Quality Assessment of LLM-Generated Multi-Language Systems: Comparative Analysis Using Design Smells

Abstract Large Language Models (LLMs) are increasingly used to generate source code, including software that integrates multiple programming languages. While prior research has examined correctness and security aspects of LLM-generated code, little is known about its structural quality in multi-language settings, particularly with respect to cross-language design smells. Multi-language systems, such as those based on the Java Native Interface (JNI), introduce additional architectural complexity due to the need for coordination across language boundaries.In this paper, we present an exploratory empirical study investigating the prevalence and distribution of multi-language design smells in LLM-generated codes. We design a set of 80 JNI-based programming tasks covering 16 multi-language smells and generate implementations using four state-of-the-art LLMs: ChatGPT-5.2 , Claude Sonnet 4.6 , Gemini 3.1 pro , and Qwen3-Coder . The generated code is analyzed using a rule-based detection framework to identify the presence and distribution of multli-language design smells.Our results show that multi-language design smells are prevalent across all models, with up to 60% of generated implementations containing at least one smell. We observe significant variation across both models and smell types. Some smells occur consistently across all LLMs, suggesting shared structural tendencies, while others exhibit strong model-dependent behavior. These findings indicate that structural deficiencies in LLM-generated multi-language code are systematic and not uniformly distributed.Overall, this study highlights that architectural quality remains an underexplored dimension of LLM-based code generation, particularly in multi-language contexts. The results suggest that existing evaluation practices, which focus primarily on correctness, may overlook important structural issues. We argue for the need to incorporate multi-language structural quality assessment into future evaluation frameworks for LLM-generated codes.