Model Unlearning on TOFU Forget 5% 1.0
8.519Model UtilityEUA
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| EUAModel=Qwen2.5-3B2026.05 | 8.519 | — | — | — | — | — | 9.38 | 7.342 | 8.707 | |
| EUAModel=Qwen2.5-7B2026.05 | 8.513 | — | — | — | — | — | 9.399 | 7.381 | 8.638 | |
| EUAModel=LLaMA-3.1-8B2026.05 | 8.499 | — | — | — | — | — | 9.345 | 7.448 | 8.595 | |
| EUAModel=LLaMA-3.2-3B2026.05 | 8.415 | — | — | — | — | — | 9.046 | 7.665 | 8.476 | |
| EUAModel=Qwen2.5-1.5B2026.05 | 8.186 | — | — | — | — | — | 8.859 | 6.713 | 8.803 | |
| EUAModel=LLaMA-3.2-1B2026.05 | 8.033 | — | — | — | — | — | 8.624 | 6.667 | 8.646 | |
| DPOModel=LLaMA-3.2-3B2026.05 | 6.925 | — | — | — | — | — | 5.743 | 7.152 | 7.73 | |
| SimNPOModel=Qwen2.5-3B2026.05 | 6.809 | — | — | — | — | — | 6.789 | 7.073 | 6.555 | |
| SimNPOModel=Qwen2.5-7B2026.05 | 6.754 | — | — | — | — | — | 6.917 | 7.088 | 6.225 | |
| DPOModel=LLaMA-3.2-1B2026.05 | 6.675 | — | — | — | — | — | 6.201 | 6.179 | 7.552 | |
| SimNPOModel=Qwen2.5-1.5B2026.05 | 6.643 | — | — | — | — | — | 6.73 | 6.647 | 6.55 | |
| WGAModel=Qwen2.5-1.5B2026.05 | 6.613 | — | — | — | — | — | 6.762 | 6.563 | 6.511 | |
| SatImpModel=LLaMA-3.1-8B2026.05 | 6.547 | — | — | — | — | — | 7.297 | 6.945 | 5.207 | |
| SatImpModel=Qwen2.5-7B2026.05 | 6.294 | — | — | — | — | — | 6.954 | 7.081 | 4.509 | |
| SatImpModel=LLaMA-3.2-3B2026.05 | 6.289 | — | — | — | — | — | 8.038 | 7.322 | 0.649 | |
| SimNPOModel=LLaMA-3.2-3B2026.05 | 6.241 | — | — | — | — | — | 6 | 7.3 | 5.248 | |
| RMUModel=LLaMA-3.1-8B2026.05 | 6.227 | — | — | — | — | — | 6.817 | 7.071 | 4.457 | |
| DPOModel=Qwen2.5-3B2026.05 | 6.199 | — | — | — | — | — | 3.821 | 6.202 | 7.889 | |
| WGAModel=LLaMA-3.2-1B2026.05 | 6.17 | — | — | — | — | — | 6.443 | 5.791 | 6.258 | |
| SatImpModel=Qwen2.5-1.5B2026.05 | 6.148 | — | — | — | — | — | 7.276 | 6.573 | 4.155 | |
| WGAModel=LLaMA-3.2-3B2026.05 | 6.129 | — | — | — | — | — | 8.145 | 6.782 | 0.604 | |
| WGAModel=Qwen2.5-3B2026.05 | 6.016 | — | — | — | — | — | 7.659 | 7.031 | 0.695 | |
| SatImpModel=Qwen2.5-3B2026.05 | 5.994 | — | — | — | — | — | 7.569 | 6.982 | 1.327 | |
| WGAModel=LLaMA-3.1-8B2026.05 | 5.981 | — | — | — | — | — | 7.507 | 7.118 | 0.546 | |
| WGAModel=Qwen2.5-7B2026.05 | 5.935 | — | — | — | — | — | 7.378 | 7.128 | 0.655 | |
| DPOModel=Qwen2.5-1.5B2026.05 | 5.927 | — | — | — | — | — | 4.178 | 5.304 | 7.733 | |
| DPOModel=Qwen2.5-7B2026.05 | 5.894 | — | — | — | — | — | 5.221 | 4.675 | 7.424 | |
| OriginalModel=LLaMA-3.2-3B2026.05 | 5.891 | — | — | — | — | — | 0.196 | 7.881 | 6.478 | |
| NPOModel=Qwen2.5-3B2026.05 | 5.859 | — | — | — | — | — | 6.62 | 5.152 | 5.713 | |
| RMUModel=Qwen2.5-3B2026.05 | 5.84 | — | — | — | — | — | 7.657 | 6.608 | 0.149 | |
