Toxicity Defense on Qwen2-VL
0.05Toxicity ScoreNullSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| NullSteerepsilon=unconstrained2026.03 | 0.05 | |
| ASTRAepsilon=unconstrained2026.03 | 0.07 | |
| NullSteerepsilon=64/2552026.03 | 1.98 | |
| ASTRAepsilon=64/2552026.03 | 2.39 | |
| NullSteerepsilon=32/2552026.03 | 3.51 | |
| ASTRAepsilon=32/2552026.03 | 5.45 | |
| NullSteerepsilon=16/2552026.03 | 13.36 | |
| ASTRAepsilon=16/2552026.03 | 15.52 | |
| JailGuardepsilon=32/2552026.03 | 24.68 | |
| Self-reminderepsilon=32/2552026.03 | 27.53 | |
| JailGuardepsilon=unconstrained2026.03 | 27.76 | |
| JailGuardepsilon=64/2552026.03 | 28.74 | |
| Self-reminderepsilon=unconstrained2026.03 | 29.09 | |
| JailGuardepsilon=16/2552026.03 | 29.37 | |
| Self-reminderepsilon=16/2552026.03 | 30.47 | |
| Self-reminderepsilon=64/2552026.03 | 32.84 | |
| Benign imageepsilon=16/2552026.03 | 38.52 | |
| Benign imageepsilon=32/2552026.03 | 38.52 | |
| Benign imageepsilon=64/2552026.03 | 38.52 | |
| Benign imageepsilon=unconstrained2026.03 | 38.52 | |
| Refusal Pairsepsilon=unconstrained2026.03 | 40.53 | |
| Refusal Pairsepsilon=16/2552026.03 | 46.14 | |
| Refusal Pairsepsilon=32/2552026.03 | 46.83 | |
| Refusal Pairsepsilon=64/2552026.03 | 46.83 | |
| ECSOepsilon=16/2552026.03 | 50.09 | |
| w/o defenseepsilon=16/2552026.03 | 50.5 | |
| ECSOepsilon=32/2552026.03 | 50.68 | |
| ECSOepsilon=unconstrained2026.03 | 51.57 | |
| w/o defenseepsilon=32/2552026.03 | 51.62 | |
| w/o defenseepsilon=unconstrained2026.03 | 53.43 | |
| w/o defenseepsilon=64/2552026.03 | 55.59 | |
| ECSOepsilon=64/2552026.03 | 56.08 | |
| Jailbreak Templatesepsilon=32/2552026.03 | 63.35 | |
| Jailbreak Templatesepsilon=16/2552026.03 | 66.74 | |
| Jailbreak Templatesepsilon=64/2552026.03 | 67.15 | |
| Jailbreak Templatesepsilon=unconstrained2026.03 | 68.29 |