HTMLCure: Browser Interaction Exposes LLM-Generated HTML Failures That Screenshots Miss

arXiv·medium signal

Wu et al. introduce HTMLCure, a framework that evaluates LLM-generated HTML by actually interacting with it — executing pages across viewports and interaction states (scroll, hover, click, resize, gameplay). Many LLM-generated pages render once correctly but fail under interaction. HTMLCure records deterministic browser traces and uses them to guide targeted repairs, turning 'looks right in a screenshot' into 'works in a browser.'