Reproduce “Catalyzing Interactive HTML Generation Capabilities via High-Fidelity Reinforcement Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xujia Wang, Yixuan Liu, Shaohan Huang, Furu Wei, Lei Hou, Bin Xu, Juanzi Li

WebCatalyst trains models to generate interactive HTML by assessing browser-run interactions and visuals, penalizing visual degradation against a reference and tracing successful runtime actions to code tokens. After 640 processed requests, Cookie-Bench Mini Games Total rises from 68.8 to 75.3, while WebGen-Bench Accuracy rises from 48.8 to 56.9. Results cover one backbone and these benchmarks, not arbitrary web applications.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.