arXiv 2610.23230
RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars
AI19 upvotes
Authors: Haonan Han, Yukang Lin, Minliang Lin, Haoxian Tan, Jun-Yan He, Rui Yang, Zhi-Qi Cheng, Jie Hu +2
Overview
The authors build RAP, a dataset of over 25K verified paper-poster pairs from machine learning conferences (2022-2026), where each human poster is converted into styled HTML/CSS with refillable slots. They also release a 150-paper held-out benchmark and use the corpus for two approaches: PosterRAG, a training-free multi-agent pipeline that retrieves and refills exemplars, and an SFT+RL model.
Introduction
Generating a conference poster requires summarizing a paper, designing its layout, and writing code, but existing benchmarks are small, annotated only with layout boxes, and force evaluation through VLM judges or human checklists. Human-crafted posters also go unused as retrievable design libraries or as verified training data.
Results
The abstract reports no benchmark scores; its key contribution is a 25K+ paper-poster dataset plus a 150-paper benchmark in which every sample can be checked by rendering against its source poster. This makes objective render-and-compare evaluation possible in place of subjective judging.
Conclusion
RAP turns human-made posters into a verifiable, reusable resource for both retrieval-based generation and model training.