Evaluating Programming Language Confusion (2503.13620v1)

Abstract: LLMs for code (Code LLMs) have gained significant traction in software engineering, achieving state-of-the-art performance on various programming tasks including code completion, generation, repair, and translation. These models have demonstrated remarkable capabilities in understanding programming concepts, implementing algorithms, and even bridging different programming languages, fundamentally transforming how developers interact with coding environments. Despite these advances, Code LLMs often struggle with programming language confusion--producing code in unintended languages despite explicit instructions or obvious context. We systematically evaluate this phenomenon across diverse programming contexts. Our study assesses seven popular general and Code LLMs across multiple natural and programming languages, analyzing their behavior using four datasets (HumanEval, HumanEval-xl, MBPP, TP3) for code generation and one dataset (CodeNet) for code translation. The study results reveal that language confusion occurs across all evaluated models, with StarCoder and CodeLlama exhibiting the highest confusion rates. Even high-performing models fail to maintain language consistency throughout generated solutions, particularly when handling complex algorithmic problems. We identify key factors contributing to this confusion, including syntactic similarities between programming languages and inconsistent prompt formatting. Interestingly, we find evidence suggesting that LLMs consistently exhibit strategic language migration behaviors, prioritizing languages where they can produce more syntactically correct code even when explicitly instructed otherwise. This phenomenon is particularly pronounced in code generation tasks, where models show strong migration patterns toward Python and between syntactically similar language pairs.