Downstream generation on HotpotQA
96.6ASR (Substring-based)Baseline
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| BaselinePoison Method=Joint-GCG, Top-k=52026.03 | 96.6 | 94.9 | 1.7 | |
| BaselinePoison Method=PoisonedRAG, Top-k=52026.03 | 94 | 96 | 1 | |
| BaselinePoison Method=Macro Avg, Top-k=52026.03 | 88.2 | 91.2 | 0.9 | |
| BaselinePoison Method=LIAR-RAG, Top-k=52026.03 | 73.9 | 82.6 | 0 | |
| RAGuardPoison Method=Joint-GCG, Top-k=52026.03 | 44.1 | 43 | 28.8 | |
| RAGuardPoison Method=Macro Avg, Top-k=52026.03 | 38 | 37.7 | 35.1 | |
| RAGuardPoison Method=PoisonedRAG, Top-k=52026.03 | 35 | 36 | 33 | |
| RAGuardPoison Method=LIAR-RAG, Top-k=52026.03 | 34.8 | 34 | 43.5 | |
| GMTPPoison Method=Joint-GCG, Top-k=52026.03 | 10.2 | 3.4 | 13.6 | |
| GMTPPoison Method=PoisonedRAG, Top-k=52026.03 | 9 | 3 | 10 | |
| GRADAPoison Method=Joint-GCG, Top-k=52026.03 | 8.5 | 11.9 | 23.7 | |
| GRADAPoison Method=PoisonedRAG, Top-k=52026.03 | 8 | 8 | 23 | |
| GMTPPoison Method=Macro Avg, Top-k=52026.03 | 6.4 | 2.1 | 16.6 | |
| ProGRankAttack Type=PoisonedRAG2026.03 | 6 | 1.8 | 28.4 | |
| GRADAPoison Method=Macro Avg, Top-k=52026.03 | 5.5 | 8.1 | 28.6 | |
| ProGRankPoison Method=PoisonedRAG, Top-k=52026.03 | 3.2 | 0 | 30 | |
| ProGRankAttack Type=Macro Avg2026.03 | 3 | 1.6 | 37.2 | |
| ProGRankAttack Type=Joint-GCG2026.03 | 2.8 | 2.9 | 45.1 | |
| ProGRankAttack Type=LIAR-RAG2026.03 | 2.6 | 0.8 | 38.1 | |
| ProGRankPoison Method=Macro Avg, Top-k=52026.03 | 1.1 | 0 | 39 | |
| GRADAPoison Method=LIAR-RAG, Top-k=52026.03 | 0 | 4.3 | 39.1 | |
| GMTPPoison Method=LIAR-RAG, Top-k=52026.03 | 0 | 0 | 26.1 | |
| ProGRankPoison Method=LIAR-RAG, Top-k=52026.03 | 0 | 0 | 40 | |
| ProGRankPoison Method=Joint-GCG, Top-k=52026.03 | 0 | 0 | 46.9 |