List your company
ResearchFormerly WizWand

Diagram Understanding on AI2D (EM)

79.11Exact MatchBaseline
65.340468.915272.4976.0648Feb 10, 2026
Updated 8mo ago

Evaluation results

MethodLinks
BaselineBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=100%, Compression Ratio=0%2026.0279.11
IDPrunerBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Hybrid strategies2026.0275.94
SCOPEBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Hybrid strategies2026.0275.84
VisionSelectorBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Importance-based2026.0275.19
VisPrunerBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Hybrid strategies2026.0274.29
VisionZipBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Importance-based2026.0274.19
HiPruneBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Importance-based2026.0273.83
DivPruneBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Diversity-based2026.0273.06
FastVBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Importance-based2026.0272.7
IDPrunerBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Hybrid strategies2026.0271.79
DARTBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=25%, Compression Ratio=75%, Method Category=Diversity-based2026.0271.08
VisionSelectorBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Importance-based2026.0270.5
SCOPEBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Hybrid strategies2026.0269.75
HiPruneBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Importance-based2026.0267.75
VisPrunerBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Hybrid strategies2026.0267.75
DivPruneBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Diversity-based2026.0267.71
VisionZipBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Importance-based2026.0267.65
DARTBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Diversity-based2026.0267.49
FastVBase Model=Qwen2.5-VL-3B-Instruct, Token Retention Rate=10%, Compression Ratio=90%, Method Category=Importance-based2026.0265.87