Indirect Prompt Injection Defense on Inj-TriviaQA
0.11Naive ASRStruQ
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| StruQBase Model=Llama3.1-8B-Instruct2025.04 | 0.11 | 0.56 | 4 | 86.44 | 70.33 | |
| Robustness via ReferencingBase Model=Llama3.1-8B-Instruct2025.04 | 0.11 | 1.44 | 0.22 | 0.67 | 0.56 | |
| StruQBase Model=Llama3-8B-Instruct2025.04 | 0.78 | 1.78 | 11.89 | 28.78 | 49.44 | |
| Robustness via ReferencingBase Model=Llama3-8B-Instruct2025.04 | 1.22 | 4 | 2.67 | 7 | 3.78 | |
| Robustness via ReferencingBase Model=Qwen2-7B-Instruct2025.04 | 1.56 | 5 | 1 | 4.33 | 6.56 | |
| SpotlightBase Model=Llama3-8B-Instruct2025.04 | 1.67 | 16.11 | 26.11 | 71.89 | 64.33 | |
| StruQBase Model=Qwen2-7B-Instruct2025.04 | 2.44 | 0.89 | 7.56 | 93.33 | 89.89 | |
| SandwichBase Model=Llama3.1-8B-Instruct2025.04 | 11.22 | 18.56 | 21.78 | 26 | 38.33 | |
| SpotlightBase Model=Llama3.1-8B-Instruct2025.04 | 11.44 | 31.22 | 34.56 | 85.33 | 91.44 | |
| SandwichBase Model=Llama3-8B-Instruct2025.04 | 13 | 23 | 33.56 | 31.89 | 37.56 | |
| SandwichBase Model=Qwen2-7B-Instruct2025.04 | 13.44 | 22.33 | 20 | 45.56 | 48.33 | |
| InstructionalBase Model=Llama3-8B-Instruct2025.04 | 18.56 | 38.56 | 50.89 | 75 | 62.11 | |
| SpotlightBase Model=Qwen2-7B-Instruct2025.04 | 19.89 | 40.89 | 27.56 | 98.56 | 94.67 | |
| NoneBase Model=Llama3-8B-Instruct2025.04 | 20.67 | 50.56 | 57.67 | 80.44 | 80.33 | |
| InstructionalBase Model=Llama3.1-8B-Instruct2025.04 | 20.89 | 51.67 | 51.33 | 83.78 | 82.67 | |
| ReminderBase Model=Llama3-8B-Instruct2025.04 | 23.11 | 47.89 | 55.11 | 60.56 | 60.33 | |
| NoneBase Model=Llama3.1-8B-Instruct2025.04 | 23.22 | 64.67 | 58 | 89.67 | 85.11 | |
| NoneBase Model=Qwen2-7B-Instruct2025.04 | 26.67 | 58.33 | 49.78 | 96 | 91.78 | |
| ReminderBase Model=Llama3.1-8B-Instruct2025.04 | 27.22 | 66.78 | 63 | 85.67 | 85 | |
| InstructionalBase Model=Qwen2-7B-Instruct2025.04 | 30 | 56.22 | 50.67 | 96.33 | 90.33 | |
| ReminderBase Model=Qwen2-7B-Instruct2025.04 | 35.11 | 67.67 | 53.22 | 96.67 | 86.11 |