List your company
ResearchFormerly WizWand

Harmful Request Defense on AdvBench

0ASRSafeKey
-1.49648.604318.70528.8057Dec 12, 2025Jan 13, 2026Feb 14, 2026Mar 19, 2026Apr 20, 2026May 22, 2026Jun 24, 2026
Updated 4mo ago

Evaluation results

MethodLinks
SafeKeyModel Scale=Qwen3-4B2026.010—
ReasoningGuardModel Scale=Qwen3-4B2026.010—
Self-ReminderModel Scale=Qwen3-4B2026.010—
Self-GuardModel Scale=Qwen3-4B2026.010—
STAR-1Model Scale=Qwen3-8B2026.010—
SafeKeyModel Scale=Qwen3-8B2026.010—
ReasoningGuardModel Scale=Qwen3-8B2026.010—
Self-ReminderModel Scale=Qwen3-8B2026.010—
Self-GuardModel Scale=Qwen3-8B2026.010—
STAR-1Model Scale=Qwen3-14B2026.010—
SafeKeyModel Scale=Qwen3-14B2026.010—
Alpha-steerModel Scale=Qwen3-14B2026.010—
ReasoningGuardModel Scale=Qwen3-14B2026.010—
Self-ReminderModel Scale=Qwen3-14B2026.010—
Self-GuardModel Scale=Qwen3-14B2026.010—
W-DOORBackbone=Qwen2.5-7B-Instruct2025.120—
GRPO+PolicyBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.060—
PolicyAlignBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.060—
SFTBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.060—
PolicyAlignBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.060—
ICLBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.060—
PolicyAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.060—
DPO-SBackbone=Llama3-8B-Instruct2025.120.06—
NSPOBackbone=Llama3-8B-Instruct2025.120.06—
ICLBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.060.19—
AlphaAlignBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=RL-based alignment2026.060.19—
NSPOBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.060.19—
BaseBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Base2026.060.19—
SFTBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.060.19—
AlphaAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=RL-based alignment2026.060.19—
STAR-1Model Scale=Qwen3-4B2026.010.2—
NSPOBackbone=Qwen2.5-7B-Instruct2025.120.29—
PeCANBackbone=Llama3-8B-Instruct2025.120.38—
ICLBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.060.38—
GRPO+PolicyBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.060.38—
NSPOBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.060.38—
GRPO+PolicyBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.060.38—
BaseBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Base2026.060.58—
InstructModel Scale=Qwen3-4B2026.010.6—
BFPOBackbone=Llama3-8B-Instruct2025.120.64—
W-DOORBackbone=Llama3-8B-Instruct2025.120.75—
BFPOBackbone=Qwen2.5-7B-Instruct2025.120.96—
Alpha-steerModel Scale=Qwen3-4B2026.011—
MoCANBackbone=Llama3-8B-Instruct2025.121.09—
MoCANBackbone=Qwen2.5-7B-Instruct2025.121.22—
Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct2025.121.28—
Llama3-8B-InstructBackbone=Llama3-8B-Instruct2025.121.36—
DPO-HBackbone=Qwen2.5-7B-Instruct2025.121.47—
InstructModel Scale=Qwen3-14B2026.011.5—
DPO-SBackbone=Qwen2.5-7B-Instruct2025.121.67—
InstructModel Scale=Qwen3-8B2026.011.7—
AlphaAlignBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=RL-based alignment2026.061.92—
PeCANBackbone=Qwen2.5-7B-Instruct2025.121.99—
Alpha-steerModel Scale=Qwen3-8B2026.012—
BaseBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Base2026.062.12—
SFTBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.062.12—
NSPOBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.062.12—
DPO-MixBackbone=Qwen2.5-7B-Instruct2025.122.24—
DPO-MixBackbone=Llama3-8B-Instruct2025.123.84—
SafeRLHFBackbone=Qwen2.5-7B-Instruct2025.126.64—
SafeRLHFBackbone=Llama3-8B-Instruct2025.1214.39—
SafeChainModel Scale=Qwen3-4B2026.0125.4—
SafeChainModel Scale=Qwen3-14B2026.0126—
SafeChainModel Scale=Qwen3-8B2026.0127.5—
DPO-HBackbone=Llama3-8B-Instruct2025.1237.41—