Knowledge corruption traceback on HotpotQA
98PrecisionRAGForensics
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RAGForensics2025.08 | 98 | 98 | 172.1 | |
| RAGOrigin2025.08 | 97 | 97 | 4.1 | |
| AttnTrace2025.08 | 95 | 95 | 8.1 | |
| RAGForensics2025.08 | 0.98 | 0.98 | 172.1 | |
| RAGOrigin2025.08 | 0.97 | 0.97 | 4.1 | |
| AttnTraceLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.95 | 0.95 | 8.1 | |
| AttnTrace2025.08 | 0.95 | 0.95 | 8.1 | |
| TracLLMLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.8 | 0.8 | 135.3 | |
| STCLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.77 | 0.77 | 2.1 | |
| DAALLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.75 | 0.75 | 0.8 | |
| ShapleyLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.75 | 0.75 | 145.5 | |
| LIME/CCLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.74 | 0.74 | 170.2 | |
| Self-CLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.68 | 0.68 | 0.9 | |
| AT2LLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.66 | 0.66 | 1.5 | |
| GradientLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.33 | 0.33 | 1.6 | |
| LOOLLM=Llama-3.1-8B-Instruct, Attack Type=Knowledge corruption2025.08 | 0.27 | 0.27 | 27.1 |