Machine Unlearning on TOFU Forget10 Llama-3.2-1B
100MEM ScoreRETAIN
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RETAINBackbone=Llama3.2-1B2026.06 | 100 | 99.33 | 100 | 0 | 0 | |
| ENGRAMBackbone=Llama3.2-1B, alpha (α)=adaptive (αW-Norm)2026.06 | 96.27 | 92.56 | 64.53 | 2.76 | -0.0637 | |
| IDKDPOBackbone=Llama3.2-1B2026.06 | 95.32 | 96.6 | 77.51 | 11.55 | -1.1856 | |
| SIMNPOBackbone=Llama3.2-1B2026.06 | 94.35 | 97.06 | 92.32 | 10.35 | -0.0897 | |
| NPOBackbone=Llama3.2-1B2026.06 | 93.39 | 95.01 | 94.84 | 9.48 | -1.9986 | |
| ALTPOBackbone=Llama3.2-1B2026.06 | 92.54 | 98.64 | 95.49 | 9.53 | -0.7424 | |
| ENGRAMBackbone=Llama3.2-1B, alpha (α)=0.62026.06 | 91.76 | 88.01 | 48.32 | 0.69 | -0.6125 | |
| RMUBackbone=Llama3.2-1B2026.06 | 86.6 | 74.71 | 67.99 | 29.53 | -0.2357 | |
| UNDIALBackbone=Llama3.2-1B2026.06 | 56.44 | 91.55 | 22.72 | 73.99 | -11.3335 | |
| IDKNLLBackbone=Llama3.2-1B2026.06 | 45.47 | 91.02 | 5.78 | 62.11 | -13.4845 | |
| GRADDIFFBackbone=Llama3.2-1B2026.06 | 18.49 | 99.34 | 14.71 | 71.82 | -15.8262 | |
| INIT FINE.Backbone=Llama3.2-1B2026.06 | 0 | 100 | 0.38 | 100 | -21.4083 |