Abstract
Large language models (LLM) provide opportunities to generate user interfaces directly from natural-language intent, but the emerging generative-UI literature remains focused primarily on web and chat contexts rather than smart systems. Smart-system interfaces are characterized by stricter demands, necessitating grounding in concrete device registries, validation against real-world action consequences, and explainability to non-developer users. This paper identifies and compares six distinct strategies currently used for LLM-driven UI generation (direct structured-output prompting, template-augmented generation, retrieval-augmented generation, model-driven hybrid pipelines, constrained generation, and multi-agent decomposition) using criteria tailored to smart-system contexts: output consistency, safety and accuracy validation, explainability, latency, implementation complexity, generalisability, and failure handling. The comparison is then adapted to three representative domains: smart-home consumer IoT, smart industrial supervisory control, and smart cultural heritage (virtual and AR museums, AI-assisted reconstruction interfaces, and smart exhibit displays). For each domain, a primary strategy is recommended, modifications required to meet specific domain constraints are demonstrated, and an adapted reference architecture is presented. A decision matrix and a selection flowchart are developed to guide the selection of an LLM-based UI generation strategy for specific smart-system deployments. The paper closes by outlining several open challenges, including the verification of accuracy-critical generated interfaces and the integration of static generation with runtime adaptation.

This publication has 18 references indexed in Scilit: