RAG Poisoning Attack Mitigation on RQA-MC
64ASR (PIA)Vigilant Prompt
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Vigilant PromptLLM=Llama2-C2025.06 | 64 | 29.4 | 49.8 | |
| Perplexity FilterLLM=Llama2-C2025.06 | 34.8 | 28.8 | 43.8 | |
| Vigilant PromptLLM=Mistral-7B2025.06 | 32.2 | 28.6 | 54.2 | |
| Reranking (t5)LLM=Llama2-C2025.06 | 18 | 17 | 16 | |
| Perplexity FilterLLM=Mistral-7B2025.06 | 17.6 | 31.6 | 55.2 | |
| AV Filter(α=∞)LLM=Llama2-C2025.06 | 16.8 | 12.6 | 8.2 | |
| Vigilant PromptLLM=GPT-4o2025.06 | 16.2 | 23.6 | 34.6 | |
| Perplexity FilterLLM=GPT-4o2025.06 | 7.6 | 23.2 | 37 | |
| AV Filter(α=∞)LLM=Mistral-7B2025.06 | 7.2 | 8.4 | 3.4 | |
| Reranking (ColBERTv2)LLM=Llama2-C2025.06 | 6 | 13 | 15 | |
| AV Filter(α=∞)LLM=GPT-4o2025.06 | 5.4 | 6.8 | 2.4 | |
| Reranking (t5)LLM=Mistral-7B2025.06 | 5 | 15 | 12 | |
| Reranking (ColBERTv2)LLM=Mistral-7B2025.06 | 3 | 10 | 14 | |
| Reranking (t5)LLM=GPT-4o2025.06 | 2 | 7.2 | 4.6 | |
| Reranking (ColBERTv2)LLM=GPT-4o2025.06 | 0.4 | 6 | 4.8 |