Discourse-level opinion manipulation on Black-box RAG Qwen-Embedding
16.33RASRDiscourseFlip
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DiscourseFlipLLM=Llama3.1, Target=PRO2026.05 | 16.33 | 12.12 | 53.7 | 5.79 | |
| DiscourseFlipLLM=Qwen3, Target=PRO2026.05 | 16.33 | 13.79 | 63.76 | 7.71 | |
| DiscourseFlipLLM=Llama3.1, Target=CON2026.05 | 13.42 | 13.66 | 63.03 | 8.25 | |
| DiscourseFlipLLM=Qwen3, Target=CON2026.05 | 13.42 | 13.71 | 63.3 | 5.86 | |
| Unic-RAGLLM=Llama3.1, Target=PRO2026.05 | 1.97 | 7.27 | 8.84 | 0.7 | |
| Unic-RAGLLM=Qwen3, Target=PRO2026.05 | 1.97 | 7.34 | 11.09 | 1.1 | |
| Unic-RAGLLM=Llama3.1, Target=CON2026.05 | 1.8 | 6.62 | 4.16 | 0.6 | |
| Unic-RAGLLM=Qwen3, Target=CON2026.05 | 1.8 | 6.61 | 4.16 | 0.37 | |
| PoisonedRAGLLM=Llama3.1, Target=CON2026.05 | 1.42 | 4.83 | 0 | 0.32 | |
| PoisonedRAGLLM=Qwen3, Target=CON2026.05 | 1.42 | 5.92 | 0 | 0.11 | |
| Topic-FlipRAGLLM=Llama3.1, Target=CON2026.05 | 1.03 | 4.59 | 0 | 0.65 | |
| Topic-FlipRAGLLM=Qwen3, Target=CON2026.05 | 1.03 | 5.65 | 0 | 0.05 | |
| PoisonedRAGLLM=Llama3.1, Target=PRO2026.05 | 0.8 | 4.48 | 0 | 0.48 | |
| PoisonedRAGLLM=Qwen3, Target=PRO2026.05 | 0.8 | 5.84 | 0 | 0.73 | |
| Topic-FlipRAGLLM=Llama3.1, Target=PRO2026.05 | 0.74 | 4.52 | 0 | 0.07 | |
| Topic-FlipRAGLLM=Qwen3, Target=PRO2026.05 | 0.74 | 5.75 | 0 | 0.19 |