mUXSmellDetection: An LLM-Assisted Framework for Automated UX Smell Detection in Mobile Apps

Mobile applications increasingly support critical services, making high-quality user experience (UX) essential. Nevertheless, scalable UX evaluation remains challenging because it relies on expert judgment and heterogeneous evidence. This study introduces mUXSmellDetection, a catalog-driven, multi-modal framework for automated UX smell detection in Java-based Android applications. A Mobile UX Smells Catalog was developed, comprising 40 standardized UX smells with explicit detection rules and evidence requirements. The framework integrates runtime interaction logs collected through hybrid in-app and system-level logging, GUI screenshots, and Java AST-based code metadata, which are transformed into structured prompts for large language models (LLMs). The approach was evaluated on ten real-world Android applications using GPT-4o, GPT-5, and GPT-5 Thinking. Detection results were compared to an expert baseline established by three UX experts. GPT-5 achieved the highest overall performance, with 94.29% Precision, 89.19% Recall, and 91.67% F1-score. These findings indicate that multi-modal evidence enables scalable and practical automated UX smell detection.