The Role of Chain-of-Thought in Complex Vision-Language Reasoning Task

Benchmark Model Rank Results
visual-reasoning-on-winogroundGPT-4V (CoT, pick b/w two options)–Text Score: 75.25Image Score: 68.75Group Score: 58.75
visual-reasoning-on-winogroundGPT-4V (pick b/w two options)–Text Score: 69.25Image Score: 46.25Group Score: 39.25