Question Answering on Natural Questions (NQ) (test) (Robust Accuracy)
68Robust AccuracySampling + MIS
Evaluation Results
| Method | Links | |
|---|---|---|
| Sampling + MISModel=GPT-4o-mini, Poison Position=50, k=502025.09 | 68 | |
| Sampling + MISModel=GPT-4o-mini, Poison Position=25, k=502025.09 | 67.6 | |
| Vanilla RAGModel=GPT-4o-mini, Poison Position=50, k=502025.09 | 63.6 | |
| Sampling + MISModel=GPT-4o-mini, Poison Position=1, k=502025.09 | 63.6 | |
| Vanilla RAGModel=GPT-4o-mini, Poison Position=25, k=502025.09 | 62.8 | |
| RobustRAGModel=GPT-4o-mini, Poison Position=25, k=502025.09 | 62.4 | |
| RobustRAGModel=GPT-4o-mini, Poison Position=50, k=502025.09 | 62.2 | |
| Sampling + MISModel=Mistral-7B, Poison Position=25, k=502025.09 | 61.3 | |
| InstructRAGModel=Mistral-7B, Poison Position=1, k=502025.09 | 60.8 | |
| Sampling + MISModel=Mistral-7B, Poison Position=50, k=502025.09 | 60.4 | |
| RobustRAGModel=GPT-4o-mini, Poison Position=1, k=502025.09 | 60.4 | |
| AstuteRAGModel=GPT-4o-mini, Poison Position=50, k=502025.09 | 60 | |
| InstructRAGModel=Mistral-7B, Poison Position=25, k=502025.09 | 59.8 | |
| Vanilla RAGModel=GPT-4o-mini, Poison Position=1, k=502025.09 | 59 | |
| InstructRAGModel=Llama3.2-3B, Poison Position=1, k=502025.09 | 58.8 | |
| AstuteRAGModel=GPT-4o-mini, Poison Position=25, k=502025.09 | 58.2 | |
| Sampling + MISModel=Llama3.2-3B, Poison Position=25, k=502025.09 | 57.6 | |
| InstructRAGModel=Mistral-7B, Poison Position=50, k=502025.09 | 57 | |
| Sampling + MISModel=Llama3.2-3B, Poison Position=50, k=502025.09 | 57 | |
| Vanilla RAGModel=Mistral-7B, Poison Position=1, k=502025.09 | 55.8 | |
| Sampling + MISModel=Llama3.2-3B, Poison Position=1, k=502025.09 | 55.2 | |
| InstructRAGModel=Llama3.2-3B, Poison Position=25, k=502025.09 | 55 | |
| Sampling + MISModel=Mistral-7B, Poison Position=1, k=502025.09 | 54.6 | |
| AstuteRAGModel=Llama3.2-3B, Poison Position=25, k=502025.09 | 54.2 | |
| AstuteRAGModel=Llama3.2-3B, Poison Position=1, k=502025.09 | 53.8 | |
| AstuteRAGModel=GPT-4o-mini, Poison Position=1, k=502025.09 | 53.8 | |
| InstructRAGModel=GPT-4o-mini, Poison Position=50, k=502025.09 | 53.8 | |
| RobustRAGModel=Llama3.2-3B, Poison Position=50, k=502025.09 | 53.2 | |
| InstructRAGModel=GPT-4o-mini, Poison Position=1, k=502025.09 | 52.8 | |
| InstructRAGModel=GPT-4o-mini, Poison Position=25, k=502025.09 | 52.8 | |
| RobustRAGModel=Llama3.2-3B, Poison Position=25, k=502025.09 | 52.2 | |
| Vanilla RAGModel=Llama3.2-3B, Poison Position=1, k=502025.09 | 52 | |
| RobustRAGModel=Llama3.2-3B, Poison Position=1, k=502025.09 | 51.4 | |
| AstuteRAGModel=Mistral-7B, Poison Position=25, k=502025.09 | 50.6 | |
| AstuteRAGModel=Llama3.2-3B, Poison Position=50, k=502025.09 | 50.2 | |
| AstuteRAGModel=Mistral-7B, Poison Position=1, k=502025.09 | 49.8 | |
| AstuteRAGModel=Mistral-7B, Poison Position=50, k=502025.09 | 48.2 | |
| InstructRAGModel=Llama3.2-3B, Poison Position=50, k=502025.09 | 47.6 | |
| RobustRAGModel=Mistral-7B, Poison Position=50, k=502025.09 | 47.2 | |
| RobustRAGModel=Mistral-7B, Poison Position=25, k=502025.09 | 46.4 | |
| RobustRAGModel=Mistral-7B, Poison Position=1, k=502025.09 | 46 | |
| Vanilla RAGModel=Mistral-7B, Poison Position=25, k=502025.09 | 43.4 | |
| Vanilla RAGModel=Llama3.2-3B, Poison Position=25, k=502025.09 | 43.4 | |
| Vanilla RAGModel=Llama3.2-3B, Poison Position=50, k=502025.09 | 42.2 | |
| Vanilla RAGModel=Mistral-7B, Poison Position=50, k=502025.09 | 38 |