iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning

Benchmark Model Rank Results
chart-question-answering-on-chartqaiVGR-Qwen3-VL-32B (ours)#61:1 Accuracy: 90.4
visual-question-answering-vqa-on-realworldqaiVGR-Qwen3-VL-32B (ours)#19Accuracy: 76.3
vision-language-understanding-on-mmstariVGR-Qwen3-VL-32B (ours)#23Accuracy: 75.1
multimodal-reasoning-on-mmmuiVGR-Qwen3-VL-32B (ours)#43Accuracy: 67.7