Document understanding is usually broken into separate tasks: layout analysis, optical character recognition, reading-order prediction, and image captioning. That separation leaves scientific figures, captions, and visual semantics only weakly connected, and there is no single measure of whether a model can produce a page-level parse that is both structurally faithful and semantically accessible. To address this, the authors introduce Read-Parse-Describe, a unified benchmark that asks models to output an ordered list of layout elements with bounding boxes, semantic labels, reading-order indices, recognized text, and alternative text for visual elements.

The dataset is built from 2,402 documents, yielding 18,163 continuous pages, 125,218 layout elements, and 11,338 visual instances. The paper also proposes a plug-and-play space-awareness enhancer that reuses multi-layer visual features and injects grid-aware spatial guidance into vision-language tokens, aiming to improve layout-aware generation.

Experiments across end-to-end vision-language models and pipeline parsers show that existing systems still struggle to jointly recover reading order and visual semantics. The enhanced model achieves strong joint performance, suggesting that the field should move beyond isolated modules toward unified models that read, parse, and describe complex documents. The dataset and code are planned for public release.