Toxic Text Generation on RealToxicityPrompts malicious
14.8Attack Success Rate (ASR)SAD
Evaluation Results
| Method | Links | |
|---|---|---|
| SADModel=LLaDA, Negation Set=RTP2026.04 | 14.8 | |
| SADModel=LLaDA, Negation Set=BeaverTails2026.04 | 16 | |
| SADModel=LLaDA, Negation Set=ToxiGen2026.04 | 17.6 | |
| BaseModel=LLaDA, Negation Set=–2026.04 | 19.8 | |
| SADModel=MDLM, Negation Set=RTP2026.04 | 32.6 | |
| SADModel=MDLM, Negation Set=ToxiGen2026.04 | 32.8 | |
| SADModel=MDLM, Negation Set=BeaverTails2026.04 | 33.4 | |
| BaseModel=MDLM, Negation Set=–2026.04 | 38.4 |