DRAGON Benchmark Tests Visual Grounding in Diagram QA

Researchers have launched DRAGON, a new benchmark designed to assess evidence-based visual reasoning within diagrams. Diagram question answering (DQA) involves models interpreting organized visual formats like charts, maps, infographics, circuit schematics, and scientific diagrams. Although recent vision-language models (VLMs) often demonstrate high accuracy in answering these questions, achieving the correct answer does not ensure that the models utilize the diagram regions that substantiate their predictions. Instead, they may depend on textual links or dataset artifacts, bypassing the visual evidence necessary for answer verification. This shortcoming hinders effective evaluation of diagram reasoning and diminishes interpretability. DRAGON mitigates this by requiring models to identify bounding boxes for visual elements that support the answer, including components that contain answers and relevant text. The benchmark strives for a more thorough evaluation of AI systems' diagram comprehension.