List your company
ResearchFormerly WizWand

Multimodal Question Answering on MMBench CN

84.6AccuracyQwen3-VL-8B
47.36857.03466.776.366Jun 16, 2024Oct 18, 2024Feb 19, 2025Jun 24, 2025Oct 26, 2025Feb 27, 2026Jul 2, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Qwen3-VL-8BRetained Tokens=1024 Tokens (100%)2026.0784.6
MergeMixbaseline=SFT Vision2025.1081.18
VisionThink-7B2025.1081.01
SFT Vision2025.1081.01
Qwen2.5-VL-Ins-7B2025.1080.41
Qwen2.5-VL-7BToken Budget=100%2025.0880.1
CDPrunerRetained Tokens=256 Tokens (↓ 75.0%)2026.0780
EADPRetained Tokens=512 Tokens (↓ 50.0%)2026.0779.7
HiPruneBase Model=Qwen2.5-VL-7B, Token Budget=33.3%2025.0879.5
DivPruneRetained Tokens=512 Tokens (↓ 50.0%)2026.0779.4
CDPrunerRetained Tokens=512 Tokens (↓ 50.0%)2026.0779.3
HiPrune++Base Model=Qwen2.5-VL-7B, Token Budget=33.3%2025.0879.2
EADPRetained Tokens=256 Tokens (↓ 75.0%)2026.0778.9
DivPruneRetained Tokens=256 Tokens (↓ 75.0%)2026.0778.8
EADPRetained Tokens=128 Tokens (↓ 87.5%)2026.0778.4
HiPruneBase Model=Qwen2.5-VL-7B, Token Budget=22.2%2025.0877.6
CDPrunerRetained Tokens=128 Tokens (↓ 87.5%)2026.0777.5
HiPrune++Base Model=Qwen2.5-VL-7B, Token Budget=22.2%2025.0877
DivPruneRetained Tokens=128 Tokens (↓ 87.5%)2026.0776.6
HiPruneBase Model=Qwen2.5-VL-7B, Token Budget=11.1%2025.0873.5
HiPrune++Base Model=Qwen2.5-VL-7B, Token Budget=11.1%2025.0873.3
Qwen2.5-VL-3BToken Budget=100%2025.0873
QMoSLoRAModel=InternLM2 + ViT, Initialization=Kaiming uniform, Quantization=4-bit2024.0672.6
HiPruneBase Model=Qwen2.5-VL-3B, Token Budget=33.3%2025.0871.8
QMoSLoRAModel=InternLM2 + ViT, Initialization=Orthogonal, Quantization=4-bit2024.0671.2
HiPrune++Base Model=Qwen2.5-VL-3B, Token Budget=33.3%2025.0870.9
HiPruneBase Model=Qwen2.5-VL-3B, Token Budget=22.2%2025.0869.1
QLoRAModel=InternLM2 + ViT, Quantization=4-bit2024.0668.9
HiPrune++Base Model=Qwen2.5-VL-3B, Token Budget=22.2%2025.0868.9
MoSLoRAModel=LLaMA-3 + ViT, Initialization=Orthogonal2024.0668.2
LoRAModel=LLaMA-3 + ViT2024.0667.8
MoSLoRAModel=LLaMA-3 + ViT, Initialization=Kaiming uniform2024.0667.5
HiPruneBase Model=Qwen2.5-VL-3B, Token Budget=11.1%2025.0865.1
HiPrune++Base Model=Qwen2.5-VL-3B, Token Budget=11.1%2025.0864.9
CoVFTBackbone=LLaVA-1.5-13B2026.0364.52
LLaVA-1.5Backbone=LLaVA-1.5-13B2026.0363.6
FreezeBackbone=LLaVA-1.5-13B2026.0363.57
Full fine-tuningBackbone=LLaVA-1.5-13B2026.0363.49
BitFitBackbone=LLaVA-1.5-13B2026.0362.54
CoVFTBackbone=LLaVA-1.5-7B2026.0360.4
FreezeBackbone=LLaVA-1.5-7B2026.0360.3
Full fine-tuningBackbone=LLaVA-1.5-7B2026.0359.88
BitFitBackbone=LLaVA-1.5-7B2026.0359.02
LLaVA-1.5Backbone=LLaVA-1.5-7B2026.0358.3
SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0658.2
VPTBackbone=LLaVA-1.5-7B2026.0358.16
LLaVA-1.5-7BBackbone=LLaVA-1.5-7B, Retained Tokens=5762026.0658.1
SVPTBackbone=LLaVA-1.5-7B2026.0357.9
LoRABackbone=LLaVA-1.5-7B2026.0357.56
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0657.3
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0656.4
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=1282026.0655.2
PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=1282026.0654.6
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0654
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0653.9
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0653.4
SparseVLMBackbone=LLaVA-1.5-7B, Retained Tokens=642026.0652.6
PruMerge+Backbone=LLaVA-1.5-7B, Retained Tokens=642026.0652.3
SPAREBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0650.3
CDPrunerBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0649.6
DARTBackbone=LLaVA-1.5-7B, Retained Tokens=322026.0648.8