Paper2Poster (PosterAgent), an open-source multi-agent system that automatically generates editable academic posters from a scientific paper's PDF, has been released. It is available on GitHub with an accompanying arXiv paper, and has been accepted to the NeurIPS 2025 Dataset and Benchmark Track.
NeurIPS 2025 · Datasets & Benchmarks Track
Paper2Poster: a PDF becomes an editable academic poster — automatically
An open-source multi-agent pipeline (PosterAgent) condenses a full paper into a single, visually coherent poster — matching human quality while cutting token cost by 87%.
87%
less token consumption vs GPT-4o systems
100
AI conference papers in the benchmark dataset
3,800+
GitHub stars gathered
MIT
fully open-source license
The compression problem
One paper must collapse into a single page. Below, the average scale of an input paper vs its one-page poster output.
22.6 pages
~12,000 words · 22.6 figures
→
1 page
editable .pptx poster
The PosterAgent pipeline
1 · PARSER
Extracts structured assets — figures, tables and text.
→
2 · PLANNER
Lays out text–visual pairs via a binary-tree structure.
→
3 · PAINTER ⟳ COMMENTER
VLM feedback loop fixes rendering and removes overflow.
A "visual-in-the-loop" approach — interleaving figures and text into a coherent, editable page.
Why it lands well
Structural coherence & readability close to human-made posters
Parallel generation and VLM feedback for high-quality output
Easy to try via Gradio demo; Docker, YAML styling, auto logos
Where it still falls short
Hierarchical understanding of long papers can miss
Fine-grained readability sometimes weak
Local run needs OpenAI API key, LibreOffice, poppler
Evaluation suite
PosterAgent (4o_4o) outperformed GPT-4o systems on nearly all metrics — measured across four dimensions:
Visual Quality
Textual Coherence
VLM-as-Judge
PaperQuiz
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…