Toxicity Defense on LLaVA v1.5
22.35Toxicity ScoreNullSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| NullSteerepsilon=unconstrained2026.03 | 22.35 | |
| ASTRAepsilon=unconstrained2026.03 | 25.1 | |
| NullSteerepsilon=32/2552026.03 | 31.82 | |
| NullSteerepsilon=16/2552026.03 | 32.55 | |
| ASTRAepsilon=32/2552026.03 | 34.76 | |
| ASTRAepsilon=16/2552026.03 | 36.02 | |
| NullSteerepsilon=64/2552026.03 | 39.32 | |
| ASTRAepsilon=64/2552026.03 | 43.13 | |
| Jailbreak Templatesepsilon=32/2552026.03 | 50.21 | |
| Jailbreak Templatesepsilon=16/2552026.03 | 52.61 | |
| Jailbreak Templatesepsilon=unconstrained2026.03 | 54.9 | |
| Jailbreak Templatesepsilon=64/2552026.03 | 55.48 | |
| Refusal Pairsepsilon=64/2552026.03 | 60.36 | |
| Refusal Pairsepsilon=unconstrained2026.03 | 62.46 | |
| Refusal Pairsepsilon=16/2552026.03 | 66.72 | |
| ECSOepsilon=unconstrained2026.03 | 66.81 | |
| Refusal Pairsepsilon=32/2552026.03 | 66.82 | |
| ECSOepsilon=64/2552026.03 | 71.32 | |
| ECSOepsilon=32/2552026.03 | 73.14 | |
| JailGuardepsilon=unconstrained2026.03 | 73.76 | |
| ECSOepsilon=16/2552026.03 | 73.77 | |
| Benign imageepsilon=16/2552026.03 | 75 | |
| Benign imageepsilon=32/2552026.03 | 75 | |
| Benign imageepsilon=64/2552026.03 | 75 | |
| Benign imageepsilon=unconstrained2026.03 | 75 | |
| JailGuardepsilon=64/2552026.03 | 75.76 | |
| JailGuardepsilon=16/2552026.03 | 77.6 | |
| JailGuardepsilon=32/2552026.03 | 77.77 | |
| Self-reminderepsilon=unconstrained2026.03 | 80.93 | |
| w/o defenseepsilon=16/2552026.03 | 83.7 | |
| Self-reminderepsilon=16/2552026.03 | 83.92 | |
| Self-reminderepsilon=32/2552026.03 | 83.97 | |
| Self-reminderepsilon=64/2552026.03 | 84.19 | |
| w/o defenseepsilon=32/2552026.03 | 84.4 | |
| w/o defenseepsilon=unconstrained2026.03 | 85.44 | |
| w/o defenseepsilon=64/2552026.03 | 85.54 |