Discourse-level opinion manipulation on Black-box RAG DPR
23.28RASRDiscourseFlip
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DiscourseFlipLLM=Llama3.1, Target=PRO2026.05 | 23.28 | 13.97 | 64.65 | 7.58 | |
| DiscourseFlipLLM=Qwen3, Target=PRO2026.05 | 23.28 | 15.49 | 71.71 | 9.34 | |
| DiscourseFlipLLM=Llama3.1, Target=CON2026.05 | 21.02 | 17.32 | 78.35 | 10.92 | |
| DiscourseFlipLLM=Qwen3, Target=CON2026.05 | 21.02 | 17.54 | 79.04 | 9.82 | |
| Topic-FlipRAGLLM=Llama3.1, Target=CON2026.05 | 8.11 | 6.67 | 21.34 | 2.7 | |
| Topic-FlipRAGLLM=Qwen3, Target=CON2026.05 | 8.11 | 6.74 | 20.75 | 1.42 | |
| PoisonedRAGLLM=Llama3.1, Target=CON2026.05 | 7.68 | 9.1 | 28.11 | 4.08 | |
| PoisonedRAGLLM=Qwen3, Target=CON2026.05 | 7.68 | 8.83 | 25.17 | 1.98 | |
| Topic-FlipRAGLLM=Llama3.1, Target=PRO2026.05 | 4.71 | 5.12 | 0 | 1.15 | |
| Topic-FlipRAGLLM=Qwen3, Target=PRO2026.05 | 4.71 | 5.9 | 0 | 0.44 | |
| PoisonedRAGLLM=Llama3.1, Target=PRO2026.05 | 4.24 | 5.97 | 0 | 2.14 | |
| PoisonedRAGLLM=Qwen3, Target=PRO2026.05 | 4.24 | 6.76 | 0 | 1.93 | |
| Unic-RAGLLM=Llama3.1, Target=CON2026.05 | 4.14 | 7.67 | 26.1 | 1.95 | |
| Unic-RAGLLM=Qwen3, Target=CON2026.05 | 4.14 | 7.07 | 18.94 | 0.93 | |
| Unic-RAGLLM=Llama3.1, Target=PRO2026.05 | 2.26 | 6.61 | 1.49 | 1.03 | |
| Unic-RAGLLM=Qwen3, Target=PRO2026.05 | 2.26 | 7.22 | 8.84 | 1.69 |