Figures are central to scientific communication, and Visual Question Answering (VQA) has become a common way to test whether large multimodal models (LMMs) actually understand them. A new paper on arXiv investigates this in the context of astronomy, a domain where figures carry dense, specialized information.

The authors examine the accuracy of several LMMs on an astronomy VQA dataset, probing how well the models parse visual content that is far removed from everyday images. The abstract reports an investigation into these models' performance, though specific results are not detailed in the available summary.

The work matters because scientific figures require precise reading of axes, labels, and visual conventions. If LMMs struggle on such tasks, it suggests limits to their usefulness in research workflows and science communication.