RAP Collects 25 Thousand Paper–Poster Pairs to Teach AI to Build Posters in HTML/CSS

Researchers have introduced RAP (Reconstruct A Poster), a dataset and benchmark that turns creating a scientific poster into a visual coding task. The corpus holds more than 25 thousand verified paper–poster pairs from ML conferences held in 2022–2026. It also includes a held-out test set of 150 papers, each with its original author-made poster. Every poster is automatically converted into HTML/CSS, so each sample is a paper–poster–markup triple.

Why posters

The authors describe a poster as visual coding in miniature. An entire paper has to be squeezed onto a single page. To do that, a model must summarize text, make design decisions and write code at the same time, and the paper itself is its only source material.

Benchmarks for this task are scarce. Existing sets are small, and their annotations stop at layout block boundaries. As a result, quality has to be judged either by a VLM acting as a judge or by people working from checklists. There is no objective "render it and compare it to the reference" comparison. There is a second problem too: methods trained on such data design layouts from scratch, and ready-made author posters are used neither as a design library for retrieval nor as a verified training set.

How RAP works

At the center of the pipeline is fully automatic derendering. Each poster is converted into styled HTML/CSS with typed slots that can be refilled. Because the markup comes from a real poster, it can be rendered and compared with the original. This makes every sample verifiable.

All pairs are also annotated with a three-level content taxonomy. The description does not say how the levels are structured.

• Size: 25 thousand+ paper–poster pairs from 2022–2026 conferences.
• Test set: 150 held-out papers, each with a human-made poster.
• Format: a paper–poster–markup triple, verifiable by rendering.
• Markup: HTML/CSS with typed slots for refilling.
• Taxonomy: three levels of content description.

Two ways to use the corpus

Using the same data, the authors demonstrate two approaches. The first is PosterRAG, a multi-agent pipeline that needs no fine-tuning. It finds a suitable author-made poster in the corpus and refills its slots with the new paper's content. The layout and styles stay unchanged. The design comes from humans, and the model handles the content.

The second approach is an SFT+RL recipe, which trains a standalone model on the same triples. Supervised fine-tuning gives the model a basic ability to write poster markup, and reinforcement learning improves the result. Here, render-based verification could serve as an objective signal. That is an assumption, not a fact stated in the abstract. There is also an online demo at reconstruct-a-poster.github.io.

The abstract gives no quality numbers: no comparison of PosterRAG with from-scratch generation, and no results for the trained model. For now, the value of the work lies in the data and in how it frames the task. How strong the proposed methods are will become clear from the full paper.

Where this is useful

For researchers working on document and interface generation, RAP offers something rare: a reference against which output can be compared automatically. There is no need to assemble an expert panel or to trust another model's judgment. The "find a good example and fill the slots" approach carries over to neighboring formats: slides, landing pages, reports and infographics.

For paper authors, the practical benefit is more modest for now: the demo lets them see what an automatically built poster looks like. But with a ready corpus and an open training recipe, conference tools will likely appear quickly. Those building agent pipelines should take a look at the PosterRAG design. It separates design from content, which reduces the risk of broken layouts that plagues from-scratch layout generation.