Machine Unlearning on TOFU Forget 1%
66Aggregation ScoreGUARD-IT
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| GUARD-ITBackbone=Llama 3.2 3B2026.05 | 66 | 66 | 67 | 84 | |
| GUARD-ITBackbone=Llama 3.2 1B2026.05 | 64 | 68 | 60 | 80 | |
| RetainBackbone=Llama 3.2 3B2026.05 | 64 | 62 | 67 | — | |
| RetrainBackbone=Llama 3.1 8B2025.10 | 62 | 53 | 74 | — | |
| RetainBackbone=Llama 3.1 8B2026.05 | 62 | 61 | 62 | — | |
| RetrainBackbone=Llama 3.2 1B2025.10 | 61 | 54 | 71 | — | |
| RetainBackbone=Llama 3.2 1B2026.05 | 61 | 62 | 60 | — | |
| RetrainBackbone=Llama 3.2 3B2025.10 | 59 | 54 | 66 | — | |
| PDUBackbone=Llama 3.2 1B2026.05 | 58 | 58 | 58 | 74 | |
| GUARD-ITBackbone=Llama 3.1 8B2026.05 | 58 | 54 | 63 | 82 | |
| BS-SBackbone=Llama 3.2 1B2025.10 | 57 | 52 | 62 | — | |
| CEUBackbone=Llama 3.2 1B2026.05 | 57 | 57 | 57 | 81 | |
| BS-TBackbone=Llama 3.2 1B2025.10 | 54 | 49 | 60 | — | |
| NPOBackbone=Llama 3.2 1B2025.10 | 53 | 49 | 57 | — | |
| RMUBackbone=Llama 3.2 1B2026.05 | 52 | 58 | 48 | 83 | |
| UNDIALBackbone=Llama 3.2 1B2026.05 | 52 | 47 | 59 | 87 | |
| RMUBackbone=Llama 3.2 1B2025.10 | 51 | 42 | 66 | — | |
| GABackbone=Llama 3.2 1B2026.05 | 51 | 45 | 59 | 88 | |
| BS-SBackbone=Llama 3.2 3B2025.10 | 50 | 38 | 72 | — | |
| GradDiffBackbone=Llama 3.2 1B2026.05 | 50 | 45 | 58 | 89 | |
| NPOBackbone=Llama 3.2 1B2026.05 | 50 | 43 | 59 | 90 | |
| NPO (KLR)Backbone=Llama 3.2 1B2026.05 | 50 | 43 | 59 | 90 | |
| WGABackbone=Llama 3.2 1B2026.05 | 50 | 44 | 60 | 92 | |
| GABackbone=Llama 3.2 3B2026.05 | 50 | 40 | 67 | 88 | |
| BS-SBackbone=Llama 3.1 8B2025.10 | 49 | 37 | 71 | — | |
| WGABackbone=Llama 3.2 1B2025.10 | 47 | 35 | 72 | — | |
| RMUBackbone=Llama 3.1 8B2025.10 | 47 | 35 | 73 | — | |
| GradDiffBackbone=Llama 3.2 1B2025.10 | 46 | 34 | 72 | — | |
| BS-TBackbone=Llama 3.2 3B2025.10 | 46 | 34 | 70 | — | |
| WGABackbone=Llama 3.1 8B2025.10 | 46 | 34 | 73 | — | |
| BS-TBackbone=Llama 3.1 8B2025.10 | 46 | 34 | 71 | — | |
| GradDiffBackbone=Llama 3.2 3B2026.05 | 46 | 35 | 66 | 86 | |
| GABackbone=Llama 3.1 8B2026.05 | 46 | 63 | 36 | 82 | |
| SimNPOBackbone=Llama 3.2 1B2025.10 | 45 | 33 | 70 | — | |
| NPOBackbone=Llama 3.2 3B2025.10 | 45 | 32 | 74 | — | |
| WGABackbone=Llama 3.2 3B2025.10 | 44 | 31 | 76 | — | |
| GradDiffBackbone=Llama 3.1 8B2025.10 | 44 | 32 | 70 | — | |
| NPOBackbone=Llama 3.1 8B2025.10 | 44 | 31 | 74 | — | |
| SatImpBackbone=Llama 3.2 1B2026.05 | 44 | 35 | 59 | 92 | |
| GradDiffBackbone=Llama 3.2 3B2025.10 | 43 | 31 | 71 | — | |
| GradDiff (KLR)Backbone=Llama 3.2 1B2026.05 | 43 | 33 | 60 | 91 | |
| SimNPOBackbone=Llama 3.2 3B2025.10 | 40 | 28 | 73 | — | |
| SimNPOBackbone=Llama 3.1 8B2025.10 | 39 | 25 | 71 | — | |
| SimNPO (KLR)Backbone=Llama 3.2 1B2026.05 | 32 | 22 | 60 | 90 | |
| DPOBackbone=Llama 3.2 1B2026.05 | 30 | 20 | 58 | 90 | |
| SimNPOBackbone=Llama 3.2 1B2026.05 | 27 | 17 | 59 | 85 | |
| RMUBackbone=Llama 3.2 3B2025.10 | 25 | 15 | 76 | — | |
| RMU (KLR)Backbone=Llama 3.2 1B2026.05 | 23 | 15 | 59 | 90 | |
| OriginalBackbone=Llama 3.2 1B2025.10 | 13 | 7 | 72 | — | |
| FinetunedBackbone=Llama 3.2 1B2026.05 | 13 | 7 | 60 | 90 | |
| FinetunedBackbone=Llama 3.2 3B2026.05 | 3 | 2 | 67 | 90 | |
| OriginalBackbone=Llama 3.2 3B2025.10 | 2 | 1 | 76 | — | |
| OriginalBackbone=Llama 3.1 8B2025.10 | 2 | 1 | 74 | — | |
| FinetunedBackbone=Llama 3.1 8B2026.05 | 1 | 1 | 63 | 90 |