InfiniteVL: Synergizing Linear and Sparse Attention for Highly-Efficient, Unlimited-Input Vision-Language Models

Benchmark Model Rank Results
visual-question-answering-vqa-on-textvqa-valInfiniteVL-Base (extra data)#15Accuracy: 78.5
visual-question-answering-vqa-on-mmbench-enInfiniteVL-4B (extra data)#25Accuracy: 79.0
visual-question-answering-on-docvqa-testInfiniteVL-Base (extra data)#33ANLS: 0.917
visual-question-answering-vqa-on-blinkInfiniteVL-Base#45Accuracy: 47.6
optical-character-recognition-ocr-on-ocrbenchInfiniteVL-Base (extra data)#52Score: 798
chart-question-answering-on-chartqaInfiniteVL-4B (extra data)#601:1 Accuracy: 82.0
visual-question-answering-vqa-on-realworldqaInfiniteVL-Base#66Accuracy: 67.3
multimodal-reasoning-on-mathvistaInfiniteVL-Base#81Accuracy: 65.4
vision-language-understanding-on-mmstarInfiniteVL-Base#89Accuracy: 55.6
multimodal-reasoning-on-mmmuInfiniteVL-Base#115Accuracy: 44.0