RAP compiles 25,000 paper-poster pairs to teach AI how to lay out posters in HTML/CSS
Designing a conference poster sits at the intersection of visual coding, design, and information compression. Researchers at Meituan collected 25,000+ verified pairs from conferences held between 2022 and 2026 and proposed two generation methods: PosterRAG, a training-free multi-agent pipeline, and SFT+RL, a recipe for training a dedicated model. The dataset includes a three-level content taxonomy and a benchmark of 150 held-out test papers. Each example is converted into a paper-poster-markup triplet: paper → poster → stylized HTML/CSS with typed, fillable slots. This setup lets researchers automatically verify quality by rendering the output. • RAP: 25K+ paper-poster pairs with a three-level content taxonomy • PosterRAG: a training-free multi-agent pipeline for filling slots • SFT+RL: a recipe for training a dedicated model on the triplets
AI-processed from Hugging Face Daily Papers; edited by Hamidun News
Why turn a poster into code
Designing an academic poster sits at the intersection of visual coding, information compression, and design. You have to fit the essence of a multi-page paper onto a single sheet while preserving hierarchy and aesthetics. Solid benchmarks for this task remain scarce: existing datasets have few examples and are annotated only with layout markup. That forces researchers to evaluate results using VLM-as-judge setups or checklists instead of objective rendering comparisons, while generation methods design layouts from scratch — leaving author-made posters untapped as a source of knowledge or a verified training signal.
RAP: 25,000 triplets from conferences held 2022–2026
Researchers at Meituan introduced RAP (Reconstruct A Poster) — a dataset and benchmark that frames poster generation as a visual coding task. The dataset includes:
• 25,000+ verified paper-poster pairs from machine learning conferences (2022–2026) • a three-level content taxonomy • a held-out benchmark of 150 test papers with their author-made posters
The key feature is a fully automated derendering pipeline that converts each poster into stylized HTML/CSS code with typed, fillable slots. The result is a paper-poster-markup triplet that can be verified by rendering it against the original. Every example doubles as both an inference case and a training example.
Two methods: from PosterRAG to SFT+RL
PosterRAG is a training-free multi-agent pipeline. It retrieves a similar poster example and fills its slots with content from the new paper, leaving the layout and styling untouched. This requires minimal compute and enables instant adaptation.
SFT+RL is a recipe for training a dedicated model on the triplets. Supervised fine-tuning plus reinforcement learning lets the model learn to structure content and fill slots using the full dataset.
Related work from 2026
The ecosystem of tools for visual generation is growing alongside RAP:
• AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards • InfoVis: Unified Slide Image Generation and Editing • From Prompting to Composing: A Spatial Canvas Interface for Poster Generation • Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation • Beyond Instruction-Driven Editing: Source-Grounded Problem Discovery with User-Governed Repair for Scientific Posters • ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation • Xiaomi-OCR-0 Technical Report
A demo is available at [reconstruct-a-poster.github.io](reconstruct-a-poster.github.io).
Authors: Haonan Han, Yukang Lin, Minliang Lin, Haoxian Tan, Jun-Yan He, Rui Yang, Zhi-Qi Cheng, Jie Hu, Guanglu Wan, Xiu Li (Meituan).
- Oct 29, 2026Sonnet 4.6 with Sourcegraph MCP Server Beats Fable 5: 0.698 vs 0.568
- Oct 29, 2026Stale Facts in Agent Memory Drop Qwen2.5 Accuracy from 100% to 44.6%
- Oct 29, 2026Skills for AI agents hurt weak models: a meta-skill recovered up to 14 points
- Oct 29, 2026AI agent on GPT-5.5 audits 2600 TravelLine autotests in two weeks
Key Questions Answered
Why turn a poster into code?
Designing an academic poster sits at the intersection of visual coding, information compression, and design. You have to fit the essence of a multi-page paper onto a single sheet while preserving hierarchy and aesthetics. Solid benchmarks for this task remain scarce: existing datasets have few examples and are annotated only with layout markup. That forces researchers to evaluate results using VLM-as-judge setups or checklists instead of objective rendering comparisons, while generation methods design layouts from scratch — leaving author-made posters untapped as a source o
What are the key details on RAP: 25,000 triplets from conferences held 2022–2026?
Researchers at Meituan introduced RAP (Reconstruct A Poster) — a dataset and benchmark that frames poster generation as a visual coding task. The dataset includes: 25,000+ verified paper-poster pairs from machine learning conferences (2022–2026) a three-level content taxonomy a held-out benchmark of 150 test papers with their author-made posters The key feature is a fully automated derendering pipeline that converts each poster into stylized HTML/CSS code with typed, fillable slots. The result is a paper-poster-markup triplet that can be verified by rendering it against the
What are the key details on Two methods: from PosterRAG to SFT+RL?
PosterRAG is a training-free multi-agent pipeline. It retrieves a similar poster example and fills its slots with content from the new paper, leaving the layout and styling untouched. This requires minimal compute and enables instant adaptation. SFT+RL is a recipe for training a dedicated model on the triplets. Supervised fine-tuning plus reinforcement learning lets the model learn to structure content and fill slots using the full dataset.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.