List your company

Safety Evaluation on MM-SafetyBench

0Average ASRRAI
-3.98422.90849.876.692Oct 20, 2025Nov 25, 2025Jan 1, 2026Feb 6, 2026Mar 15, 2026Apr 20, 2026May 27, 2026

Evaluation results

MethodLinks
RAIModel=Qwen3-VL-8B2026.020000000000000000000————————
CBModel=LLaVA-v1.6-Mistral-7B2026.040——————————————————————————
DefenSeeTarget Model=Llava 22025.120.03——————————————————————————
AdaShield-ATarget Model=Llava 22025.120.12——————————————————————————
AdaShield-STarget Model=Llava 22025.120.12——————————————————————————
CrossGuardCategory=MLLM Guardrails2025.100.38——————————————————————————
SafeSteerModels=Qwen3-VL2026.050.69——————————————————————————
MLLM-ProtectorModels=Qwen3-VL2026.051.32——————————————————————————
DefenSeeTarget Model=MiniGPT42025.121.7——————————————————————————
SafeSteerModels=Qwen2.5-VL2026.051.75——————————————————————————
BeamModel=LLaVA-v1.6-Mistral-7B2026.041.96——————————————————————————
shiftDCModel=Qwen3-VL-8B2026.022.240.310.620.9304.772.732.754.136.610.731.21.473.334.583.960.390.481.36————————
MRDModels=Qwen3-VL2026.052.39——————————————————————————
SafetyModel=LLaVA-v1.6-Mistral-7B2026.042.49——————————————————————————
VanillaModels=Qwen3-VL2026.052.73——————————————————————————
ECSOModels=Qwen3-VL2026.052.85——————————————————————————
AdashieldModel=Qwen3-VL-8B2026.023.360.460.931.3907.164.094.136.199.911.11.82.2156.885.940.590.722.05————————
RAIModel=LLaVA-1.6-7B2026.023.620010.36012.832.457.252011.20.334.1506.830.650.213.97————————
RAIModel=DeepSeek-VL-7B2026.024.322.056.287.10.357.654.520.563.652.811.225.733.942.767.016.691.597.656.26————————
CoCAModel=Qwen3-VL-8B2026.026.730.931.852.78014.328.188.2512.3819.822.213.64.421013.7511.871.171.434.09————————
ECSOModel=Qwen3-VL-8B2026.026.880.951.92.84014.648.368.4412.6620.262.253.684.5210.2214.0612.131.21.464.19————————
AdaShield-ATarget Model=MiniGPT42025.127.2——————————————————————————
OriginalModel=Qwen3-VL-8B2026.027.471.032.063.09015.919.099.1713.7622.022.4544.9111.1115.2813.191.31.594.55————————
MLLM-ProtectorModels=Qwen2.5-VL2026.058.02——————————————————————————
shiftDCModel=DeepSeek-VL-7B2026.028.484.0212.3113.920.68158.871.17.165.52.3911.237.735.4213.7513.123.121512.27————————
HiddenDetectCategory=MLLM Guardrails2025.108.65——————————————————————————
AdaShield-STarget Model=MiniGPT42025.129.82——————————————————————————
Qwen2.5-VL-7BCategory=Offline MLLMs2025.1010——————————————————————————
ECSOModels=Qwen2.5-VL2026.0512.65——————————————————————————
AdashieldModel=DeepSeek-VL-7B2026.0212.726.0318.4720.881.0222.513.31.6510.738.263.5916.8411.68.1320.6219.694.6822.518.41————————
MRDModels=Qwen2.5-VL2026.0512.86——————————————————————————
Claude-3.5-SonnetCategory=Online MLLMs2025.1013.08——————————————————————————
shiftDCModel=LLaVA-1.6-7B2026.0214.838.6621.9621.964.7721.1420.452.4815.4111.565.3418.0415.468.3321.2520.216.4321.8221.62————————
VanillaModels=Qwen2.5-VL2026.0515.13——————————————————————————
GPT-4oCategory=Online MLLMs2025.1016.15——————————————————————————
JailDAMCategory=MLLM Guardrails2025.1016.54——————————————————————————
BaseModel=LLaVA-v1.6-Mistral-7B2026.0417.1——————————————————————————
AM3SafetyBackbone=Qwen2.5-VL-7B2026.0117.26——————————————————56.0171.23——————
Gemma-3-12BInteraction Protocol=Image-tool interaction2026.0521.8——————————————————————————
AdashieldModel=LLaVA-1.6-7B2026.0222.2412.9932.9432.947.1631.730.683.7223.1217.348.0127.0523.1912.531.8730.319.6432.7332.44————————
Qwen3-VL-8BInteraction Protocol=Image-tool interaction2026.0522.3——————————————————————————
Gemma-3-27BInteraction Protocol=Image-tool interaction2026.0523.3——————————————————————————
Qwen3-VL-235BInteraction Protocol=Image-tool interaction2026.0523.8——————————————————————————
CoCAModel=DeepSeek-VL-7B2026.0225.4312.0636.9441.752.044526.63.321.4716.527.1833.6823.1916.2541.2539.389.354536.82————————
ECSOModel=DeepSeek-VL-7B2026.022612.3337.7642.682.094627.193.3821.9416.887.3434.4323.7116.6242.1640.259.564637.64————————
ECSOModels=LLaVA-1.5-7b2026.0526.59——————————————————————————
SPA-VLBackbone=Qwen2.5-VL-7B2026.0127.56——————————————————44.8753.68——————
Qwen3-VL-32BInteraction Protocol=Image-tool interaction2026.0528.2——————————————————————————
OriginalModel=DeepSeek-VL-7B2026.0228.2613.441.0446.392.275029.553.6723.8518.357.9837.4225.7718.0645.8343.7510.395040.91————————
