List your company

Jailbreak Defense on jailbreak defense dataset

0ASRGradient-guided Token Masking
-3.890822.372148.63574.8979May 24, 2026

Evaluation results

MethodLinks
Gradient-guided Token Maskingε (perturbation limit)=162026.050————————
Gradient-guided Token Maskingε (perturbation limit)=642026.050————————
GUARDε (perturbation limit)=162026.050.91————————
GUARDε (perturbation limit)=642026.050.91————————
GUARDε (perturbation limit)=∞2026.050.91————————
Gradient-guided Token Maskingε (perturbation limit)=322026.050.91————————
Gradient-guided Token Maskingε (perturbation limit)=∞2026.050.91————————
GUARDε (perturbation limit)=322026.051.82————————
ECSOε (perturbation limit)=162026.056.36————————
ECSOε (perturbation limit)=322026.059.09————————
ECSOε (perturbation limit)=642026.0510————————
CIDERε (perturbation limit)=642026.0510.9————————
ECSOε (perturbation limit)=∞2026.0513.63————————
ASTRAε (perturbation limit)=162026.0513.64————————
ASTRAε (perturbation limit)=322026.0515.45————————
CIDERε (perturbation limit)=322026.0517.27————————
ASTRAε (perturbation limit)=∞2026.0520————————
CIDERε (perturbation limit)=162026.0520.91————————
ASTRAε (perturbation limit)=642026.0531.81————————
W/O DEFENSEε (perturbation limit)=642026.0568.18————————
W/O DEFENSEε (perturbation limit)=162026.0578.18————————
W/O DEFENSEε (perturbation limit)=322026.0582.73————————
CIDERε (perturbation limit)=∞2026.0594.55————————
W/O DEFENSEε (perturbation limit)=∞2026.0597.27————————
AlphaSteerBase Model=Llama-3.1-8B-Instruct, Defense Method=AlphaSteer2026.04—100965997899810091.29
AlphaSteerBase Model=Qwen2.5-7B-Instruct, Defense Method=AlphaSteer2026.04—100100979795959897.43
BiPOBase Model=Llama-3.1-8B-Instruct, Defense Method=BiPO2026.04—100100100100949910099
BiPOBase Model=Qwen2.5-7B-Instruct, Defense Method=BiPO2026.04—100100100100999910099.71
CASTBase Model=Llama-3.1-8B-Instruct, Defense Method=CAST2026.04—9251679981969683.14
CASTBase Model=Qwen2.5-7B-Instruct, Defense Method=CAST2026.04—25273396919910067.29
FineSteerBase Model=Llama-3.1-8B-Instruct, Defense Method=FineSteer2026.04—10010093100951009998.14
FineSteerBase Model=Qwen2.5-7B-Instruct, Defense Method=FineSteer2026.04—100100981009710010099.28
Jailbreak AntidoteBase Model=Llama-3.1-8B-Instruct, Defense Method=Jailbreak Antidote2026.04—10097010086936377
Jailbreak AntidoteBase Model=Qwen2.5-7B-Instruct, Defense Method=Jailbreak Antidote2026.04—91426905417347.14
Llama-3.1-8B-InstructBase Model=Llama-3.1-8B-Instruct, Defense Method=None2026.04—924805875452849.43
Qwen2.5-7B-InstructBase Model=Qwen2.5-7B-Instruct, Defense Method=None2026.04—2521227119420.57
SurgicalBase Model=Llama-3.1-8B-Instruct, Defense Method=Surgical2026.04—10076619888906782.86
SurgicalBase Model=Qwen2.5-7B-Instruct, Defense Method=Surgical2026.04—77816710079887071.71
TruthFlowBase Model=Llama-3.1-8B-Instruct, Defense Method=TruthFlow2026.04—9690479886917383
TruthFlowBase Model=Qwen2.5-7B-Instruct, Defense Method=TruthFlow2026.04—9910090100971009997.85