Applying personalized LLM strategies to dynamic mobile GUI environments

Develop methods to effectively apply both fine-tuning and training-free personalization strategies for large language model agents—specifically prompt engineering, retrieval-augmented generation, and external memory—to dynamic smartphone graphical user interface environments in order to achieve robust personalized assistance under changing UI states and interactions.

Background

The paper surveys two primary paradigms for personalization in LLM agents: parameter-based fine-tuning methods and training-free methods such as prompt engineering, retrieval-augmented generation, and external memory. These approaches have shown effectiveness in static tasks (e.g., travel planning).

However, smartphone GUI tasks are dynamic and interface-driven, with rapidly changing UI states, personalized content, and varied interaction paths. The authors explicitly note that transferring these personalization strategies to dynamic mobile GUI environments remains unresolved, highlighting a critical gap in current methods and evaluations.