MiniAppBench: Evaluating the Shift from Text to Interactive HTML Responses in LLM-Powered Assistants

Researchers introduce MiniAppBench, a new benchmark for evaluating Large Language Models' ability to generate interactive HTML applications rather than static text responses. The benchmark includes 500 real-world tasks and an agentic evaluation framework called MiniAppEval that uses browser automation for testing.

• →Human-AI interaction is shifting from static text to dynamic, interactive HTML-based applications called MiniApps.
• →Current benchmarks fail to evaluate LLMs' capabilities for generating interactive applications with custom logic.
• →MiniAppBench provides 500 tasks across six domains sourced from real-world applications with 10M+ generations.
• →MiniAppEval framework uses browser automation to perform human-like testing across three evaluation dimensions.
• →Current LLMs still struggle significantly with generating high-quality interactive MiniApps.