List your company
ResearchFormerly WizWand

Visual Question Answering on GQA (Accuracy and Average Performance %)

62.4AccuracyVanilla
35.3642.3849.456.42May 28, 2026Jun 2, 2026Jun 8, 2026Jun 14, 2026Jun 19, 2026Jun 25, 2026Jul 1, 2026
Updated 4mo ago

Evaluation results

MethodLinks
VanillaAverage Token Budget=576, Compression Level=100%2026.0562.4100
VisionZip ‡Average Token Budget=192, Compression Level=↓ 66.7%2026.0561.698.4
MutualAverage Token Budget=192, Compression Level=↓ 66.7%2026.0561.499.4
CLSAverage Token Budget=192, Compression Level=↓ 66.7%2026.0561.499.3
CLSAverage Token Budget=128, Compression Level=↓ 77.8%2026.0560.898.6
MutualAverage Token Budget=128, Compression Level=↓ 77.8%2026.0560.798.4
VisionZipAverage Token Budget=192, Compression Level=↓ 66.7%2026.0560.398.2
VisionZip ‡Average Token Budget=128, Compression Level=↓ 77.8%2026.056097
MutualAverage Token Budget=64, Compression Level=↓ 88.9%2026.0559.896.8
CLSAverage Token Budget=64, Compression Level=↓ 88.9%2026.0559.696.6
VisionZipAverage Token Budget=128, Compression Level=↓ 77.8%2026.0558.796.1
VisionZip ‡Average Token Budget=64, Compression Level=↓ 88.9%2026.0557.795.1
VisionZipAverage Token Budget=64, Compression Level=↓ 88.9%2026.0555.891.8
LeVLJEPALanguage Model=Llama-1B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0744.6—
LeVLJEPALanguage Model=Qwen-1.5B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0743.7—
InfoNCELanguage Model=Llama-1B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0742.7—
SigLIPLanguage Model=Llama-1B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0742.4—
InfoNCELanguage Model=Qwen-1.5B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0742.2—
SigLIPLanguage Model=Qwen-1.5B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0741.6—
RandomLanguage Model=Qwen-1.5B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0737—
RandomLanguage Model=Llama-1B, Vision Encoder=ViT-B/16, Pre-training Data=Datacomp-L, Bridge Architecture=MLP, Evaluation Protocol=Frozen backbone and LLM2026.0736.4—