Title: I-WebGenBench: Assessing Interactivity in LLM-Generated Scientific Web Applications

Abstract: The latest breakthroughs in visual language models have empowered autonomous agents to handle complex reasoning, document comprehension, and tool utilization. Despite these advancements, current document-focused agents are primarily limited to converting papers into static outputs like summaries, slides, or webpages. These static formats often fall short when addressing technical papers that rely on dynamic mechanisms and state transitions. To address this gap, we introduce a Paper-to-Interactive-System Agent designed to transform research papers into fully executable, interactive web platforms. This agent operates autonomously, processing a PDF paper from start to finish without human oversight. The workflow encompasses paper comprehension, system modeling, and the synthesis of interactive webpages, allowing users to actively manipulate inputs and witness dynamic behaviors in real time.

To assess the efficacy of this approach, we present a benchmark comprising 19 research papers, each paired with expert-developed interactive systems that serve as the ground truth. Additionally, we propose PaperVoyager, a structured generation framework that explicitly captures interaction logic and underlying mechanisms during the synthesis phase. Our experimental results demonstrate that PaperVoyager substantially enhances the quality of the generated interactive systems, establishing a novel paradigm for understanding scientific papers through interactivity.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC