Discourse-level opinion manipulation on Black-box RAG BGE
44.48RASRDiscourseFlip
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DiscourseFlipLLM=Llama3.1, Target=PRO2026.05 | 44.48 | 27.09 | 94.81 | 16.68 | |
| DiscourseFlipLLM=Qwen3, Target=PRO2026.05 | 44.48 | 26.85 | 94.62 | 7.48 | |
| DiscourseFlipLLM=Llama3.1, Target=CON2026.05 | 38.58 | 32.52 | 97.65 | 25.64 | |
| DiscourseFlipLLM=Qwen3, Target=CON2026.05 | 38.58 | 30.34 | 96.77 | 21.12 | |
| Topic-FlipRAGLLM=Llama3.1, Target=CON2026.05 | 24.06 | 12.77 | 85.91 | 11.35 | |
| Topic-FlipRAGLLM=Qwen3, Target=CON2026.05 | 24.06 | 11.19 | 76.25 | 6.57 | |
| Topic-FlipRAGLLM=Llama3.1, Target=PRO2026.05 | 20.35 | 8.97 | 60.54 | 4.77 | |
| Topic-FlipRAGLLM=Qwen3, Target=PRO2026.05 | 20.35 | 8.49 | 51.69 | 3.29 | |
| PoisonedRAGLLM=Llama3.1, Target=CON2026.05 | 13.95 | 14.6 | 67.78 | 7.65 | |
| PoisonedRAGLLM=Qwen3, Target=CON2026.05 | 13.95 | 12.6 | 56.83 | 4.89 | |
| Unic-RAGLLM=Llama3.1, Target=PRO2026.05 | 12.01 | 10.99 | 65.44 | 7.24 | |
| Unic-RAGLLM=Qwen3, Target=PRO2026.05 | 12.01 | 8.93 | 46.34 | 3.44 | |
| Unic-RAGLLM=Llama3.1, Target=CON2026.05 | 9.67 | 9.74 | 57.37 | 5.09 | |
| Unic-RAGLLM=Qwen3, Target=CON2026.05 | 9.67 | 10.2 | 56.29 | 5.68 | |
| PoisonedRAGLLM=Llama3.1, Target=PRO2026.05 | 8.8 | 8.51 | 21.53 | 2.64 | |
| PoisonedRAGLLM=Qwen3, Target=PRO2026.05 | 8.8 | 7.96 | 15 | 1.05 |