TLDR: AutoPage is a novel multi-agent AI system that automates the creation of interactive project webpages from academic papers. It uses a collaborative, coarse-to-fine pipeline with “Checker” agents and optional human feedback to ensure high content and visual quality. The system is highly efficient, generating pages in under 15 minutes for less than $0.1, and is evaluated using a new benchmark called PageBench.
Researchers often spend valuable time creating project webpages to make their academic papers more accessible and engaging. This manual process can be repetitive and lead to inconsistent quality. A new multi-agent system called AutoPage aims to automate this task, transforming dense academic papers into dynamic, interactive project webpages efficiently and affordably.
AutoPage tackles the challenge of webpage creation by breaking it down into a structured, collaborative process. Unlike traditional end-to-end automation methods that might produce generic or unappealing layouts, AutoPage employs a “coarse-to-fine” pipeline. This means it first plans the overall narrative and structure of the webpage, then progressively refines the details, including multimodal content and interactive elements.
How AutoPage Works: A Collaborative Approach
The system operates through three main phases, each designed to ensure high quality and accuracy:
1. Narrative Planning and Structuring: AutoPage begins by parsing the original PDF paper into a structured format. A “Paper Content Parser” extracts key information, organizing text, figures, and tables. Following this, a “Page Content Planner” designs a web-optimized narrative flow, creating a logical outline for the project page. This initial blueprint is then verified for completeness and soundness.
2. Multimodal Content Generation: In this phase, the system populates the planned structure with rich content. A “Text Content Generator” synthesizes human-readable paragraphs from the paper’s core information. Then, a “Visual Content Generator” selects and places the most relevant figures and tables, ensuring they align semantically with the accompanying text. To combat potential AI inaccuracies, a “Content Checker” verifies consistency between text and visuals. Authors also have the option for “Human-in-the-Loop Refinement” to provide feedback and make adjustments, ensuring the content perfectly matches their vision.
3. Interactive Page Rendering: Once the content modules are approved, AutoPage moves to rendering the interactive webpage. A “Page Template Matcher” helps users select a design template from a curated library, or the system can choose one autonomously. The “HTML Generator” then compiles the final HTML, CSS, and JavaScript files, integrating interactive features. A final “HTML Checker” inspects the page for layout and visual integrity, and another “Human-in-the-Loop” checkpoint allows authors to fine-tune styles and make final corrections.
PageBench: A New Standard for Evaluation
To rigorously validate AutoPage and encourage further research in this area, the creators also developed PageBench. This is the first benchmark dataset specifically designed for automated paper-to-page generation. It includes over 1,500 academic papers paired with their corresponding human-created project pages. PageBench uses a comprehensive evaluation protocol that assesses both content quality (readability, semantic fidelity, information accuracy) and visual quality (visual content accuracy, layout, and aesthetic appeal).
Also Read:
- Beyond Simple Success: Graph Evaluation for Web Agents
- Making AI Research Reproducible: The Executable Knowledge Graph Approach
Impressive Results and Efficiency
Experiments show that AutoPage significantly improves upon existing end-to-end methods, enhancing both content and visual quality. It also helps to narrow the performance gap between different underlying AI models. The system excels as a “visual architect,” creating aesthetically pleasing and well-structured webpages. A user study confirmed that AutoPage-generated pages are highly preferred by human evaluators, achieving the highest average score. Crucially, AutoPage is remarkably efficient and cost-effective, capable of generating a complete, high-quality project page in under 15 minutes for less than $0.1, depending on the model used.
This innovative system offers a powerful solution for researchers, freeing them from the tedious task of webpage creation and allowing them to focus more on their core research. For more details, you can refer to the full research paper here.
