LLMs on code smell detection and refactoring: a systematic literature review

Abstract

Large Language Models (LLMs) are increasingly being applied to software maintenance tasks such as code smell detection and code refactoring, yet the evidence on their use remains fragmented across heterogeneous tasks, evaluation settings, and tool architectures. In particular, there is still limited consolidated knowledge about the prompt-based use of frozen-parameter LLMs for these activities. This paper presents a systematic literature review of how such models have been applied and evaluated in code smell detection and refactoring. Following a predefined protocol, we combined database searches and snowballing and used source-grounded AI assistance during screening under human oversight, yielding a final corpus of 65 studies. The review shows that this is a recent and rapidly expanding research area, concentrated around a small number of model families, programming languages, and canonical smells and refactorings. Across the reviewed studies, standalone prompting was generally less reliable in structurally demanding scenarios, whereas hybrid and repository-aware approaches were more often associated with stronger outcomes. Retrieval-augmented generation, validation layers, and iterative workflows repeatedly emerged as key mechanisms for improving reliability. Beyond technical performance, the evidence also indicates that practical usefulness depends on transparency, iterative interaction, and visible developer control. Overall, this review consolidates a still-dispersed body of work and provides a focused basis for future research and for the design of context-aware, validated, and human-supervised LLM-based maintenance tools.