List your company

Over-refusal Evaluation Suite (XSTest, WildJailbreak, WildGuard, OKTest, OR-Bench)

3.2XSTest Refusal Rate (%)Pure
2.566.8811.215.52Feb 14, 2026

Evaluation results

MethodLinks
PureModel Size=Qwen3-4B2026.023.201.95.3319.796.04
PureModel Size=Qwen3-8B2026.023.20.481.275.3315.095.07
PureModel Size=Qwen3-14B2026.023.61.431.487.3315.625.89
ShortModel Size=Qwen3-4B2026.024.41.432.2212.6741.9312.53
SafeChainModel Size=Qwen3-14B2026.024.42.383.929.3310.616.13
ShortModel Size=Qwen3-8B2026.024.83.334.2310.3340.8612.71
ShortModel Size=Qwen3-14B2026.025.23.814.0212.3346.8514.44
SafeChainModel Size=Qwen3-8B2026.026.403.177.679.255.3
IFPOModel Size=Qwen3-14B2026.027.21.433.69.677.285.84
IFPOModel Size=Qwen3-4B2026.027.62.384.349.675.996
IFPOModel Size=Qwen3-8B2026.027.62.384.349.675.994.33
STAR-1-mixModel Size=Qwen3-14B2026.0281.94.1312.6748.2915
SafeChainModel Size=Qwen3-4B2026.028.85.244.979.3312.818.23
AgentModel Size=Qwen3-4B2026.0210.46.676.881479.1523.42
STAR-1-mixModel Size=Qwen3-8B2026.0210.80.484.2310.3350.3415.24
BCModel Size=Qwen3-14B2026.0211.24.765.7114.6748.616.99
AgentModel Size=Qwen3-8B2026.0211.67.627.315.3371.7222.71
BCModel Size=Qwen3-8B2026.0212.41.94.4412.3343.8214.98
STAR-1Model Size=Qwen3-14B2026.0212.82.866.141755.1218.78
BCModel Size=Qwen3-4B2026.0213.63.814.6612.3347.2316.33
STAR-1Model Size=Qwen3-8B2026.02143.816.6715.6754.5118.93
AgentModel Size=Qwen3-14B2026.021419.0511.7516.6780.5928.41
STAR-1-mixModel Size=Qwen3-4B2026.02161.434.9711.3357.1618.18
STAR-1Model Size=Qwen3-4B2026.0219.21.97.4118.3363.3122.03