Code readability assessment and improvement: human and automated perspectives

Abstract

Readability is a crucial characteristic in software development because developers often spend a significant amount of time reading and understanding code. However, there are still challenges both in accurately assessing code readability and in automatically applying refactoring operations to improve it. As large language models (LLMs) revolutionize the way code is written, a new opportunity emerges: to evaluate not only the readability of the code generated by these models but also their ability to refactor existing code into more readable versions. This study evaluates the readability of the code generated from three LLMs (ChatGPT, DeepSeek, and Gemini), comparing the code snippets they generate with those written by humans on Stack Overflow. Based on this evaluation, a hybrid approach is presented that combines automatic static analysis tools (ASATs) with LLM-based refactorings to improve code readability. In addition, a taxonomy of readability improvements is extracted from real-world GitHub repositories written in Java and Python. The results indicate that LLMs generate code with significantly fewer readability warnings compared to Stack Overflow recommendations. Combining LLMs and ASATs offers valuable potential for readability-oriented refactorings, especially when improving code through a single interaction with the LLMs. Code-generation-focused models, such as the DeepSeek Coder, achieved the best performance in refining source code. In real-world code, developers tend to improve readability by clarifying the intent of the code and reducing redundancy and verbosity. While Python and Java developers share concerns such as improving naming, some differences are observed: Java developers focus more on removing redundant and unnecessary code, whereas Python developers prioritize code documentation and correct formatting using standards such as the PEP 8 guidelines.