CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

Benchmark Model Rank Results
visual-question-answering-vqa-on-mmt-benchCoLT#6Accuracy: 67.4
visual-question-answering-vqa-on-seed-seedbenchCoLT#8Accuracy: 77.5
vision-language-understanding-on-mmstarCoLT#45Accuracy: 68.9
chart-question-answering-on-chartqaCoLT#831:1 Accuracy: 74.7