Share your thoughts, 1 month free Claude Pro on usSee more

Safety evaluation on UnsafeBench

89F1 Score

BLPO

Updated 23d ago

Evaluation Results

Method	Links
BLPO 2026.02		89	90
GuardReasoner-VL 2026.06		84.17	-
BLPO 2026.02		83	84
OPRO 2026.02		81	83
SingGuard-4B 2026.06		80.02	-
BLPO 2026.02		80	82
SingGuard-8B 2026.06		78.84	-
TextGrad 2026.02		78	81
SingGuard-2B 2026.06		77.82	-
Qwen3-VL-235B 2025.12		74.4	80.9
Qwen2.5-VL-72B 2025.12		73.2	77.7
OMNIGUARD-3B 2025.12		72.3	81.1
OMNIGUARD-7B 2025.12		72.2	81.1
Qwen3-VL-235B 2026.06		72.05	-
OPRO 2026.02		72	76
TextGrad 2026.02		71	75
APO-image 2026.02		70	76
APO-image 2026.02		70	71
SafeGuard-VL-RL 2026.06		69.06	-
OPRO 2026.02		68	70
TextGrad 2026.02		68	73
APO-image 2026.02		67	70
No Optim. 2026.02		65	67
No Optim. 2026.02		65	69
LlavaGuard-v1.2-7B 2025.12		63.4	77
No Optim. 2026.02		63	65
LLaVAShield 2026.06		60.87	-
LlavaGuard 2026.06		57.79	-
Qwen2.5-VL-7B 2025.12		55.7	40.1
GPT-4o 2025.12		55.2	74.7
GPT-5.1 2026.06		54.27	-
ShieldGemma-2 2026.06		53.19	-
Gemini3-Pro 2026.06		49.6	-
LlamaGuard4 2026.06		48.52	-
Qwen3-VL-8B 2026.06		48	-
Qwen2.5-Omni-7B 2025.12		47.3	67.8
LlamaGuard3-Vision 2026.06		43.23	-
Qwen3-VL-4B 2026.06		43.11	-
LLaMA Guard 3V 2025.12		0	61.9
GPT-4o 2026.07		-	76.34
Claude 3 Opus 2026.07		-	73.39
Gemini 1.5 Pro 2026.07		-	76.04
InternVL2-8B 2026.07		-	66.42
InternVL2.5-8B 2026.07		-	76.29
Qwen2.5-9B 2026.07		-	68.88
MiShield-8B 2026.07		-	77.86