Abstract

Generative UI, where interfaces are dynamically constructed by Large Language Models (LLMs) at runtime, promises unparalleled personalization. However, the foundational assumption—that providing a user persona to an LLM actually results in a measurably different interface—lacks empirical validation. This study investigates the "persona effect" in Agentic Computing by using the A2UI framework to generate 500 personalized dashboards for a diverse set of synthetic personas. Through quantitative analysis of layout structure, styling, and content curation, we discovered a "dual-signal pattern." While personas strongly dictate narrative and content choices, they have a surprisingly weak influence on numerical structure and core layout templates. This finding reveals significant constraints in current LLM-driven UI generation and directly impacts how developers must prompt and architect systems for true "UI on Demand."