List your company
ResearchFormerly WizWand

Multimodal Safety Evaluation on MM-SafetyBench

2.73Safety ScoreQwen2.5-VL-3B + SFT+DPO+GRPO
-0.05720.66641.392.1136Mar 3, 2026Mar 21, 2026Apr 9, 2026Apr 28, 2026May 16, 2026Jun 4, 2026Jun 23, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen2.5-VL-3B + SFT+DPO+GRPOBase Model=Qwen2.5-VL-3B, Stage=SFT+DPO+GRPO2026.032.732.622.93
Qwen2.5-VL-7B + TISBase Model=Qwen2.5-VL-7B, Enhancement=TIS2026.032.732.422.75
Qwen2.5-VL-7B + SFT+DPO+GRPOBase Model=Qwen2.5-VL-7B, Stage=SFT+DPO+GRPO2026.032.732.72.94
Qwen2.5-VL-3B + SFT+GRPOBase Model=Qwen2.5-VL-3B, Stage=SFT+GRPO2026.032.712.642.89
Qwen2.5-VL-7B + SaFeR-VLMBase Model=Qwen2.5-VL-7B, Method=SaFeR-VLM2026.032.72.132.42
Qwen2.5-VL-7B + SFT+GRPOBase Model=Qwen2.5-VL-7B, Stage=SFT+GRPO2026.032.72.652.93
Qwen2.5-VL-3B + SaFeR-VLMBase Model=Qwen2.5-VL-3B, Method=SaFeR-VLM2026.032.642.242.33
Qwen2.5-VL-3B + TISBase Model=Qwen2.5-VL-3B, Enhancement=TIS2026.032.622.172.55
Qwen2.5-VL-7B + VLGuardBase Model=Qwen2.5-VL-7B, Enhancement=VLGuard2026.032.61.612.01
Qwen2.5-VL-7B + SFT+DPOBase Model=Qwen2.5-VL-7B, Stage=SFT+DPO2026.032.552.532.93
Qwen2.5-VL-3B + VLGuardBase Model=Qwen2.5-VL-3B, Enhancement=VLGuard2026.032.541.61.92
Qwen2.5-VL-7B + SFTBase Model=Qwen2.5-VL-7B, Stage=SFT2026.032.52.282.84
Qwen2.5-VL-3B + SFT+DPOBase Model=Qwen2.5-VL-3B, Stage=SFT+DPO2026.032.422.342.8
Qwen2.5-VL-3B + SPA-VLBase Model=Qwen2.5-VL-3B, Enhancement=SPA-VL2026.032.342.462.08
Qwen2.5-VL-3B + SFTBase Model=Qwen2.5-VL-3B, Stage=SFT2026.032.312.292.73
Qwen2.5-VL-7B + SIABase Model=Qwen2.5-VL-7B, Enhancement=SIA2026.032.12.171.97
Qwen2.5-VL-7B + SPA-VLBase Model=Qwen2.5-VL-7B, Enhancement=SPA-VL2026.032.082.352.19
Qwen2.5-VL-7B + ECSOBase Model=Qwen2.5-VL-7B, Enhancement=ECSO2026.031.332.312.02
Qwen2.5-VL-7BBase Model=Qwen2.5-VL-7B2026.031.312.311.98
Yuvion VLModel Scale=8B, Decoding Mode=Instruct2026.061——
Yuvion VLModel Scale=8B, Decoding Mode=Reasoning2026.061——
Yuvion VLModel Scale=32B, Decoding Mode=Instruct2026.061——
Yuvion VLModel Scale=32B, Decoding Mode=Reasoning2026.060.996——
Qwen2.5-VL-3B + SIABase Model=Qwen2.5-VL-3B, Enhancement=SIA2026.030.971.761.37
Qwen3-VLModel Scale=32B, Decoding Mode=Instruct2026.060.747——
Qwen3.5-PlusModel Scale=397B2026.060.635——
GPT-5.4Model Scale=≈1.5T2026.060.627——
Qwen3-VLModel Scale=8B, Decoding Mode=Instruct2026.060.607——
K2.5Model Scale=≈1T2026.060.579——
Qwen3-VLModel Scale=32B, Decoding Mode=Thinking2026.060.575——
Opus-4.6Model Scale=≈800B2026.060.571——
Qwen3-VLModel Scale=8B, Decoding Mode=Thinking2026.060.465——
Qwen2.5-VL-3B + ECSOBase Model=Qwen2.5-VL-3B, Enhancement=ECSO2026.030.42.251.62
Qwen2.5-VL-3BBase Model=Qwen2.5-VL-3B2026.030.052.371.7