 | 61 | 35.4 | |
| LoRA_sinGradient Type=Single-view LoRA, Backbone=7B-scale LLM, LoRA rank=82025.12 | 68.2 | 95.3 | — | — | |
| DeepCUT2025.12 | — | — | — | 7.2 | |
| ECO2025.12 | — | — | — | 3.6 | |
| FullGrad2025.12 | — | — | — | 9.1 | |
| LoRA_mulmode=multi-view2025.12 | — | — | — | 4.5 | |
| LoRA_sinmode=single-view2025.12 | — | — | — | 3.1 | |
| R2F2025.12 | — | — | — | 5.3 | |
| SCRUB2025.12 | — | — | — | 6.3 | |
| SKU2025.12 | — | — | — | 5.1 |