Evaluating Large Language Models for Arduino Code Generation

Abstract

Large language models (LLMs), also known as generative AI, have transformed code generation by translating natural language prompts into executable code. Yet, their capabilities in generating code for resource-constrained devices such as Arduino, which are used in the Internet of Things and embedded systems, remained underexplored. This study evaluates six state-of-the-art LLMs for generating correct, efficient, and high-quality Arduino code. The evaluation was performed across five dimensions, namely functional correctness, runtime efficiency, memory usage, code quality, similarity to human-written code, and multi-round error correction. The results reveal that ChatGPT-4o achieves the highest zero-shot functional correctness and aligns closely with human code in readability and similarity. On the other hand, Gemini 2.0 Flash generates faster-executing code but at the cost of higher code complexity and lower similarity. DeepSeek-V3 balances correctness with superior flash memory optimization, whereas Claude 3.5 Sonnet struggles with prompt adherence. Finally, multi-round error correction improves correctness across all six models. Overall, the f indings underscore that none of the evaluated LLMs consistently outperforms all evaluation criteria. Hence, model choice must align with project priorities; as shown, ChatGPT-4o excels in functional correctness, whereas Gemini 2.0 excels in execution time, and DeepSeek-V3 in memory efficiency. This study provides a systematic evaluation of code generated with LLMs for Arduino, which, to the best of our knowledge, has not been previously studied across multiple models and performance metrics, thereby establishing a foundation for future research and contributing to enhancing the trustworthiness and effectiveness of LLM-generated code.