Paper2Web: Let's Make Your Paper Alive!
2510.15842v1
cs.CL, cs.CV
2025-10-21
Авторы:
Yuhang Chen, Tianpeng Lv, Siyi Zhang, Yixiang Yin, Yao Wan, Philip S. Yu, Dongping Chen
Abstract
Academic project websites can more effectively disseminate research when they
clearly present core content and enable intuitive navigation and interaction.
However, current approaches such as direct Large Language Model (LLM)
generation, templates, or direct HTML conversion struggle to produce
layout-aware, interactive sites, and a comprehensive evaluation suite for this
task has been lacking. In this paper, we introduce Paper2Web, a benchmark
dataset and multi-dimensional evaluation framework for assessing academic
webpage generation. It incorporates rule-based metrics like Connectivity,
Completeness and human-verified LLM-as-a-Judge (covering interactivity,
aesthetics, and informativeness), and PaperQuiz, which measures paper-level
knowledge retention. We further present PWAgent, an autonomous pipeline that
converts scientific papers into interactive and multimedia-rich academic
homepages. The agent iteratively refines both content and layout through MCP
tools that enhance emphasis, balance, and presentation quality. Our experiments
show that PWAgent consistently outperforms end-to-end baselines like
template-based webpages and arXiv/alphaXiv versions by a large margin while
maintaining low cost, achieving the Pareto-front in academic webpage
generation.
Ссылки и действия
Дополнительные ресурсы: