Digital libraries increasingly host scientific publications whose figures contain primary research findings. Yet making that visual content searchable, accessible, and reusable remains an open challenge, according to a new arXiv paper.
The paper introduces a multi-panel visual question answering dataset built for large language model chart analysis. The authors aim to address the gap between text-based scholarly content and the visual information embedded in figures.
Because scientific charts often combine multiple panels, the dataset focuses on the kind of complex, multi-part visuals common in research papers. The work points toward better tools for navigating and extracting knowledge from the visual record of science.