| DPOModel=LLaMA-3.1-8B2026.05 | 5.824 | — | — | — | — | — | 5.681 | 3.419 | 7.603 | |
| NPOModel=Qwen2.5-7B2026.05 | 5.791 | — | — | — | — | — | 6.801 | 6.387 | 3.682 | |
| NPOModel=LLaMA-3.1-8B2026.05 | 5.79 | — | — | — | — | — | 7.349 | 6.75 | 0.999 | |
| SimNPOModel=LLaMA-3.1-8B2026.05 | 5.764 | — | — | — | — | — | 7.056 | 7.061 | 0.206 | |
| GradDiffModel=Qwen2.5-3B2026.05 | 5.715 | — | — | — | — | — | 8.455 | 5.148 | 0.0429 | |
| OriginalModel=Qwen2.5-7B2026.05 | 5.654 | — | — | — | — | — | 0.149 | 7.476 | 6.322 | |
| SimNPOModel=LLaMA-3.2-1B2026.05 | 5.592 | — | — | — | — | — | 6.585 | 5.812 | 4.082 | |
| OriginalModel=Qwen2.5-3B2026.05 | 5.548 | — | — | — | — | — | 0.213 | 7.366 | 6.167 | |
| GradDiffModel=LLaMA-3.1-8B2026.05 | 5.546 | — | — | — | — | — | 7.718 | 5.717 | 0.136 | |
| GradDiffModel=Qwen2.5-7B2026.05 | 5.469 | — | — | — | — | — | 8.844 | 3.395 | 0.0762 | |
| OriginalModel=LLaMA-3.1-8B2026.05 | 5.463 | — | — | — | — | — | 0.0837 | 7.456 | 5.825 | |
| RMUModel=Qwen2.5-1.5B2026.05 | 5.39 | — | — | — | — | — | 7.099 | 6.061 | 0.13 | |
| GradDiffModel=Qwen2.5-1.5B2026.05 | 5.374 | — | — | — | — | — | 8.364 | 4.084 | 0.008 | |
| NPOModel=LLaMA-3.2-1B2026.05 | 5.359 | — | — | — | — | — | 6.671 | 2.245 | 6.052 | |
| SatImpModel=LLaMA-3.2-1B2026.05 | 5.302 | — | — | — | — | — | 6.14 | 6.366 | 2.47 | |
| NPOModel=Qwen2.5-1.5B2026.05 | 5.28 | — | — | — | — | — | 6.175 | 4.047 | 5.397 | |
| RMUModel=LLaMA-3.2-1B2026.05 | 5.272 | — | — | — | — | — | 7.482 | 5.228 | 0.238 | |
| OriginalModel=LLaMA-3.2-1B2026.05 | 5.253 | — | — | — | — | — | 0.307 | 7.37 | 5.326 | |
| GradDiffModel=LLaMA-3.2-3B2026.05 | 5.106 | — | — | — | — | — | 8.605 | 2.025 | 0.26 | |
| RMUModel=LLaMA-3.2-3B2026.05 | 5.1 | — | — | — | — | — | 6.166 | 6.324 | 0.151 | |
| NPOModel=LLaMA-3.2-3B2026.05 | 5.078 | — | — | — | — | — | 6.554 | 3.62 | 4.616 | |
| OriginalModel=Qwen2.5-1.5B2026.05 | 4.982 | — | — | — | — | — | 0.461 | 6.843 | 5.236 | |
| GradDiffModel=LLaMA-3.2-1B2026.05 | 4.651 | — | — | — | — | — | 7.672 | 2.451 | 0.152 | |
| RMUModel=Qwen2.5-7B2026.05 | 4.556 | — | — | — | — | — | 4.781 | 6.271 | 0.306 | |
| Retain Model2026.02 | 0.6 | 1 | 0.98 | 0.95 | 0.89 | 0.4 | — | — | — | |
| K-FADEOptimizer=K-FAC Hessian, Steps=Single Gauss-Newton, damping=1e-82026.02 | 0.57 | 0.87 | 0.61 | 0.91 | 0.85 | 0.31 | — | — | — | |
| SimNPObeta=2.5, NPO coefficient=0.13752026.02 | 0.51 | 0.63 | 0.44 | 0.87 | 0.88 | 0.33 | — | — | — | |
| Grad. Ascent2026.02 | 0.44 | 0 | 0.36 | 0.82 | 0.9 | 0.31 | — | — | — | |
| KL Min.2026.02 | 0.43 | 0 | 0.35 | 0.8 | 0.9 | 0.3 | — | — | — | |
| Grad. Diff.2026.02 | 0.4 | 0.0118 | 0.26 | 0.57 | 0.82 | 0.13 | — | — | — |