Priming Attack Robustness on AdvBench No Attack (test)
0ASR (GPT-4o)MOSA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| MOSABase Model=LLaDA2025.10 | 0 | 0 | 0 | |
| RA w/o inter (ablation)Base Model=LLaDA2025.10 | 0 | 0 | 0 | |
| RA (ours)Base Model=LLaDA2025.10 | 0 | 0 | 0 | |
| MOSABase Model=LLaDA 1.52025.10 | 0 | 0 | 0 | |
| RA w/o inter (ablation)Base Model=LLaDA 1.52025.10 | 0 | 0 | 0 | |
| RA (ours)Base Model=LLaDA 1.52025.10 | 0 | 0 | 0 | |
| RA w/o inter (ablation)Base Model=MMaDA2025.10 | 0 | 0 | 0 | |
| RA (ours)Base Model=MMaDA2025.10 | 0 | 0 | 0 | |
| OriginalBase Model=LLaDA2025.10 | 2 | 2 | 4 | |
| OriginalBase Model=LLaDA 1.52025.10 | 2 | 2 | 6 | |
| DPOBase Model=LLaDA2025.10 | 4 | 24 | 10 | |
| SFTBase Model=LLaDA 1.52025.10 | 6 | 6 | 6 | |
| DPOBase Model=LLaDA 1.52025.10 | 6 | 30 | 8 | |
| MOSABase Model=MMaDA2025.10 | 8 | 12 | 12 | |
| SFTBase Model=LLaDA2025.10 | 16 | 6 | 4 | |
| DPOBase Model=MMaDA2025.10 | 34 | 64 | 57 | |
| SFTBase Model=MMaDA2025.10 | 60 | 54 | 36 | |
| OriginalBase Model=MMaDA2025.10 | 78 | 72 | 56 |