Question Answering on NQ Poison Attack (test)
50.4Attack Success RateVanilla
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VanillaLLM=Mistral-7B2025.06 | 50.4 | — | 23 | |
| VanillaLLM=Llama 2-C2025.06 | 48 | — | 26.2 | |
| VanillaLLM=GPT-4o2025.06 | 34.6 | — | 28.6 | |
| RobustRAGPos=1, top-k=502026.05 | 34 | 61.8 | — | |
| AV FilterLLM=Llama 2-C, alpha=∞2025.06 | 25.8 | — | 32.4 | |
| Vanilla RAGPos=1, top-k=502026.05 | 23 | 62.2 | — | |
| DecodingLLM=Llama 2-C2025.06 | 23 | — | 26 | |
| DecodingLLM=Mistral-7B2025.06 | 20 | — | 43 | |
| AV FilterLLM=Llama 2-C, alpha=52025.06 | 17.6 | — | 37.2 | |
| InstructRAGPos=1, top-k=502026.05 | 16.8 | 41.2 | — | |
| AV FilterLLM=Llama 2-C, alpha=102025.06 | 15.6 | — | 36.4 | |
| AV FilterLLM=Mistral-7B, alpha=52025.06 | 14.6 | — | 43 | |
| ReliabilityRAGPos=1, top-k=502026.05 | 13.4 | 64.4 | — | |
| AV FilterLLM=Mistral-7B, alpha=102025.06 | 13.2 | — | 44 | |
| AV FilterLLM=Mistral-7B, alpha=∞2025.06 | 12.4 | — | 42.2 | |
| KeywordLLM=Mistral-7B2025.06 | 10 | — | 42 | |
| KeywordLLM=Llama 2-C2025.06 | 9 | — | 37 | |
| RADARPos=1, top-k=502026.05 | 7 | 65.4 | — | |
| AV FilterLLM=GPT-4o, alpha=52025.06 | 6.8 | — | 36.8 | |
| Vanilla RAGPos=25, top-k=502026.05 | 6.6 | 67.5 | — | |
| Vanilla RAGPos=50, top-k=502026.05 | 5.8 | 67.4 | — | |
| AV FilterLLM=GPT-4o, alpha=∞2025.06 | 5.4 | — | 38.6 | |
| KeywordLLM=GPT-4o2025.06 | 4.8 | — | 53 | |
| RADARPos=25, top-k=502026.05 | 4.6 | 69.6 | — | |
| InstructRAGPos=50, top-k=502026.05 | 4.2 | 56.4 | — | |
| RADARPos=50, top-k=502026.05 | 4.2 | 69.2 | — | |
| AV FilterLLM=GPT-4o, alpha=102025.06 | 4 | — | 35.8 | |
| ReliabilityRAGPos=25, top-k=502026.05 | 3.4 | 69.2 | — | |
| ReliabilityRAGPos=50, top-k=502026.05 | 3.4 | 68.4 | — | |
| AstuteRAGPos=1, top-k=502026.05 | 3.2 | 25.2 | — | |
| InstructRAGPos=25, top-k=502026.05 | 3.2 | 55.2 | — | |
| RobustRAGPos=25, top-k=502026.05 | 2.6 | 66.4 | — | |
| RobustRAGPos=50, top-k=502026.05 | 2.6 | 66.6 | — | |
| AstuteRAGPos=25, top-k=502026.05 | 1.4 | 33.4 | — | |
| AstuteRAGPos=50, top-k=502026.05 | 1.4 | 31.6 | — |