Recent advances in large language models (LLMs) have shown significant capabilities in code translation, often evaluated using benchmarks like CodeTransOcean. However, these evaluations typically focus on simple, function-level translations without considering dependencies, which does not reflect the complexities of real-world software development. Further, their effectiveness in translating to newer, lower-resource languages like Rust in realistic scenarios is still under-explored. To address this gap, we introduce first repository-level code translation benchmark comprising 375 tasks targeting Rust, complete with relevant dependencies. Using this benchmark, we study four state-of-the-art LLMs, analyzing their erroneous outputs to understand their performance in more complex translation scenarios. Our findings reveal that LLMs exhibit substantially worse performance (41.5%-56.2% Pass@1 drop of GPT-4) on repository-level translations compared to simpler tasks, highlighting limitations in existing evaluation methods. The model that performed the best is Claude-3.5, demonstrating the strongest translation capabilities in both basic functionality accuracy and several relevant additional abilities. Additionally, we discover that LLMs struggle with identifying language differences in complex tasks, and that increased dependencies correlate with greater translation difficulty.
翻译:暂无翻译