Toxicity Defense on MiniGPT-4
2.89Toxicity ScoreNullSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| NullSteerepsilon=unconstrained2026.03 | 2.89 | |
| NullSteerepsilon=64/2552026.03 | 3.26 | |
| ASTRAepsilon=unconstrained2026.03 | 4.48 | |
| ASTRAepsilon=64/2552026.03 | 4.51 | |
| NullSteerepsilon=32/2552026.03 | 6.45 | |
| ASTRAepsilon=32/2552026.03 | 8.84 | |
| NullSteerepsilon=16/2552026.03 | 9.52 | |
| ASTRAepsilon=16/2552026.03 | 11.3 | |
| JailGuardepsilon=16/2552026.03 | 16.51 | |
| JailGuardepsilon=32/2552026.03 | 18.93 | |
| Jailbreak Templatesepsilon=16/2552026.03 | 19.73 | |
| JailGuardepsilon=64/2552026.03 | 20.93 | |
| JailGuardepsilon=unconstrained2026.03 | 21.23 | |
| Jailbreak Templatesepsilon=unconstrained2026.03 | 22.78 | |
| Jailbreak Templatesepsilon=32/2552026.03 | 25.03 | |
| Refusal Pairsepsilon=16/2552026.03 | 25.76 | |
| Jailbreak Templatesepsilon=64/2552026.03 | 30.1 | |
| Refusal Pairsepsilon=32/2552026.03 | 30.28 | |
| Benign imageepsilon=16/2552026.03 | 30.65 | |
| Benign imageepsilon=32/2552026.03 | 30.65 | |
| Benign imageepsilon=64/2552026.03 | 30.65 | |
| Benign imageepsilon=unconstrained2026.03 | 30.65 | |
| Refusal Pairsepsilon=64/2552026.03 | 31.99 | |
| ECSOepsilon=32/2552026.03 | 32.42 | |
| ECSOepsilon=16/2552026.03 | 34.59 | |
| Refusal Pairsepsilon=unconstrained2026.03 | 35.71 | |
| ECSOepsilon=64/2552026.03 | 38.54 | |
| Self-reminderepsilon=16/2552026.03 | 38.97 | |
| w/o defenseepsilon=16/2552026.03 | 39.73 | |
| ECSOepsilon=unconstrained2026.03 | 42.86 | |
| Self-reminderepsilon=64/2552026.03 | 45.15 | |
| w/o defenseepsilon=32/2552026.03 | 48.52 | |
| Self-reminderepsilon=32/2552026.03 | 48.71 | |
| Self-reminderepsilon=unconstrained2026.03 | 50.12 | |
| w/o defenseepsilon=unconstrained2026.03 | 52.12 | |
| w/o defenseepsilon=64/2552026.03 | 54.7 |