LLM-Based Automated Program Repair: Bug Complexity, Fault Localization, and Cost-Efficiency

A new empirical study from arXiv (2608.14065) examines how bug complexity, fault localization, reasoning settings, and cost-efficiency affect Large Language Model (LLM)-based Automated Program Repair (APR). The research evaluates two APR techniques, ChatRepair and CodeCorrector, using three LLMs: DeepSeek, GPT, and Llama. Through a multi-dimensional empirical framework and statistical analysis, the study assesses performance across diverse bug complexity levels and localization strategies. Preliminary results indicate that structurally complex bugs and imprecise fault localization significantly impact repair effectiveness. The study aims to fill gaps left by prior research that focused primarily on overall repair effectiveness, offering insights into the nuanced factors that influence LLM-based APR. This work is relevant for software engineering researchers and practitioners interested in optimizing automated repair tools. The findings highlight the need for cost-efficient approaches in LLM-based APR, as the expense of running large models can be prohibitive. The study is available on arXiv under the identifier 2608.14065.