Memorization Reduction on GSM8K
35.2Memorization Reduction (%)GSS
Evaluation Results
| Method | Links | |
|---|---|---|
| GSSCategory=Ours, Model=Qwen, Utility threshold=>= 0.92026.02 | 35.2 | |
| GSSCategory=Ours, Model=LLama, Utility threshold=>= 0.92026.02 | 33.5 | |
| Task VectorCategory=Unlearning, Model=Qwen, Utility threshold=>= 0.92026.02 | 25.4 | |
| Task VectorCategory=Unlearning, Model=LLama, Utility threshold=>= 0.92026.02 | 25.1 | |
| ROMECategory=Editing, Model=Qwen, Utility threshold=>= 0.92026.02 | 14.5 | |
| Balanced SubnetCategory=Editing, Model=Qwen, Utility threshold=>= 0.92026.02 | 12.8 | |
| LUNARCategory=Editing, Model=LLama, Utility threshold=>= 0.92026.02 | 9 | |
| MEMITCategory=Editing, Model=Qwen, Utility threshold=>= 0.92026.02 | 8.7 | |
| ROMECategory=Editing, Model=LLama, Utility threshold=>= 0.92026.02 | 7.8 | |
| DPOCategory=Unlearning, Model=Qwen, Utility threshold=>= 0.92026.02 | 7.3 | |
| SimNPOCategory=Unlearning, Model=Qwen, Utility threshold=>= 0.92026.02 | 7.2 | |
| NPOCategory=Unlearning, Model=Qwen, Utility threshold=>= 0.92026.02 | 6.9 | |
| GACategory=Unlearning, Model=LLama, Utility threshold=>= 0.92026.02 | 4.5 | |
| SimNPOCategory=Unlearning, Model=LLama, Utility threshold=>= 0.92026.02 | 2.2 | |
| NPOCategory=Unlearning, Model=LLama, Utility threshold=>= 0.92026.02 | 2.2 | |
| MEMITCategory=Editing, Model=LLama, Utility threshold=>= 0.92026.02 | 2.2 | |
| Balanced SubnetCategory=Editing, Model=LLama, Utility threshold=>= 0.92026.02 | 1.4 | |
| DPOCategory=Unlearning, Model=LLama, Utility threshold=>= 0.92026.02 | 1.3 | |
| LUNARCategory=Editing, Model=Qwen, Utility threshold=>= 0.92026.02 | 1 | |
| GACategory=Unlearning, Model=Qwen, Utility threshold=>= 0.92026.02 | 0 |