RGFL: Reasoning Guided Fault Localization for Automated Program Repair Using Large Language Models

TL;DR

This paper introduces a hierarchical reasoning approach for fault localization in automated program repair using large language models, significantly improving localization accuracy and repair success rates on Python and Java projects.

Contribution

It proposes a novel hierarchical reasoning module and a two-stage ranking scheme that enhance project-level fault localization for LLM-based repair agents.

Findings

File-level Hit@1 improved from 71.4% to 85%.

Element-level Exact Match under top-3 files increased from 36% to 69%.

End-to-end repair success improved by 12.8%.

Abstract

Fault Localization (FL) is a critical step in Automated Program Repair (APR), and its importance has increased with the rise of Large Language Model (LLM)-based repair agents. In realistic project-level repair scenarios, software repositories often span millions of tokens, far exceeding current LLM context limits. Consequently, models must first identify a small, relevant subset of code, making accurate FL essential for effective repair. We present a novel project-level FL approach that improves both file- and element-level localization. Our method introduces a hierarchical reasoning module that (i) generates structured, bug-specific explanations for candidate files and elements, and (ii) leverages these explanations in a two-stage ranking scheme combining LLM-based and embedding-based signals. We further propose a counterfactual upper-bound analysis to quantify the contribution of each localization stage to repair success. We evaluate our approach on Python and Java projects from SWE-bench Verified, Lite, and Java. Compared to state-of-the-art baselines, including Agentless and OpenHands, our method consistently improves localization accuracy. On SWE-bench Verified, file-level Hit@1 improves from 71.4% to 85%, and MRR from 81.8% to 88.8%. At the element level, Exact Match under top-3 files increases from 36% to 69%. Integrating our localization into Agentless yields a 12.8% end-to-end repair success improvement.