Evaluating Source Code Quality with Large Language Models: a comparative study

TL;DR

This study explores the potential of large language models to evaluate source code quality, comparing their assessments with traditional static analysis tools like SonarQube across open source Java projects.

Contribution

It provides an empirical comparison of GPT 3.5 Turbo and GPT 4o in assessing code quality, highlighting their capabilities and limitations.

Findings

GPT 3.5 Turbo correlates with SonarQube metrics

GPT 4o diverges from traditional assessments

LLMs show potential but have limitations in code quality evaluation

Abstract

Code quality is an attribute composed of various metrics, such as complexity, readability, testability, interoperability, reusability, and the use of good or bad practices, among others. Static code analysis tools aim to measure a set of attributes to assess code quality. However, some quality attributes can only be measured by humans in code review activities, readability being an example. Given their natural language text processing capability, we hypothesize that a Large Language Model (LLM) could evaluate the quality of code, including attributes currently not automatable. This paper aims to describe and analyze the results obtained using LLMs as a static analysis tool, evaluating the overall quality of code. We compared the LLM with the results obtained with the SonarQube software and its Maintainability metric for two Open Source Software (OSS) Java projects, one with Maintainability Rating A and the other B. A total of 1,641 classes were analyzed, comparing the results in two versions of models: GPT 3.5 Turbo and GPT 4o. We demonstrated that the GPT 3.5 Turbo LLM has the ability to evaluate code quality, showing a correlation with Sonar's metrics. However, there are specific aspects that differ in what the LLM measures compared to SonarQube. The GPT 4o version did not present the same results, diverging from the previous model and Sonar by assigning a high classification to codes that were assessed as lower quality. This study demonstrates the potential of LLMs in evaluating code quality. However, further research is necessary to investigate limitations such as LLM's cost, variability of outputs and explore quality characteristics not measured by traditional static analysis tools.