DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models

Benchmark Model Rank Results
visual-question-answering-vqa-on-hr-bench-4kDeepScan–Overall: 75.0FSP: 90.1FCP: 59.7
visual-question-answering-vqa-on-hr-bench-4kDyFo–Overall: 71.3FSP: 89.2FCP: 53.5
visual-question-answering-vqa-on-hr-bench-4kInternVL3-38B–Overall: 76.3FSP: 83.5FCP: 69.0
visual-question-answering-vqa-on-hr-bench-4kInternVL3-78B–Overall: 75.5FSP: 84.5FCP: 66.5
visual-question-answering-vqa-on-hr-bench-4kInternVL3-8B–Overall: 70.8FSP: 79.3FCP: 62.3
visual-question-answering-vqa-on-hr-bench-4kLLaVA-OneVision-72B–Overall: 66.3FSP: 76.5FCP: 56.0
visual-question-answering-vqa-on-hr-bench-4kQwen2.5-VL-72B–Overall: 79.4FSP: 88.8FCP: 70.0
visual-question-answering-vqa-on-hr-bench-4kTreeVGR–Overall: 72.7FSP: 89.5FCP: 61.5
visual-question-answering-vqa-on-hr-bench-4kZoomRefine–Overall: 71.5FSP: 88.5FCP: 55.3
visual-question-answering-vqa-on-hr-bench-8kDeepScan–Overall: 72.4FSP: 87.2FCP: 57.6
visual-question-answering-vqa-on-hr-bench-8kDyFo–Overall: 69.8FSP: 86.5FCP: 53.2
visual-question-answering-vqa-on-hr-bench-8kInternVL3-38B–Overall: 67.0FSP: 71.3FCP: 62.8
visual-question-answering-vqa-on-hr-bench-8kInternVL3-78B–Overall: 67.3FSP: 71.8FCP: 62.8
visual-question-answering-vqa-on-hr-bench-8kInternVL3-8B–Overall: 62.0FSP: 64.3FCP: 59.8
visual-question-answering-vqa-on-hr-bench-8kLLaVA-OneVision-72B–Overall: 60.9FSP: 68.8FCP: 53.0
visual-question-answering-vqa-on-hr-bench-8kQwen2.5-VL-72B–Overall: 76.3FSP: 84.3FCP: 68.3
visual-question-answering-vqa-on-hr-bench-8kTreeVGR–Overall: 69.8FSP: 84.4FCP: 57.2
visual-question-answering-vqa-on-hr-bench-8kZoomRefine–Overall: 68.6FSP: 83.9FCP: 54.0