VincentHancoder / RAP

# RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars
**Haonan Han** 1,2  
**Yukang Lin** 1  
**Minliang Lin** 2  
**Haoxian Tan** 2  
**Jun-Yan He** 2 ∗  
**Rui Yang** 3 
**Zhi-Qi Cheng** 4  
**Jie Hu** 2  
**Guanglu Wan** 2  
**Xiu Li** 1 † 
 1 Tsinghua University    2 Meituan Longcat    3 The University of Hong Kong    4 University of Washington 
 ∗ Project lead    † Corresponding author 
---
## Overview
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. **RAP** turns this into a measurable task.
- **RAP-25K** — 25,435 paper–poster pairs from NeurIPS, ICLR and ICML (2022–2026), each with a three-level content
 taxonomy (15 domains / 20 tasks / 39 methods) and a distilled introduction summary.
- **Poster derendering** — a fully automated pipeline rewrites every human-crafted poster as styled HTML/CSS with typed,
 refillable slots, so each sample is a *paper–poster–markup* triplet that can be verified by rendering.
- **PosterRAG** — a training-free multi-agent pipeline that retrieves a human-crafted template and refills its slots
 without touching layout or styling.
- **RAP-Qwen3.8-27B-RL** — Qwen3.8-27B post-trained on the triplets with full-parameter SFT and GRPO.
- **RAP-150** — the held-out benchmark: 150 papers, 10 per domain, with the exact model inputs used in the paper.
## Repository layout
| Folder | What it does | Entry point |
|---|---|---|
| `derender/` | Poster image → HTML template (Layout Detection → Font Size Estimation → Design Captioning → HTML Reconstruction) | `python -m derender.run_derender --image poster.png --out out/` |
| `posterrag/` | PosterRAG: Parser → Retrieval (categorical, narrative, aesthetic) → Mapping → Reviewer agents over the RAP-25K template pool | `python -m posterrag.run_posterrag --pdf paper.pdf --out out/` |
| `inference/` | Paper parser, prompt construction, generation and rendering for RAP-Qwen3.8-27B-RL | `python -m inference.generate benchmark|pdf ...` |
| `benchmark/rap_test150/` | The 150 benchmark prompts with thumbnails, ids and metadata | — |
| `figures/`, `docs/` | Paper figures used here; third-party notices | — |
## RAP-25K
| Conference | 2022 | 2023 | 2024 | 2025 | 2026 | Total |
|---|---:|---:|---:|---:|---:|---:|
| NeurIPS | 2,280 | 2,613 | 3,037 | 3,326 | – | **11,256** |
| ICLR | 830 | 877 | 1,447 | 2,085 | 3,265 | **8,504** |
| ICML | 882 | 1,471 | 1,751 | 1,571 | – | **5,675** |
| **Total** | 3,992 | 4,961 | 6,235 | 6,982 | 3,265 | **25,435** |
Each pair ships the poster image, the paper PDF, the annotations (`domain`, `task`, `method`, `keywords`,
`intro_summary`, title, abstract where available) and the derendered template (`layout.json`, `design_free.json`,
`generated.html` + `assets/`, `rendered.png`). The dataset card on Hugging Face documents the layout and loading.
## Poster derendering
Four stages tu