From Detection to Refactoring: A Controlled Study of LLMs for Java Code Smells

Code smells are symptoms of design and implementation choices that may hinder software maintainability and evolvability. Large Language Models (LLMs) have shown potential for code smell detection and refactoring, but further evidence is needed on how analysis context and localization guidance affect LLM performance in these tasks. We present a controlled empirical study of GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5 on a Java dataset covering 14 code smell types. Detection is evaluated in mixed-smell and smell-specific settings, while refactoring is assessed with and without explicit smell locations. Results show model- and smell-dependent behavior: GPT-5 provides the most balanced detection performance, while all models remove most target smell occurrences, particularly when localization guidance is provided. However, generated transformations may include unrelated modifications or compilation issues. These findings support developer-in-the-loop workflows in which LLM-generated refactorings are validated before acceptance.