LLMs Love Python: A Study of LLMs' Bias (arxiv.org)

A groundbreaking study has evaluated the preferences of large language models (LLMs) for programming languages and libraries in code generation. This research examined eight diverse LLMs and revealed a pronounced bias towards popular libraries like NumPy, with instances of misuse occurring in up to 45% of cases where the library was not required. Furthermore, Python emerged as the preferred language for code generation, dominating 58% of instances even in high-performance scenarios where it may not be the best choice, while other languages like Rust were not utilized at all. This study is significant for the AI/ML community as it highlights the need for a deeper understanding of LLMs' decision-making processes, particularly regarding language and library selection. The findings suggest that current LLMs prioritize familiarity and popularity over task-specific efficacy, raising important questions about their reliability in producing optimal code. As a result, the study calls for improvements in fine-tuning techniques, diversification of training data, and the establishment of benchmarks that evaluate LLM performance with respect to language and library choices. These advancements could enhance the practical utility of LLMs in software development and coding tasks.