List your company
ResearchFormerly WizWand

Safety Evaluation on JailbreakBench

0Harmful RateSafeReAct
-1.8810.8123.536.19Mar 10, 2026
Updated 7mo ago

Evaluation results

MethodLinks
SafeReActModel=R1-8B2026.030——
SafeReActModel=OT-7B2026.030——
SafeReActModel=R1-14B2026.030——
SafeReActModel=R1-7B2026.031——
Circuit-BreakerModel=R1-8B2026.032——
SafeChainModel=R1-14B2026.037——
SafeChainModel=R1-7B2026.038——
SafeChainModel=OT-7B2026.039——
Circuit-BreakerModel=R1-14B2026.0317——
OriginalModel=R1-14B2026.0323——
SafeChainModel=R1-8B2026.0327——
OriginalModel=R1-8B2026.0333——
Circuit-BreakerModel=OT-7B2026.0341——
OriginalModel=OT-7B2026.0343——
OriginalModel=R1-7B2026.0345——
Circuit-BreakerModel=R1-7B2026.0347——
ActorAttackToken Cost=8K2026.03——37.5
MDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=China-centric, Top-k Layers=4, Coeff=-0.5, Reposition=false2025.12—81—
MDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=Extended, Top-k Layers=4, Coeff=-1, Reposition=false2025.12—92—
OriginalBackbone=Llama-3-8B-UltraMedical2026.03—66—
Qwen3-Next-4B-InstructN. Params=4B, MoE=false, Reasoning=false2025.12—99—
Qwen3-Next-4B-ThinkingN. Params=4B, MoE=false, Reasoning=true2025.12—98—
Qwen3-Next-80B-A3B-ThinkingN. Params=80B, MoE=true, Reasoning=true2025.12—99—
RMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=China-centric, Top-k Layers=4, Coeff=-0.75, Reposition=false2025.12—82—
RMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=Extended, Top-k Layers=4, Coeff=-2, Reposition=false2025.12—86.87—
SafeReActBackbone=Llama-3-8B-UltraMedical2026.03—6—
STARToken Cost=37K2026.03——94
WMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=China-centric, Top-k Layers=8, Coeff=-0.25, Reposition=false2025.12—91—
WMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=Extended, Top-k Layers=8, Coeff=-0.25, Reposition=false2025.12—84—
WRMDBase Model=Qwen3-Next-4B-Instruct, N. Params=4B, MoE=false, Reasoning=false, Train Dataset=Extended, Top-k Layers=4, Coeff=-0.5, Reposition=false2025.12—55—
WRMDBase Model=Qwen3-Next-4B-Thinking, N. Params=4B, MoE=false, Reasoning=true, Train Dataset=Extended, Top-k Layers=2, Coeff=-2, Reposition=false2025.12—78—
WRMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=China-centric, Top-k Layers=8, Coeff=-0.25, Reposition=false2025.12—90—
WRMDBase Model=Qwen3-Next-80B-A3B-Thinking, N. Params=80B, MoE=true, Reasoning=true, Train Dataset=Extended, Top-k Layers=16, Coeff=-0.15, Reposition=false2025.12—99—
X-TeamingToken Cost=30K2026.03——85.5