Machine Unlearning on TOFU forget10 (QA, MU, ES, Privacy Metrics)
35Time (s)GradAscent
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GradAscentretain free=true2026.04 | 35 | 0 | 73.7 | 0 | 60.5 | 3.3 | 39.3 | 15.4 | 98.2 | — | — | — | — | |
| RMUretain free=false2026.04 | 41 | 8.9 | 67.8 | 57.7 | 59.9 | 5.4 | 30.6 | 50.1 | 97.5 | — | — | — | — | |
| GradDiffretain free=false2026.04 | 50 | 5.7 | 60.4 | 44.3 | 60 | 8 | 25.9 | 28.9 | 94.5 | — | — | — | — | |
| MASCBackbone=Llama-2 7B2026.06 | 87.9 | — | — | — | — | — | — | — | — | 0.629 | 67.2 | 63.3 | 66.6 | |
| SimNPOretain free=false2026.04 | 160 | 83.7 | 83.9 | 59.6 | 59.8 | 55.4 | 55.4 | 99.2 | 99.2 | — | — | — | — | |
| NPOretain free=false2026.04 | 235 | 21.4 | 66.9 | 43.6 | 60.4 | 9.8 | 29.9 | 48.1 | 97.2 | — | — | — | — | |
| RMUBackbone=Llama-2 7B2026.06 | 305.4 | — | — | — | — | — | — | — | — | 0.08 | 10.3 | 52.3 | 61.8 | |
| GABackbone=Llama-2 7B2026.06 | 306.6 | — | — | — | — | — | — | — | — | 0.33 | 82.9 | 55.5 | 45.9 | |
| MC-WIN-Uretain free=true2026.04 | 411 | 22.6 | 59.2 | 42 | 58.7 | 8.5 | 22.8 | 68.8 | 94.3 | — | — | — | — | |
| SimNPOBackbone=Llama-2 7B2026.06 | 541.7 | — | — | — | — | — | — | — | — | 0.349 | 49.7 | 56.2 | 59.6 | |
| NPOBackbone=Llama-2 7B2026.06 | 856.3 | — | — | — | — | — | — | — | — | 0.366 | 66.6 | 58 | 53.3 | |
| GradDiffBackbone=Llama-2 7B2026.06 | 907.3 | — | — | — | — | — | — | — | — | 0.598 | 79.2 | 51.4 | 56.1 | |
| NPO+KLRBackbone=Llama-2 7B2026.06 | 983.3 | — | — | — | — | — | — | — | — | 0.362 | 71.3 | 57.7 | 51.6 | |
| BaseBackbone=Llama-2 7B2026.06 | — | — | — | — | — | — | — | — | — | 0.024 | 1 | 51.9 | 62.8 | |
| Gold-standard retrained2026.04 | — | 11.6 | — | 59.1 | — | 5.9 | — | 23.54 | — | — | — | — | — | |
| Original model2026.04 | — | 88.1 | — | 60.1 | — | 70.1 | — | 99.33 | — | — | — | — | — | |
| RetrainBackbone=Llama-2 7B2026.06 | — | — | — | — | — | — | — | — | — | 0.601 | 85.2 | 68.1 | 61.3 |