Prompt recovery attack on zsRE
40Top-1 AccuracyKSTER
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| KSTERModel=Llama-3-Instruct, N=10, Target Editing Method=MEMIT2026.02 | 40 | 64 | 90 | 0.86 | |
| KSTERModel=Llama-3-Instruct, N=50, Target Editing Method=MEMIT2026.02 | 40 | 60 | 76 | 0.84 | |
| Prompt Recovery AttackModel=Llama-3-Instruct, Editing Algorithm=ROME2026.02 | 40 | 60 | 80 | 0.87 | |
| Prompt Recovery AttackModel=Qwen-2.5-Instruct, Editing Algorithm=ROME2026.02 | 40 | 40 | 60 | 0.8 | |
| KSTERModel=GPT-J, N=50, Target Editing Method=MEMIT2026.02 | 33 | 64 | 85 | 0.82 | |
| KSTERModel=Llama-3-Instruct, N=100, Target Editing Method=MEMIT2026.02 | 33 | 57 | 76 | 0.83 | |
| KSTERModel=GPT-J, N=100, Target Editing Method=MEMIT2026.02 | 32 | 61 | 83 | 0.82 | |
| KSTERModel=GPT-J, N=10, Target Editing Method=MEMIT2026.02 | 28 | 70 | 86 | 0.83 | |
| Prompt Recovery AttackModel=GPT-J, Editing Algorithm=ROME2026.02 | 20 | 40 | 100 | 0.85 | |
| KSTERModel=Qwen-2.5-Instruct, N=50, Target Editing Method=MEMIT2026.02 | 13 | 34 | 60 | 0.79 | |
| KSTERModel=Qwen-2.5-Instruct, N=100, Target Editing Method=MEMIT2026.02 | 8 | 29 | 55 | 0.78 | |
| KSTERModel=Qwen-2.5-Instruct, N=10, Target Editing Method=MEMIT2026.02 | 4 | 42 | 60 | 0.81 |