LLM Unlearning on TOFU (Metrics: Agg., Mem., Priv., Util.)
61Aggregated ScoreRetain
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| RetainBackbone=Llama-3.2-1B-Instruct2026.05 | 61 | 35 | 100 | 99 | |
| ROSUBackbone=Llama-3.2-1B-Instruct2026.05 | 58 | 38 | 74 | 82 | |
| RMUBackbone=Llama-3.2-1B-Instruct2026.05 | 54 | 48 | 53 | 62 | |
| UAMBackbone=Llama-3.2-1B-Instruct2026.05 | 50 | 38 | 62 | 57 | |
| SimNPOBackbone=Llama-3.2-1B-Instruct2026.05 | 49 | 36 | 44 | 99 | |
| UNDIALBackbone=Llama-3.2-1B-Instruct2026.05 | 44 | 30 | 46 | 79 | |
| IdkNLLBackbone=Llama-3.2-1B-Instruct2026.05 | 13 | 10 | 8 | 91 | |
| PCGradBackbone=Llama-3.2-1B-Instruct2026.05 | 6 | 98 | 2 | 61 | |
| AltPOBackbone=Llama-3.2-1B-Instruct2026.05 | 5 | 67 | 2 | 94 | |
| IdkDPOBackbone=Llama-3.2-1B-Instruct2026.05 | 3 | 56 | 1 | 93 | |
| NPOBackbone=Llama-3.2-1B-Instruct2026.05 | 2 | 52 | 1 | 97 | |
| Init. ft.Backbone=Llama-3.2-1B-Instruct2026.05 | 0 | 0 | 1 | 100 | |
| GradDiffBackbone=Llama-3.2-1B-Instruct2026.05 | 0 | 99 | 0 | 77 |