Robustness against priming vulnerability on JBB-Behaviors (test)
0ASR (Guardrail Model)RA (ours)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| RA (ours)Backbone=LLaDA, Attack=No Attack2025.10 | 0 | 130 | |
| RA (ours)Backbone=LLaDA, Attack=Anchoring, t_inter=12025.10 | 0 | 0 | |
| RA (ours)Backbone=LLaDA, Attack=Anchoring, t_inter=42025.10 | 0 | 170 | |
| RA (ours)Backbone=LLaDA1.5, Attack=No Attack2025.10 | 0 | 430 | |
| RA (ours)Backbone=LLaDA, Attack=Anchoring, t_inter=82025.10 | 70 | 100 | |
| OriginalBackbone=LLaDA1.5, Attack=No Attack2025.10 | 70 | 500 | |
| RA (ours)Backbone=LLaDA, Attack=PAD2025.10 | 100 | 230 | |
| RA (ours)Backbone=LLaDA, Attack=Anchoring, t_inter=162025.10 | 200 | 670 | |
| RA (ours)Backbone=MMaDA, Attack=No Attack2025.10 | 400 | 570 | |
| RA (ours)Backbone=LLaDA, Attack=First-Step GCG2025.10 | 530 | 1,230 | |
| RA (ours)Backbone=LLaDA, Attack=DiJA2025.10 | 1,170 | 1,930 | |
| RA (ours)Backbone=LLaDA1.5, Attack=Anchoring, t_inter=322025.10 | 2,730 | 1,770 | |
| RA (ours)Backbone=LLaDA, Attack=Anchoring, t_inter=322025.10 | 2,830 | 1,430 | |
| RA (ours)Backbone=MMaDA, Attack=PAD2025.10 | 2,830 | 2,330 | |
| MOSABackbone=MMaDA, Attack=No Attack2025.10 | 2,870 | 2,600 | |
| RA (ours)Backbone=MMaDA, Attack=First-Step GCG2025.10 | 4,300 | 4,830 | |
| RA (ours)Backbone=MMaDA, Attack=DiJA2025.10 | 5,800 | 3,870 | |
| OriginalBackbone=MMaDA, Attack=No Attack2025.10 | 7,400 | 6,100 | |
| RA (ours)Backbone=MMaDA, Attack=Anchoring, t_inter=322025.10 | 7,630 | 3,830 | |
| RA w/o inter (ablation)Backbone=LLaDA, Attack=Anchoring, t_inter=322025.10 | 8,730 | 6,070 |