List your company

Jailbreak Robustness on JBB-Behaviors

0.3ASR (PAIR, Guardrail Model)RA (ours)
-1.88812.88127.6542.419Oct 1, 2025

Evaluation results

MethodLinks
RA (ours)Base Model Architecture=LLaDA2025.100.310456029.367.3
MOSABase Model Architecture=LLaDA2025.10127.7655654.374
RA (ours)Base Model Architecture=LLaDA1.52025.101.3214556.738.758.3
MOSABase Model Architecture=LLaDA1.52025.10227.367.756.75774.3
RA w/o inter (ablation)Base Model Architecture=LLaDA1.52025.102.3366862.553.772.3
DPOBase Model Architecture=LLaDA2025.102.733.36957.360.784.7
DPOBase Model Architecture=LLaDA1.52025.10337.770.361.366.783.7
SFTBase Model Architecture=LLaDA2025.10439.368.7567085.3
RA w/o inter (ablation)Base Model Architecture=LLaDA2025.10425.768.360.352.368.7
OriginalBase Model Architecture=LLaDA1.52025.104.744.774.36368.391.7
OriginalBase Model Architecture=LLaDA2025.10541.777.3656889.7
SFTBase Model Architecture=LLaDA1.52025.105.346.37154.767.780.7
RA (ours)Base Model Architecture=MMaDA2025.101342.3506565.390.7
MOSABase Model Architecture=MMaDA2025.1017.746.368.365.755.383.7
RA w/o inter (ablation)Base Model Architecture=MMaDA2025.103252.4608566.386.7
DPOBase Model Architecture=MMaDA2025.1037.5607965.761.385.7
SFTBase Model Architecture=MMaDA2025.1047.770.386.578.55771.3
OriginalBase Model Architecture=MMaDA2025.10558181.785.76488.3