Adversarial Attack on Cascaded LLMs on AGNews, Overruling, and Headline 3-layer Cascade
94Cascade AccuracyInitial cascade
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Initial cascadeSetting=202026.05 | 94 | 0.49 | 82.5 | 0.27 | 87.5 | 0.3 | 94.5 | 1.37 | 96.5 | 98 | |
| Initial cascadeSetting=232026.05 | 93.5 | 0.31 | 80 | 0.092 | 89 | 0.19 | 95 | 0.46 | 94 | 95.5 | |
| Initial cascadeSetting=212026.05 | 93 | 0.49 | 82.5 | 0.27 | 87.5 | 0.3 | 94.5 | 1.37 | 95 | 94 | |
| Initial cascadeSetting=222026.05 | 92 | 0.24 | 80 | 0.092 | 89 | 0.19 | 95 | 0.46 | 92 | 89.5 | |
| Initial cascadeSetting=12026.05 | 78 | 2,771.9 | 48 | 114 | — | — | 76 | 5,037.3 | 86.5 | — | |
| Joint-target AttackSetting=20, Transfer=(1)→(20)2026.05 | 77.5 | 0.42 | 75 | 0.29 | 79.5 | 0.32 | 87 | 1.42 | 71 | 76 | |
| Joint-target AttackSetting=21, Transfer=(1)→(21)2026.05 | 74 | 0.41 | 75 | 0.29 | 79.5 | 0.32 | 87 | 1.42 | 69 | 77.5 | |
| Joint-target AttackSetting=22, Transfer=(1)→(22)2026.05 | 72.5 | 0.18 | 72 | 0.097 | 77.5 | 0.2 | 90 | 0.5 | 75.5 | 80 | |
| Joint-target AttackSetting=23, Transfer=(1)→(23)2026.05 | 70.5 | 0.17 | 72 | 0.097 | 77.5 | 0.2 | 90 | 0.5 | 71 | 74 | |
| Joint-target AttackSetting=1, Transfer=(1)→ (1)2026.05 | 21 | 2,006.2 | 16 | 125.5 | — | — | 63 | 5,728.5 | 18 | — |