List your company
ResearchFormerly WizWand

Visual Question Answering on V*

74.35AccuracyCropVLM
31.325242.495153.66564.8349May 29, 2025Jul 15, 2025Sep 1, 2025Oct 19, 2025Dec 5, 2025Jan 22, 2026Mar 11, 2026
Updated 6mo ago

Evaluation results

MethodLinks
CropVLMReward=LL, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1174.35
Qwen 2.5 VLInput Resolution=1792x17922025.1173.3
CropVLMReward=Accuracy, Backbone=Qwen 2.5 VL, Input Resolution=1792x1792, CropVLM Resolution=2048x20482025.1172.25
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=1024, Cropping Strategy=CropVLM2025.1160.73
Qwen 2.5 VL + CropVLMBase Model=Qwen 2.5 VL, Cropping Strategy=CropVLM, Input Resolution=2048x20482025.1159.69
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=2048, Cropping Strategy=CropVLM2025.1159.69
LLaVA-1.5-13B + QLIPModel=LLaVA-1.5-13B, Cropped Images=false2025.0558.6
Qwen 2.5 VL + UV-CoTBase Model=Qwen 2.5 VL, Cropping Strategy=UV-CoT, Input Resolution=336x3362025.1156.54
CropVLMBase Model=Qwen 2.5 VL 3B, Resolution=512, Cropping Strategy=CropVLM2025.1156.02
LLaVA-1.5-7b + QLIPModel=LLaVA-1.5-7b, Cropped Images=false2025.0553.4
Qwen 2.5 VL + ViCrop (rel-att)Base Model=Qwen 2.5 VL, Cropping Strategy=ViCrop (rel-att), Input Resolution=448x4482025.1153.4
Qwen 2.5 VL + ViCrop (grad-att)Base Model=Qwen 2.5 VL, Cropping Strategy=ViCrop (grad-att), Input Resolution=448x4482025.1153.4
Qwen 2.5 VLBase Model=Qwen 2.5 VL, Cropping Strategy=None, Input Resolution=448x4482025.1151.31
Qwen 2.5 VL 3BBase Model=Qwen 2.5 VL 3B, Resolution=-, Cropping Strategy=None2025.1151.31
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=Base2026.0350.79
LLaVA 1.5 + CropVLMBase Model=LLaVA 1.5, Cropping Strategy=CropVLM, Input Resolution=2048x20482025.1150.79
CropVLMBase Model=LLaVA 1.5 7B, Resolution=2048, Cropping Strategy=CropVLM2025.1150.79
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=+DIPE2026.0350.26
CropVLMBase Model=LLaVA 1.5 7B, Resolution=1024, Cropping Strategy=CropVLM2025.1149.74
HAVCBackbone=LLaVA-1.5, Signal Type=grad+entropy2026.0149.73
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=Base2026.0349.21
LLaVA 1.5 + ViCrop (rel-att)Base Model=LLaVA 1.5, Cropping Strategy=ViCrop (rel-att), Input Resolution=336x3362025.1149.21
ViCropBackbone=LLaVA-1.5, Signal Type=pure-att2026.0149
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=+DIPE2026.0348.17
LLaVA 1.5 + ViCrop (grad-att)Base Model=LLaVA 1.5, Cropping Strategy=ViCrop (grad-att), Input Resolution=336x3362025.1148.17
ViCropBackbone=LLaVA-1.5, Signal Type=rel-att2026.0147.64
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=+DIPE2026.0347.64
Vanilla RoPEPositional Encoding Method=Vanilla RoPE, DIPE Enhancement=Base2026.0347.12
ViCropBackbone=LLaVA-1.5, Signal Type=grad-att2026.0146.59
CropVLMBase Model=LLaVA 1.5 7B, Resolution=512, Cropping Strategy=CropVLM2025.1146.07
LLaVA-1.5-13BModel=LLaVA-1.5-13B, Cropped Images=false2025.0545
LLaVA 1.5 + UV-CoTBase Model=LLaVA 1.5, Cropping Strategy=UV-CoT, Input Resolution=336x3362025.1143.98
VanillaBackbone=LLaVA-1.52026.0142.93
LLaVA 1.5Base Model=LLaVA 1.5, Cropping Strategy=None, Input Resolution=336x3362025.1142.41
LLaVA 1.5 7BBase Model=LLaVA 1.5 7B, Resolution=-, Cropping Strategy=None2025.1142.41
CropVLMBase Model=GPT 4.1 nano, Resolution=2048, Cropping Strategy=CropVLM2025.1142.41
LLaVA-1.5-7bModel=LLaVA-1.5-7b, Cropped Images=false2025.0542.4
CropVLMBase Model=GPT 4.1 nano, Resolution=512, Cropping Strategy=CropVLM2025.1138.22
CropVLMBase Model=GPT 4.1 nano, Resolution=1024, Cropping Strategy=CropVLM2025.1137.7
VanillaBackbone=InstructBLIP2026.0137.69
ViCropBackbone=InstructBLIP, Signal Type=rel-att2026.0137.17
HAVCBackbone=InstructBLIP, Signal Type=grad+entropy2026.0137.17
ViCropBackbone=InstructBLIP, Signal Type=pure-grad2026.0135.07
ViCropBackbone=InstructBLIP, Signal Type=grad-att2026.0134.55
GPT 4.1 nanoBase Model=GPT 4.1 nano, Resolution=-, Cropping Strategy=None2025.1132.98