List your company
ResearchFormerly WizWand

Visual Question Answering on MMBench

92.1AccuracyDDP
63.70871.07978.4585.821Mar 27, 2024Aug 11, 2024Dec 27, 2024May 14, 2025Sep 28, 2025Feb 13, 2026Jul 1, 2026
Updated 3mo ago

Evaluation results

MethodLinks
DDPbackbone=Gemini-3-Pro2026.0492.1
Gemini 2.5 Pro2026.0490.1
Gemini-3-pro2026.0488.4
OpenVLThinkerV2Parameters=8B2026.0488.2
Vision-G12026.0488
VL-Rethinker-7BParameters=7B2026.0487.9
OpenVLThinker-7BParameters=7B2026.0487.8
MM-Eureka-7BParameters=7B2026.0487.7
Qwen3-VL-2B-Instruct + MoDA (ours)LLM=Qwen3-2B2025.0687.4
OVR-7BParameters=7B2026.0486.6
OneThinker-8BParameters=8B2026.0486.6
Qwen3-VL-2B-InstructLLM=Qwen3-2B2025.0686.5
Qwen3-VL GDPORL Strategy=GDPO2026.0486.1
Qwen3-VL GRPORL Strategy=GRPO2026.0486
Qwen3-VL-Instruct-8BParameters=8B2026.0485.1
GPT-4o2026.0484.3
VanillaSource=Qwen’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=0%2026.0684
MiMo-VL-7B +PCBackbone Model=MiMo-VL-7B, Training Paradigm (+PC)=Yes2026.0383.8
VanillaSource=Qwen’25, Token Reduction=0%, Backbone=Qwen2.5-VL-7B2026.0683.7
GPT-4o2026.0483.4
MiMo-VL-7BBackbone Model=MiMo-VL-7B, Training Paradigm (+PC)=No2026.0383.2
Qwen2.5-VL-7B +PCBackbone Model=Qwen2.5-VL-7B, Training Paradigm (+PC)=Yes2026.0382.5
InternVL-1.5Model Size=1.5-26B, Access Type=Open-source2024.0782.3
Gemini-1.5-pro2026.0482.3
IXC-2.5-7BModel Size=7B2024.0782.2
Qwen2.5-VL-7BBackbone Model=Qwen2.5-VL-7B, Training Paradigm (+PC)=No2026.0382.2
GPT-4VAccess Type=Closed-source API2024.0781.3
PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0680
IDPrunerRetained Tokens=25%2026.0279.73
SCOPERetained Tokens=25%2026.0279.38
Qwen2.5-VL-3BBackbone Model=Qwen2.5-VL-3B, Training Paradigm (+PC)=No2026.0379.1
Qwen2.5-VL-3B +PCBackbone Model=Qwen2.5-VL-3B, Training Paradigm (+PC)=Yes2026.0379.1
PDrop+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0679
PDropSource=CVPR’25, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0678.6
SFT + VIGILBase Model=Qwen2.5-VL-72B2026.0678.5
FastVSource=ECCV’24, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0677.8
SFT + DA-DPOBase Model=Qwen2.5-VL-72B2026.0677.8
SFT + SimPOBase Model=Qwen2.5-VL-72B2026.0677.5
FastV+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0677.3
SFT + DPOBase Model=Qwen2.5-VL-72B2026.0677.2
FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0676.9
SFT onlyBase Model=Qwen2.5-VL-72B2026.0676.8
SFT + HA-DPOBase Model=Qwen2.5-VL-72B2026.0676.5
PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.0675.6
PDrop+RerouteSource=Ours, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.0675.4
PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0674.2
Gemini-ProAccess Type=Closed-source API2024.0773.9
TinyLLaVA†LLM=Qwen2.5-3B2026.0773.1
Imp-3BLLM=Phi2-2.7B2026.0772.9
PDropSource=CVPR’25, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.0672.8
FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.0672.6
