RAG Robustness on NQ
57.8Robustness RACC (Paradox)AV Filter
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| AV FilterLLM=GPT-4o, Defense=AV Filter, alpha=infinity2025.06 | 57.8 | 3.2 | 60.8 | 0 | 59.8 | 3.4 | |
| AV FilterLLM=Mistral-7B, Defense=AV Filter, alpha=infinity2025.06 | 56.4 | 4 | 56.2 | 2 | 51.2 | 8 | |
| AV FilterLLM=GPT-4o, Defense=AV Filter, alpha=102025.06 | 54.6 | 6.6 | 58 | 2 | 57 | 5.4 | |
| AV FilterLLM=GPT-4o, Defense=AV Filter, alpha=52025.06 | 54 | 6.8 | 57 | 2.4 | 56.8 | 6 | |
| AV FilterLLM=Mistral-7B, Defense=AV Filter, alpha=102025.06 | 52 | 7.6 | 54.8 | 3.8 | 51 | 9.2 | |
| AV FilterLLM=Llama2-C, Defense=AV Filter, alpha=infinity2025.06 | 51.8 | 1.6 | 54.6 | 0.2 | 46.2 | 22.4 | |
| AV FilterLLM=Mistral-7B, Defense=AV Filter, alpha=52025.06 | 50.6 | 7.4 | 53.2 | 4.4 | 48.4 | 9.2 | |
| AV FilterLLM=Llama2-C, Defense=AV Filter, alpha=102025.06 | 48.4 | 10.6 | 52.2 | 4.4 | 45.6 | 18 | |
| AV FilterLLM=Llama2-C, Defense=AV Filter, alpha=52025.06 | 46.2 | 10.6 | 52 | 4.6 | 42.8 | 19.8 | |
| Vanilla RAGLLM=GPT-4o, Defense=Vanilla2025.06 | 37.4 | 22.4 | 59.4 | 2.8 | 56.8 | 2.4 | |
| Vanilla RAGLLM=Llama2-C, Defense=Vanilla2025.06 | 32 | 28.6 | 59.6 | 4.8 | 44.2 | 30.6 | |
| Vanilla RAGLLM=Mistral-7B, Defense=Vanilla2025.06 | 29.2 | 24.2 | 56.8 | 5 | 40.4 | 15.8 |