Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training

Benchmark Model Rank Results
science-question-answering-on-scienceqaMC-CoT F-Large#1Avg. Accuracy: 94.88Natural Science: 97.47Social Science: 90.44
visual-question-answering-on-a-okvqaMC-CoT#3MC Accuracy: 71