Dataset: Exploring Architectural Smells Detection Through LLMs

Architectural Smells (AS) are design patterns in software systems that violate design principles, potentially compromising maintainability and overall quality. As software complexity grows, suboptimal architectural decisions can lead to technical debt, making system evolution more difficult and costly over time.

Early detection of AS is crucial to prevent this accumulation. Among the various types of AS, this study focuses on the Hub-Like Dependency (HL) smell, which occurs when an abstraction has many incoming and outgoing dependencies, forming a hub-like structure in the dependency graph.

Traditional detection methods rely on rule-based systems, which often miss architectural issues requiring a deeper understanding of design context. Large Language Models (LLMs) offer a promising alternative, as they can analyze large amounts of code and have shown potential in detecting code smells, vulnerabilities, and suggesting refactorings. However, no prior study has specifically investigated the use of LLMs for AS detection.

This research explores the detection of HL smells using Google’s Gemini 1.5 Pro and compares its performance with Arcan, a specialized AS detection tool. Additionally, the study evaluates whether LLMs can provide explanations and refactoring suggestions for the detected smells.