RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars

RAP. A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars. By Haonan Han and nine others, from Tsinghua University, Meituan Longcat, and two other institutions. Published on the sixteenth of October. Its code repository is worked out: about fifty source files, with a readme and a licence.

Making a conference poster is visual coding in miniature. A whole paper has to be compressed onto one page, which exercises summarization, visual design and code implementation at once, using nothing beyond the paper itself. Seen as a visual coding task, though, it has had few benchmarks. Existing resources are small and annotated only with layout boxes. That forces evaluation to fall back on a VLM acting as judge, or on human checklists, instead of the objective approach of rendering an output and comparing it with a reference. The generation methods those resources support also design layouts from scratch. So the many posters people have already crafted by hand go unused, neither retrievable as a design library nor usable as verified supervision. RAP, short for Reconstruct A Poster, is a dataset and benchmark built to fill that space.

The corpus holds more than twenty-five thousand verified pairs of paper and poster. They come from various machine learning conferences, the earliest held in two thousand and twenty-two, the latest in two thousand and twenty-six. Each pair is annotated with a three-level content taxonomy. Alongside the corpus sits a held-out benchmark of one hundred and fifty test papers, and each of those comes paired with the poster its human authors made.

The key construction step is a fully automated derendering pipeline. It takes each finished poster and turns it into styled HTML and CSS, organized as typed, refillable slots. Every sample therefore becomes a triplet: the paper, the poster, and the markup. Because the markup can be rendered, each triplet can be verified by rendering it and comparing the result against the source poster. That check is exactly what box-only annotation could not offer, and it is what makes render-and-compare evaluation possible here.

The corpus is built to serve both inference and training. On the inference side, PosterRAG is a training-free multi-agent pipeline. It retrieves an exemplar poster and refills its slots with the new paper's content, without touching the layout or the styling at all. On the training side, a recipe combining supervised fine-tuning with reinforcement learning trains a dedicated model from the same triplets. Anyone working on document generation as code, or on retrieving and reusing designs, has a use for one half or the other. An online demo is available.

What sets RAP apart from a large pile of posters is the markup layer. Turning human designs into refillable code makes them reusable twice over: once as templates to retrieve and fill, and once as supervision whose correctness can be rendered and checked rather than judged by a model or a checklist.