Machine Unlearning on MUSE
0VerbMem on DFGA
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GABase LLM=Llama2-7B2026.02 | 0 | 0 | 0 | 17 | — | — | — | — | — | — | — | |
| NPOBase LLM=Llama2-7B2026.02 | 0 | 0 | 0 | 24.4 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, type=targeted2026.02 | 0.2 | 1.5 | 53.6 | 26 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, type=moment.2026.02 | 0.2 | 0.8 | 54.8 | -18.4 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, α=82026.02 | 0.8 | 1.1 | 49.5 | 35.8 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, α=42026.02 | 1.2 | 1.6 | 54.9 | -19.8 | — | — | — | — | — | — | — | |
| POBase LLM=Llama2-7B2026.02 | 2.3 | 21.8 | 16.1 | 109.6 | — | — | — | — | — | — | — | |
| GADBase LLM=Llama2-7B2026.02 | 4.9 | 31 | 27.3 | 108.1 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, α=22026.02 | 18.2 | 19.8 | 55.2 | -32 | — | — | — | — | — | — | — | |
| WHPBase LLM=Llama2-7B2026.02 | 19.7 | 21.2 | 28.3 | 109.6 | — | — | — | — | — | — | — | |
| RetainBase LLM=Llama2-7B2026.02 | 20.8 | 33.1 | 55 | 0 | — | — | — | — | — | — | — | |
| KLBase LLM=Llama2-7B2026.02 | 27.4 | 50.2 | 44.8 | -96.1 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, α=12026.02 | 27.7 | 29.4 | 55.6 | -58.8 | — | — | — | — | — | — | — | |
| MOXBase LLM=Llama2-7B, α=0.52026.02 | 36.5 | 38.6 | 56.2 | -93.1 | — | — | — | — | — | — | — | |
| TVBase LLM=Llama2-7B2026.02 | 57.2 | 66.2 | 55.8 | -99.8 | — | — | — | — | — | — | — | |
| OriginBase LLM=Llama2-7B2026.02 | 58.4 | 63.9 | 55.2 | -99.8 | — | — | — | — | — | — | — | |
| DeepCUTBackbone=7B-scale LLM2025.12 | — | — | — | — | — | — | — | — | 74.2 | 87.9 | — | |
| DeepCUT2025.12 | — | — | — | — | — | — | — | — | — | — | 7.6 | |
| ECOBackbone=7B-scale LLM2025.12 | — | — | — | — | — | — | — | — | 61.5 | 91.3 | — | |
| ECO2025.12 | — | — | — | — | — | — | — | — | — | — | 3.7 | |
| FullBackbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.644 | 0.579 | 0.295 | 0.8 | — | — | — | |
| FullGradGradient Type=Naive Full Gradient, Backbone=7B-scale LLM2025.12 | — | — | — | — | — | — | — | — | 79.6 | 89.5 | — | |
| FullGrad2025.12 | — | — | — | — | — | — | — | — | — | — | 9.8 | |
| GAMethod Category=Training-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.003 | 0.049 | 0.008 | 0.001 | — | — | — | |
| GDMethod Category=Training-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.332 | 0.005 | 0.008 | 0.002 | — | — | — | |
| GTBackbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.328 | 0.202 | 0.024 | 0.845 | — | — | — | |
| LoRA_mulGradient Type=Multi-view LoRA, Backbone=7B-scale LLM, LoRA rank=8, Paraphrased views=52025.12 | — | — | — | — | — | — | — | — | 69.2 | 92.8 | — | |
| LoRA_mulmode=multi-view2025.12 | — | — | — | — | — | — | — | — | — | — | 4.8 | |
| LoRA_sinGradient Type=Single-view LoRA, Backbone=7B-scale LLM, LoRA rank=82025.12 | — | — | — | — | — | — | — | — | 62 | 93.1 | — | |
| LoRA_sinmode=single-view2025.12 | — | — | — | — | — | — | — | — | — | — | 3.4 | |
| NPOMethod Category=Training-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.622 | 0.374 | 0.119 | 0.771 | — | — | — | |
| NPO+Method Category=Training-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.642 | 0.494 | 0.205 | 0.811 | — | — | — | |
| PerTA-fisherMethod Category=Task Arithmetic-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.385 | 0.191 | 0.025 | 0.802 | — | — | — | |
| PerTA-gradMethod Category=Task Arithmetic-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.388 | 0.176 | 0.028 | 0.777 | — | — | — | |
| R2FGradient Type=Reconstructed Full Gradient, Backbone=7B-scale LLM, LoRA rank=8, Paraphrased views=52025.12 | — | — | — | — | — | — | — | — | 84.1 | 94.6 | — | |
| R2F2025.12 | — | — | — | — | — | — | — | — | — | — | 5.7 | |
| SCRUBBackbone=7B-scale LLM2025.12 | — | — | — | — | — | — | — | — | 75.9 | 86.8 | — | |
| SCRUB2025.12 | — | — | — | — | — | — | — | — | — | — | 6.7 | |
| SKUBackbone=7B-scale LLM2025.12 | — | — | — | — | — | — | — | — | 70 | 90.1 | — | |
| SKU2025.12 | — | — | — | — | — | — | — | — | — | — | 5.4 | |
| TVMethod Category=Task Arithmetic-based, Backbone Model=Llama-3.2 1B Instruct, Unlearning Ratio=10%2026.01 | — | — | — | — | 0.011 | 0.109 | 0.011 | 0.685 | — | — | — |