Yifan Wu, Pengchuan Zhang, Wenhan Xiong, Barlas Oguz, James C. Gee, Yixin Nie
The study explores the effectiveness of the Chain-of-Thought approach, known for its proficiency in language tasks by breaking them down into sub-tasks and intermediate steps, in improving vision-language tasks that demand sophisticated perception and reasoning. We present the "Description then Decision" strategy, which is inspired by how humans process signals. This strategy significantly improves probing task performance by 50%, establishing the groundwork for future research on reasoning paradigms in complex vision-language tasks.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Visual Reasoning | Winoground | Group Score | 58.75 | GPT-4V (CoT, pick b/w two options) |
| Visual Reasoning | Winoground | Image Score | 68.75 | GPT-4V (CoT, pick b/w two options) |
| Visual Reasoning | Winoground | Text Score | 75.25 | GPT-4V (CoT, pick b/w two options) |
| Visual Reasoning | Winoground | Group Score | 39.25 | GPT-4V (pick b/w two options) |
| Visual Reasoning | Winoground | Image Score | 46.25 | GPT-4V (pick b/w two options) |
| Visual Reasoning | Winoground | Text Score | 69.25 | GPT-4V (pick b/w two options) |