LLaVA-1.5-7B (base)Category=Offline MLLMs2025.1028.85——————————————————————————
SafeSteerModels=LLaVA-1.5-7b2026.0530.37——————————————————————————
Safe RLHF-VBackbone=Qwen2.5-VL-7B2026.0131.01——————————————————53.6861.72——————
RLHF-VBackbone=Qwen2.5-VL-7B2026.0132.2——————————————————4.49.2——————
LlavaGuardCategory=MLLM Guardrails2025.1032.58——————————————————————————
Qwen3-VL-8BInteraction Protocol=Prior turn (NORMAL)2026.0533.2——————————————————————————
Qwen3-VL-8BInteraction Protocol=Direct2026.0534.2——————————————————————————
Llama-4-ScoutInteraction Protocol=Image-tool interaction2026.0534.7——————————————————————————
MLLM-ProtectorModels=LLaVA-1.5-7b2026.0535.15——————————————————————————
Qwen3-VL-235BInteraction Protocol=Prior turn (HARMFUL)2026.0535.6——————————————————————————
ECSOTarget Model=Llava 22025.1236.01——————————————————————————
Qwen3-VL-8BInteraction Protocol=Prior turn (HARMFUL)2026.0536.1——————————————————————————
Qwen3-VL-235BInteraction Protocol=Direct2026.0536.1——————————————————————————
VanillaModels=LLaVA-1.5-7b2026.0537.62——————————————————————————
Qwen3-VL-32BInteraction Protocol=Prior turn (NORMAL)2026.0538.1——————————————————————————
Qwen3-VL-235BInteraction Protocol=Prior turn (NORMAL)2026.0538.1——————————————————————————
SPA-VLBackbone=LLaVA-NeXT-7B2026.0138.21——————————————————37.7164.14——————
MRDModels=LLaVA-1.5-7b2026.0538.31——————————————————————————
Pixtral-LargeInteraction Protocol=Image-tool interaction2026.0539.1——————————————————————————
Qwen3-VL-32BInteraction Protocol=Prior turn (HARMFUL)2026.0539.6——————————————————————————
No DefenseTarget Model=Llava 22025.1239.82——————————————————————————
ECSOTarget Model=MiniGPT42025.1241.52——————————————————————————
Llama-4-ScoutInteraction Protocol=Direct2026.0541.6——————————————————————————
MM-DPOBackbone=Qwen2.5-VL-7B2026.0141.79——————————————————47.6153.24——————
Gemma-3-27BInteraction Protocol=Direct2026.0542.1——————————————————————————
Qwen3-VL-32BInteraction Protocol=Direct2026.0542.6——————————————————————————
Gemma-3-27BInteraction Protocol=Prior turn (NORMAL)2026.0543.1——————————————————————————
AM3SafetyBackbone=LLaVA-NeXT-7B2026.0143.33——————————————————74.0573.2——————
Gemma-3-12BInteraction Protocol=Prior turn (NORMAL)2026.0544.1——————————————————————————
Llama-4-ScoutInteraction Protocol=Prior turn (NORMAL)2026.0544.1——————————————————————————
RLHF-VBackbone=LLaVA-NeXT-7B2026.0144.17——————————————————9.2520.05——————
CoCAModel=LLaVA-1.6-7B2026.0244.4825.9865.8865.8814.3263.4161.367.4346.2434.6816.0154.1146.382563.7560.6219.2965.4664.87————————
Llama-4-ScoutInteraction Protocol=Prior turn (HARMFUL)2026.0544.6——————————————————————————
Gemma-3-12BInteraction Protocol=Direct2026.0545——————————————————————————
ECSOModel=LLaVA-1.6-7B2026.0245.4726.5667.3467.3414.6464.8162.737.647.2735.4516.3755.3147.4125.5665.1661.9719.7266.9166.31————————
No DefenseTarget Model=MiniGPT42025.1245.83——————————————————————————
Pixtral-LargeInteraction Protocol=Prior turn (HARMFUL)2026.0546.5——————————————————————————
Gemma-3-12BInteraction Protocol=Prior turn (HARMFUL)2026.0547.5——————————————————————————
MM-DPOBackbone=LLaVA-NeXT-7B2026.0148.21——————————————————55.0359.72——————
Gemma-3-27BInteraction Protocol=Prior turn (HARMFUL)2026.0548.5——————————————————————————
Qwen2.5-VL-7BAlignment=Original Base Model2026.0148.63——————————————————5050——————
OriginalModel=LLaVA-1.6-7B2026.0249.4228.8773.273.215.9170.4568.188.2651.3838.5317.7960.1251.5327.7870.8367.3621.4372.7372.08————————
Pixtral-LargeInteraction Protocol=Prior turn (NORMAL)2026.0551——————————————————————————
Pixtral-LargeInteraction Protocol=Direct2026.0554——————————————————————————
LLaVA-NeXT-7BAlignment=Original Base Model2026.0161.9——————————————————5050——————
Safe RLHF-VBackbone=LLaVA-NeXT-7B2026.0162.08——————————————————45.7247.46——————
Llama-Guard3-VisionCategory=MLLM Guardrails2025.1074.89——————————————————————————
Circuit BreakersModel=LLaVA-1.5-7B2026.0197.3——————————————————————————
TraceRouterModel=LLaVA-1.5-7B2026.0199.6——————————————————————————
QWEN2.5-VL-7B-INSTRUCTAllowed Domain=Base2026.05—————————————————————100100100100100100
QWEN2.5-VL-7B-INSTRUCT (PALETTE)Allowed Domain=IA2026.05—————————————————————30.292.46272.67581.8