Making a conference poster is a miniature exercise in visual coding: a full paper must be compressed onto a single page, demanding summarization, design, and implementation all at once. Yet existing benchmarks for this task are small and annotated only with layout boxes, so evaluation typically falls back on VLM-as-judge or human checklists rather than objective comparison. The new RAP dataset reframes paper-to-poster generation as a visual coding task with verifiable outputs.

The dataset contains more than 25,000 verified paper-poster pairs from machine learning conferences held between 2022 and 2026, each annotated with a three-level content taxonomy. A fully automated derendering pipeline converts every poster into styled HTML/CSS with typed, refillable slots, turning each sample into a paper-poster-markup triplet. Because the markup can be rendered and compared against the source poster, the benchmark supports objective render-and-compare evaluation instead of subjective judging.

RAP also includes a held-out benchmark of 150 test papers paired with their human-crafted posters. The authors demonstrate two uses for the corpus: PosterRAG, a training-free multi-agent pipeline that retrieves an exemplar and refills its slots without touching layout or styling, and an SFT+RL recipe that trains a dedicated model from the same triplets. This positions RAP as both an inference-time design library and a supervised training resource for future work.