Response Harmfulness Detection on HarmBench Suite (HarmBench, SafeRLHF, BeaverTails, XSTest, WildGuard) benchmark leaderboard · SOTA2 Research

Response Harmfulness Detection on HarmBench Suite (HarmBench, SafeRLHF, BeaverTails, XSTest, WildGuard)

0.8333Macro Avg F1COLAGUARD

Evaluation Results

Method	Links
COLAGUARD2026.05		0.8333	—	—	—	0.8155
GuardReasoner2026.05		0.8313	—	—	—	0.8122
COLAGUARD2026.05		0.8256	—	—	—	0.8022
GuardReasoner2026.05		0.8249	—	—	—	0.808
WildGuard2026.05		0.81	—	—	—	0.7795
GuardReasoner2026.05		0.8078	—	—	—	0.7906
MD-Judge2026.05		0.8004	—	—	—	0.7867
GPT-4o+CoT2026.05		0.7434	—	—	—	0.7445
BeaverDam2026.05		0.7348	—	—	—	0.766
LLaMA Guard 22026.05		0.717	—	—	—	0.6699
LLaMA Guard 32026.05		0.7115	—	—	—	0.6497
o1-preview2026.05		0.6954	—	—	—	0.6922
HarmBench Mistral2026.05		0.6934	—	—	—	0.667
HarmBench LLaMA2026.05		0.6632	—	—	—	0.6549
GPT-4o2026.05		0.6588	—	—	—	0.6941
Aegis Guard Permissive2026.05		0.6146	—	—	—	0.6355
LLaMA Guard2026.05		0.6	—	—	—	0.5827
Aegis Guard Defensive2026.05		0.5962	—	—	—	0.6279
ShieldGemma2026.05		0.576	—	—	—	0.5567
QwQ-preview2026.05		0.5464	—	—	—	0.5773
ShieldGemma2026.05		0.3379	—	—	—	0.2724
COLAGUARD2026.05		—	318.9	13	2.5041	—
COLAGUARD2026.05		—	342	12.9	2.3601	—
GuardReasoner2026.05		—	3,801.03	281.96	0.2122	—
GuardReasoner2026.05		—	4,407.8	289.4	0.1838	—