Jailbreak Robustness on JBB-Behaviors (test)
0ASRMOSA
Evaluation Results
| Method | Links | |
|---|---|---|
| MOSABackbone=LLaDA, Attack Type=No Attack2025.10 | 0 | |
| RA (ours)Backbone=LLaDA, Attack Type=No Attack2025.10 | 0 | |
| RA (ours)Backbone=LLaDA1.5, Attack Type=Anchoring, t_inter=42025.10 | 0.7 | |
| OriginalBackbone=LLaDA1.5, Attack Type=No Attack2025.10 | 1 | |
| RA (ours)Backbone=LLaDA1.5, Attack Type=PAD2025.10 | 1 | |
| RA (ours)Backbone=LLaDA, Attack Type=Anchoring, t_inter=42025.10 | 1.3 | |
| RA w/o inter (ablation)Backbone=LLaDA, Attack Type=No Attack2025.10 | 1.7 | |
| OriginalBackbone=LLaDA, Attack Type=No Attack2025.10 | 2 | |
| RA w/o inter (ablation)Backbone=MMaDA, Attack Type=No Attack2025.10 | 2 | |
| DPOBackbone=LLaDA, Attack Type=No Attack2025.10 | 4.3 | |
| SFTBackbone=LLaDA, Attack Type=No Attack2025.10 | 8.3 | |
| RA (ours)Backbone=LLaDA, Attack Type=First-Step GCG2025.10 | 11.3 | |
| RA (ours)Backbone=MMaDA, Attack Type=Anchoring, t_inter=42025.10 | 13 | |
| RA (ours)Backbone=LLaDA1.5, Attack Type=First-Step GCG2025.10 | 15 | |
| OriginalBackbone=LLaDA, Attack Type=Anchoring, t_inter=12025.10 | 17.3 | |
| RA (ours)Backbone=LLaDA1.5, Attack Type=Anchoring, t_inter=322025.10 | 43 | |
| RA (ours)Backbone=MMaDA, Attack Type=First-Step GCG2025.10 | 45.7 | |
| RA (ours)Backbone=LLaDA, Attack Type=Anchoring, t_inter=322025.10 | 50.7 | |
| OriginalBackbone=LLaDA, Attack Type=First-Step GCG2025.10 | 58 | |
| OriginalBackbone=LLaDA, Attack Type=PAD2025.10 | 67.3 | |
| RA (ours)Backbone=MMaDA, Attack Type=DiJA2025.10 | 70 | |
| RA (ours)Backbone=MMaDA, Attack Type=Anchoring, t_inter=322025.10 | 79.3 | |
| OriginalBackbone=MMaDA, Attack Type=No Attack2025.10 | 79.7 | |
| OriginalBackbone=LLaDA, Attack Type=DiJA2025.10 | 92 |