SFT + VIGILBase Model=Qwen2.5-VL-7B2026.0672.5
LLaVA-MoRE SigLIP-S2 + MoDA (ours)LLM=LLaMA 3.1-8B2025.0672.4
LLaVA-MoRE OpenAI CLIPLLM=LLaMA 3.1-8B2025.0672.3
LLaVA-MoRE OpenAI CLIP + MoDA (ours)LLM=LLaMA 3.1-8B2025.0672
SFT + DA-DPOBase Model=Qwen2.5-VL-7B2026.0672
LLaVA-MoRE SigLIP-S2LLM=LLaMA 3.1-8B2025.0671.8
SFT + SimPOBase Model=Qwen2.5-VL-7B2026.0671.5
SFT + DPOBase Model=Qwen2.5-VL-7B2026.0671.2
LLaVA-KDLLM=Qwen2.5-1.5B2026.0771
PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=88.9%2026.0670.8
TRAGLLM=Qwen2.5-1.5B2026.0770.8
TinyLLaVA†LLM=Qwen2.5-1.5B2026.0770.6
SFT onlyBase Model=Qwen2.5-VL-7B2026.0670.5
FastVSource=ECCV’24, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0669.9
LLaVA-CapLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0369.4
LLaVA-VTLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0369.4
LLaVA-SGLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0369.2
LLaVA-DCapLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0368.9
LLaVA-MoDLLM=Qwen2-1.5B2026.0768.9
ShareGPT-4VLLM=Vicuna-7B2025.0668.8
BunnyLLM=Phi2-2.7B2026.0768.6
FastV+RerouteSource=Ours, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.0668.1
MoE-LLaVALLM=Phi2-2.7B2026.0768
LLaVA-1.5-13B + QLIPModel=LLaVA-1.5-13B, Cropped Images=true2025.0567.9
LLaVA-1.5 (V-13B)LLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=Visual Embeddings (E)2024.0367.7
VanillaRetained Tokens=576, Compression Ratio=100%2026.0467.7
LLaVA-1.5LLM=Vicuna-13B2025.0667.7
MMFuserBackbone=LLaVA-1.5-7B2026.0267.5
TinyLLaVA†LLM=Qwen1.5-4B2026.0767.5
LLaVA-1.5-13BModel=LLaVA-1.5-13B, Cropped Images=false2025.0567.4
PDropSource=CVPR’25, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.0667.4
TinyLLaVALLM=Phi2-2.7B2026.0766.9
Dense ConnectorBackbone=LLaVA-1.5-7B2026.0266.8
EvoComp (l = 0)Retained Tokens=64, Compression Ratio=88.9%, Compressor layers (l)=0, Transfer setting=from LLaVA-1.5-7B to LLaVA-1.5-13B2026.0466.8
TRAGLLM=Qwen1.5-1.8B2026.0766.6
FastVSource=ECCV’24, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.0666.5
TGIFBackbone=LLaVA-1.5-7B2026.0266.4
LLaVA-1.5-7BToken Retention Rate=576, Base Model=LLaVA-1.5-7B2025.0666.31
PDrop+RerouteSource=Ours, Token Reduction=88.9%, Backbone=Qwen2.5-VL-7B2026.0666
KD†LLM=Qwen1.5-1.8B2026.0765.8
PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.0665.7
DARTRetained Tokens=64, Compression Ratio=88.9%2026.0465.5
EvoComp (l = 2)Retained Tokens=64, Compression Ratio=88.9%, Compressor layers (l)=2, Transfer setting=from LLaVA-1.5-7B to LLaVA-1.5-13B2026.0465.5
SparseVLM-v1.5Token Retention Rate=192, Base Model=LLaVA-1.5-7B2025.0665.41
SparseVLM-v1.5Token Retention Rate=128, Base Model=LLaVA-1.5-7B2025.0665.25
VisionDropToken Retention Rate=192, Base Model=LLaVA-1.5-7B2025.0665.19
VisPrunerToken Retention Rate=192, Base Model=LLaVA-1.5-7B2025.0664.91
VisionZipRetained Tokens=64, Compression Ratio=88.9%2026.0464.9
LLaVA-1.5 + MoDA (ours)LLM=Vicuna-7B2025.0664.8