VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions

Benchmark Model Rank Results
chart-question-answering-on-chartqaVISOR–1:1 Accuracy: 65.2
object-hallucination-evaluation-on-popeVISOR–F1: 87.6
visual-question-answering-on-docvqa-testVISOR–ANLS: 0.689
visual-question-answering-vqa-on-realworldqaVISOR–Accuracy: 54.6