RAG Poisoning Attack Mitigation on RQA
1ASR (PIA)Reranking (ColBERTv2)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Reranking (ColBERTv2)LLM=GPT-4o2025.06 | 1 | 10.6 | 10 | |
| Reranking (ColBERTv2)LLM=Mistral-7B2025.06 | 2 | 5 | 8 | |
| Reranking (t5)LLM=Mistral-7B2025.06 | 2.4 | 8.6 | 7 | |
| AV Filter(α=∞)LLM=Mistral-7B2025.06 | 2.4 | 8.6 | 3.8 | |
| Reranking (ColBERTv2)LLM=Llama2-C2025.06 | 4 | 9 | 15 | |
| AV Filter(α=∞)LLM=Llama2-C2025.06 | 5 | 6.6 | 1.8 | |
| Reranking (t5)LLM=GPT-4o2025.06 | 7 | 11.2 | 10.6 | |
| AV Filter(α=∞)LLM=GPT-4o2025.06 | 7 | 9.2 | 4.6 | |
| Perplexity FilterLLM=Mistral-7B2025.06 | 14.4 | 25 | 31.8 | |
| Perplexity FilterLLM=GPT-4o2025.06 | 15 | 28.4 | 47.2 | |
| Vigilant PromptLLM=GPT-4o2025.06 | 15 | 23.8 | 38.2 | |
| Reranking (t5)LLM=Llama2-C2025.06 | 21 | 10 | 14 | |
| Vigilant PromptLLM=Mistral-7B2025.06 | 27.4 | 21 | 27.8 | |
| Perplexity FilterLLM=Llama2-C2025.06 | 42 | 17.6 | 41.6 | |
| Vigilant PromptLLM=Llama2-C2025.06 | 89.6 | 16.4 | 36.2 |