ChatGPT의 동시성 결함 진단 및 수정에 관한 효과 분석 연구

초록

Recent software development has increasingly adopted large language models (LLMs). Beyond code generation, LLMs are gainingattention as tools for diagnosing and fixing software defects. Among them, OpenAI's ChatGPT has demonstrated strong capabilities acrossvarious programming languages and bug types. This study evaluates ChatGPT's effectiveness in diagnosing and repairing concurrencybugs, which have received relatively limited attention in prior research. Concurrency bugs were categorized into seven types, and a datasetof 35 code examples (five per type) was used for evaluation. Six prompting strategies were applied to analyze the impact of promptdesign on performance. The results show that ChatGPT achieved an average diagnosis success rate of 91.4% and a fix success rate of60.5%, with particularly strong performance in Dead Lock, Race Condition, and Atomicity Violation cases.

키워드