Prompt recovery attack on CounterFact
60Top-1 AccuracyPrompt Recovery Attack
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Prompt Recovery AttackModel=GPT-J, Editing Algorithm=ROME2026.02 | 60 | 80 | 100 | 86 | |
| Prompt Recovery AttackModel=Llama-3-Instruct, Editing Algorithm=ROME2026.02 | 60 | 80 | 100 | 89 | |
| KSTERModel=GPT-J, N=10, Target Editing Method=MEMIT2026.02 | 57 | 86 | 96 | 88 | |
| KSTERModel=Llama-3-Instruct, N=10, Target Editing Method=MEMIT2026.02 | 53 | 90 | 96 | 88 | |
| KSTERModel=GPT-J, N=50, Target Editing Method=MEMIT2026.02 | 51 | 90 | 99 | 88 | |
| KSTERModel=Llama-3-Instruct, N=100, Target Editing Method=MEMIT2026.02 | 51 | 81 | 94 | 88 | |
| KSTERModel=GPT-J, N=100, Target Editing Method=MEMIT2026.02 | 50 | 92 | 99 | 88 | |
| KSTERModel=Llama-3-Instruct, N=50, Target Editing Method=MEMIT2026.02 | 49 | 82 | 94 | 88 | |
| KSTERModel=Qwen-2.5-Instruct, N=100, Target Editing Method=MEMIT2026.02 | 47 | 86 | 97 | 86 | |
| KSTERModel=Qwen-2.5-Instruct, N=50, Target Editing Method=MEMIT2026.02 | 42 | 81 | 95 | 86 | |
| KSTERModel=Qwen-2.5-Instruct, N=10, Target Editing Method=MEMIT2026.02 | 35 | 83 | 96 | 85 | |
| Prompt Recovery AttackModel=Qwen-2.5-Instruct, Editing Algorithm=ROME2026.02 | 20 | 40 | 100 | 84 |