Large Language Models for Fault Localization: An Empirical Study

Large language models (LLMs) have demonstrated remarkable capabilities in code-related tasks, particularly in automated program repair. However, the effectiveness of such repairs is highly dependent on the performance of upstream fault localization, for which comprehensive evaluations are currently lacking. This paper presents a systematic empirical study on LLMs in the statement-level code fault localization task. We evaluate representative open-source models (Qwen2.5-coder-32b-instruct, DeepSeek-V3) and closed-source models (GPT-4.1 mini, Gemini-2.5-flash) to assess their fault localization capabilities on the HumanEval-Java and Defects4J datasets. The study investigates the impact of different prompting strategies--including standard prompts, few-shot examples, and chain-of-reasoning--on model performance, with a focus on analysis across accuracy, time efficiency, and economic cost dimensions. Our experimental results show that incorporating bug report context significantly enhances model performance. Few-shot learning shows potential for improvement but exhibits noticeable diminishing marginal returns, while chain-of-thought reasoning's effectiveness is highly contingent on the model's inherent reasoning capabilities. This study not only highlights the performance characteristics and trade-offs of different models in fault localization tasks, but also offers valuable insights into the strengths of current LLMs and strategies for improving fault localization effectiveness.