List your company
ResearchFormerly WizWand

Science Question Answering on ScienceQA SQA-IMG (test)

91.5SQA-IMG AccuracyQwen
46.717658.343869.9781.5962Aug 1, 2025Sep 2, 2025Oct 5, 2025Nov 6, 2025Dec 9, 2025Jan 10, 2026Feb 12, 2026
Updated 6mo ago

Evaluation results

MethodLinks
QwenToken Retention Rate=100%, Base Model=Qwen2.5-VL-32B-Instruct2025.0891.5—
HiPruneToken Retention Rate=22.2%, Base Model=Qwen2.5-VL-32B-Instruct2025.0887—
VisionZipToken Retention Rate=22.2%, Base Model=Qwen2.5-VL-32B-Instruct2025.0886.5—
HiPrune++Token Retention Rate=22.2%, Base Model=Qwen2.5-VL-32B-Instruct2025.0886.5—
FastVToken Retention Rate=22.2%, Base Model=Qwen2.5-VL-32B-Instruct2025.0884.4—
VisionZipToken Retention Rate=11.1%, Base Model=Qwen2.5-VL-32B-Instruct2025.0883.5—
FastVToken Retention Rate=11.1%, Base Model=Qwen2.5-VL-32B-Instruct2025.0882.9—
HiPrune++Token Retention Rate=11.1%, Base Model=Qwen2.5-VL-32B-Instruct2025.0882.9—
HiPruneToken Retention Rate=11.1%, Base Model=Qwen2.5-VL-32B-Instruct2025.0882.6—
HiPruneToken Retention Rate=5.6%, Base Model=Qwen2.5-VL-32B-Instruct2025.0882.6—
HiPrune++Token Retention Rate=5.6%, Base Model=Qwen2.5-VL-32B-Instruct2025.0881.8—
VisionZipToken Retention Rate=5.6%, Base Model=Qwen2.5-VL-32B-Instruct2025.0881.5—
FastVToken Retention Rate=5.6%, Base Model=Qwen2.5-VL-32B-Instruct2025.0878.7—
ScalSelectSelection Budget=400K2026.0267.72—
Full-FinetuneSelection Budget=625K (Full)2026.0267.63—
COINCIDESelection Budget=100K2026.0266.83—
ScalSelectSelection Budget=300K2026.0266.53—
ScalSelectSelection Budget=200K2026.0265.39—
ScalSelectSelection Budget=100K2026.0265.29—
RandomSelection Budget=100K2026.0265.2—
RDS+Selection Budget=100K2026.0264.45—
PRISMSelection Budget=100K2026.0264.06—
ScalSelectSelection Budget=50K2026.0263.11—
LengthSelection Budget=100K2026.0252.7—
PerplexitySelection Budget=100K2026.0248.44—
BLIP-2Vision Encoder=ViT-g, Resolution=224, LLM=Vicuna-13B, Data=129M2025.01—61
InstructBLIPVision Encoder=ViT-g, Resolution=224, LLM=Vicuna-7B, Data=130M2025.01—60.5
InstructBLIPVision Encoder=ViT-g, Resolution=224, LLM=Vicuna-13B, Data=130M2025.01—63.1
LLaVA-1.5Vision Encoder=CLIP-L, Resolution=336, LLM=Vicuna-7B, Data=1.2M2025.01—66.8
LLaVA-1.5Vision Encoder=CLIP-L, Resolution=336, LLM=Vicuna-13B, Data=1.2M2025.01—71.6
LLaVA-HRVision Encoder=ConvNeXt-L, CLIP-L, Resolution=1024/448, LLM=Vicuna-7B, Data=1.2M2025.01—65.1
LLAVA-HRVision Encoder=ConvNeXt-L, CLIP-L, Resolution=1024/448, LLM=Vicuna-13B, Data=1.2M2025.01—68.1
LLaVA-NeXTVision Encoder=CLIP-L, Resolution=1344, LLM=Vicuna-7B, Data=1.6M2025.01—70.1
LLaVA-UHDVision Encoder=CLIP-L, Resolution=1008, LLM=Vicuna-13B, Data=1.2M2025.01—67.7
LLaVA-UHD v2Vision Encoder=CLIP-L, Resolution=1008, LLM=Vicuna-7B, Data=1.4M2025.01—67.6
MG-LLaVAVision Encoder=ConvNeXt-L, CLIP-L, RAMPlus, OWL-ViTv2-L, Resolution=768/336, LLM=Vicuna-7B, Data=2.5M2025.01—70.8
MM1Vision Encoder=CLIP-L, Resolution=1792, LLM=7B, Data=1B2025.01—72.6
mPLUG-Owl2Vision Encoder=CLIP-L, Resolution=448, LLM=Llama-2-7B, Data=400M2025.01—68.7
PIIP-LLaVAVision Encoder=ConvNeXt-B, CLIP-L, Resolution=1024/336, LLM=Vicuna-7B, Data=1.2M2025.01—68.1
PIIP-LLaVAVision Encoder=ConvNeXt-L, CLIP-L, Resolution=1024/336, LLM=Vicuna-7B, Data=1.2M2025.01—68.3
PIIP-LLaVAVision Encoder=ConvNeXt-B, CLIP-L, Resolution=1024/336, LLM=Vicuna-13B, Data=1.2M2025.01—71.1
PIIP-LLaVAVision Encoder=ConvNeXt-L, CLIP-L, Resolution=1024/336, LLM=Vicuna-13B, Data=1.2M2025.01—69.8
PIIP-LLaVAVision Encoder=ConvNeXt-L, CLIP-L, Resolution=1024/336, LLM=Vicuna-7B, Data=2.7M2025.01—95
QwenVL-ChatVision Encoder=ViT-G, Resolution=448, LLM=Qwen-7B, Data=1.4B2025.01—68.2
SlimeVision Encoder=CLIP-L, Resolution=2016, LLM=Vicuna-7B, Data=2M2025.01—76.8
SPHINX-intern2Vision Encoder=ConvNeXt-XXL, DINOv2-g, Resolution=448, LLM=InternLM2-7B, Data=16M2025.01—70.4