RAP dataset pairs 25K+ papers with posters for visual coding
Original title: RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars
Each sample is a paper-poster-HTML triplet; a 150-paper held-out set supports render-and-compare evaluation.
Researchers introduce RAP (Reconstruct A Poster), a dataset and benchmark for turning a paper into a conference poster as visual coding. It holds more than 25,000 verified paper-poster pairs from machine learning conferences held between 2022 and 2026, plus a held-out set of 150 test papers each paired with a human poster. An automated derendering pipeline converts posters into styled HTML/CSS with typed, refillable slots. PosterRAG retrieves an exemplar and refills slots without changing layout or styling; an SFT+RL recipe trains a model on the same triplets.
- Pairs come from machine learning conferences held between 2022 and 2026.
- Annotations use a three-level content taxonomy.
- The held-out benchmark has 150 test papers with human posters.
- An online demo is at reconstruct-a-poster.github.io.
From the original article
1 min read at huggingface.co
Making a conference poster is visual coding in miniature: an entire paper must be compressed onto a single page, which exercises summarization, visual design, and code implementation at once, using nothing beyond the paper itself.
By Haonan Han, Yukang Lin, Minliang Lin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.