The ScienceQA paper from amazon supports this. In a task where images and text were used for prediction tasks (and often the images showed physical things like

The ScienceQA paper from amazon supports this. In a task where images and text were used for prediction tasks (and often the images showed physical things like pulling a door while the text was a question about physics etc), even the items with only text and no images see higher accuracy.