Chain-of-thought has been taking off in LLMs, but there has not been as much of an emphasis on visual reasoning. Many tasks are difficult to solve in one shot, but a human may be able to reduce the complexity by sketching out steps to the solution, either on paper or in the mind's eye. This research aims to recreate this process by adding an extra visual thinking layer at the end of a VLM and benchmarking it against existing VLMs and LLMs.
mashathepotato/visual-reasoning
Folders and files
| Name | Name | Last commit date | ||
|---|---|---|---|---|