List your company
ResearchFormerly WizWand

Visual Question Answering on POPE

89.6AccuracyInternVL3*
81.07283.28685.587.714Dec 28, 2023May 25, 2024Oct 21, 2024Mar 19, 2025Aug 15, 2025Jan 11, 2026Jun 10, 2026
Updated 4mo ago

Evaluation results

MethodLinks
InternVL3*#Parameters=2B2026.0389.6
VanillaSource=Qwen’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=0%2026.0689.6
Qwen3-VL*#Parameters=2B2026.0389.4
ARGUSVLM#Parameters=2B2026.0389.3
ViCrop (Top-3)Model=Qwen2.5-VL-3B-Instruct, Evaluation protocol=zero-shot2025.1288.5
Visual FunnelModel=Qwen2.5-VL-3B-Instruct, Evaluation protocol=zero-shot2025.1288.5
ViCropModel=Qwen2.5-VL-3B-Instruct, Evaluation protocol=zero-shot2025.1288.4
Visual FunnelModel=LLaVA-1.5-7B, Evaluation protocol=zero-shot2025.1288.3
DeepSeekVL#Parameters=2B2026.0388.3
GR3D-8BTraining Stage=Stage 12026.0588.23
Qwen2.5-VL*#Parameters=3B2026.0388.1
NVILA-Lite-8B2026.0588.1
JanusFlowLLM Param=1.3B, Type=Unified2024.1188
VanillaBackbone=QWEN-2.5-VL 7B, Average Tokens=100%2026.0287.9
DiDi-Merging2026.0587.83
VanillaSource=Qwen’25, Token Reduction=0%, Backbone=Qwen2.5-VL-7B2026.0687.8
UIO-2XXLModel Size=6.8B2023.1287.7
SmolVLM*#Parameters=2B2026.0387.7
NuwaBackbone=QWEN-2.5-VL 7B, Average Tokens=75%2026.0287.52
SmolVLMParameters=2B2025.1287.5
ViCrop (Top-3)Model=LLaVA-1.5-7B, Evaluation protocol=zero-shot2025.1287.5
NuwaBackbone=QWEN-2.5-VL 7B, Average Tokens=50%2026.0287.46
ViCropModel=LLaVA-1.5-7B, Evaluation protocol=zero-shot2025.1287.4
LLaVA-SGLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0387.3
ARGUSVLM#Parameters=256M2026.0387.3
UIO-2XLModel Size=3.2B2023.1287.2
PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0687.2
LLaVA-VTLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0387.1
Visual FunnelModel=InstructBLIP-7B, Evaluation protocol=zero-shot2025.1287.1
Qwen2.5-VL-3B-InstructVisual input strategy=No Cropping, Evaluation protocol=zero-shot2025.1287.1
NuwaBackbone=QWEN-2.5-VL 7B, Average Tokens=25%2026.0287.06
JanusLLM Param=1.3B, Type=Unified2024.1187
ViCrop (Top-3)Model=InstructBLIP-7B, Evaluation protocol=zero-shot2025.1287
GR3D-8BTraining Stage=Stage 22026.0587
LLaVA-DCapLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0386.9
FREE-Merging2026.0586.73
ViCropModel=InstructBLIP-7B, Evaluation protocol=zero-shot2025.1286.6
MMER2026.0586.58
LLaVA-CapLLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=E + T2024.0386.4
LLaVA-1.5-13B-HD2025.1286.3
ARGUSVLM#Parameters=500M2026.0386.3
mPLUG-Owl2Model Size=8.2B2023.1286.2
Individuals2026.0586.17
LLaVa-1.5Model Size=7.2B2023.1285.9
LLaVa-1.5Model Size=13B2023.1285.9
LLaVA-1.5 (V-7B)LLM=Vicuna-7B, #PT=558K, #IT=665K, Representation=Visual Embeddings (E)2024.0385.9
LLaVA-1.5 (V-13B)LLM=Vicuna-13B, #PT=558K, #IT=665K, Representation=Visual Embeddings (E)2024.0385.9
LLaVA-v1.5LLM Param=7B, Type=Und. Only2024.1185.9
LLaVA-1.5Parameters=7B2025.1285.9
