List your company

Toxicity Defense on Qwen2-VL

0.05Toxicity ScoreNullSteer
-2.679615.745234.1752.5948Mar 23, 2026

Evaluation results

MethodLinks
NullSteerepsilon=unconstrained2026.030.05
ASTRAepsilon=unconstrained2026.030.07
NullSteerepsilon=64/2552026.031.98
ASTRAepsilon=64/2552026.032.39
NullSteerepsilon=32/2552026.033.51
ASTRAepsilon=32/2552026.035.45
NullSteerepsilon=16/2552026.0313.36
ASTRAepsilon=16/2552026.0315.52
JailGuardepsilon=32/2552026.0324.68
Self-reminderepsilon=32/2552026.0327.53
JailGuardepsilon=unconstrained2026.0327.76
JailGuardepsilon=64/2552026.0328.74
Self-reminderepsilon=unconstrained2026.0329.09
JailGuardepsilon=16/2552026.0329.37
Self-reminderepsilon=16/2552026.0330.47
Self-reminderepsilon=64/2552026.0332.84
Benign imageepsilon=16/2552026.0338.52
Benign imageepsilon=32/2552026.0338.52
Benign imageepsilon=64/2552026.0338.52
Benign imageepsilon=unconstrained2026.0338.52
Refusal Pairsepsilon=unconstrained2026.0340.53
Refusal Pairsepsilon=16/2552026.0346.14
Refusal Pairsepsilon=32/2552026.0346.83
Refusal Pairsepsilon=64/2552026.0346.83
ECSOepsilon=16/2552026.0350.09
w/o defenseepsilon=16/2552026.0350.5
ECSOepsilon=32/2552026.0350.68
ECSOepsilon=unconstrained2026.0351.57
w/o defenseepsilon=32/2552026.0351.62
w/o defenseepsilon=unconstrained2026.0353.43
w/o defenseepsilon=64/2552026.0355.59
ECSOepsilon=64/2552026.0356.08
Jailbreak Templatesepsilon=32/2552026.0363.35
Jailbreak Templatesepsilon=16/2552026.0366.74
Jailbreak Templatesepsilon=64/2552026.0367.15
Jailbreak Templatesepsilon=unconstrained2026.0368.29