SmolVLM*#Parameters=500M2026.0385.9
EUPE-ViT-BBackbone=ViT-B2026.0385.9
FerretModel Size=7.2B2023.1285.8
VILA-ULLM Param=7B, Type=Unified2024.1185.8
HAVCBackbone=LLaVA-1.5, Signal Type=grad+entropy2026.0185.8
SmolVLMParameters=500M2025.1285.8
PEcore-BBackbone=ViT-B2026.0385.8
PDrop+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0685.8
VL-JEPASFTParameters=1.6B2025.1285.7
DINOv3-ViT-BBackbone=ViT-B2026.0385.7
LLaVA-1.5-7BVisual input strategy=No Cropping, Evaluation protocol=zero-shot2025.1285.6
MRoPEPositional Encoding Method=MRoPE, DIPE Enhancement=+DIPE2026.0385.58
PDropSource=CVPR’25, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0685.5
BLIP-2LLM=Vicuna-13B, #PT=129M, Representation=Visual Embeddings (E)2024.0385.3
Emu3-ChatLLM Param=8B, Type=Und. Only2024.1185.2
FastV+RerouteSource=Ours, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0685.2
LLAVA-PhiLLM Param=2.7B, Type=Und. Only2024.1185
SigLIP2-BBackbone=ViT-B2026.0385
MobileVLMLLM Param=2.7B, Type=Und. Only2024.1184.9
ViCropBackbone=LLaVA-1.5, Signal Type=grad-att2026.0184.9
ShikraModel Size=7.2B2023.1284.7
MobileVLM-V2LLM Param=2.7B, Type=Und. Only2024.1184.7
ViCropBackbone=LLaVA-1.5, Signal Type=rel-att2026.0184.7
InstructBLIP-7BVisual input strategy=No Cropping, Evaluation protocol=zero-shot2025.1284.7
RADIOv2.5-BBackbone=ViT-B2026.0384.7
PDropSource=CVPR’25, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0684.7
FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=66.7%2026.0684.7
FastVSource=ECCV’24, Token Reduction=66.7%, Backbone=Qwen2.5-VL-7B2026.0684.6
MobileVLMLLM Param=1.4B, Type=Und. Only2024.1184.5
MobileVLM#Parameters=1.7B2026.0384.5
DUNE-BBackbone=ViT-B2026.0384.5
PEspatial-BBackbone=ViT-B2026.0384.4
MobileVLM-V2LLM Param=1.4B, Type=Und. Only2024.1184.3
MobileVLM-V2#Parameters=1.7B2026.0384.3
LLaVA-v1.5-Phi-1.5LLM Param=1.3B, Type=Und. Only2024.1184.1
ViCropBackbone=LLaVA-1.5, Signal Type=pure-att2026.0184
VanillaBackbone=LLaVA-1.52026.0183.9
MiniGemini#Parameters=2.3B2026.0383.9
PDrop+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.0683.9
MRoPE-IPositional Encoding Method=MRoPE-I, DIPE Enhancement=+DIPE2026.0383.72
Vicuna-DCapLLM=Vicuna-13B, #IT=665K, Representation=Text (T)2024.0383.4
Video-LLaVAParameters=7B2025.1283.4
PDrop+RerouteSource=Ours, Token Reduction=77.8%, Backbone=Qwen2.5-VL-7B2026.0683.1
ViCropBackbone=InstructBLIP, Signal Type=pure-grad2026.0182.9
HAVCBackbone=InstructBLIP, Signal Type=grad+entropy2026.0182.7
FastV+RerouteSource=Ours, Backbone=Qwen3.5-9B-Hybrid, Average Gated Attention Token Reduction=77.8%2026.0682.5
ViCropBackbone=InstructBLIP, Signal Type=rel-att2026.0182.4
Vicuna-SGLLM=Vicuna-13B, #IT=665K, Representation=Text (T)2024.0382.3
ViCropBackbone=InstructBLIP, Signal Type=grad-att2026.0182
Vicuna-VTLLM=Vicuna-13B, #IT=665K, Representation=Text (T)2024.0381.9
VanillaBackbone=InstructBLIP2026.0181.4