一句话：提出自动程序修复中的“信息迷失”问题，并用逻辑分解、细粒度上下文检索与块到整体推理加以缓解。
自动程序修复（APR）能减少人工调试负担并提升软件可靠性，但传统方法依赖精确的故障定位，代价高且往往不可得。考虑到许多缺陷边界模糊、修复跨度较大，本文在宽松定位约束下对 APR 展开初步研究。研究发现，由于上下文混乱或不足，大语言模型（LLM）在该设定下表现不佳，作者将这一挑战定义为“信息迷失”（Information Disorientation）。为应对该问题，论文提出一种新的 APR 方法，集成逻辑分解、细粒度上下文检索与块到整体（block-to-whole）推理三个环节，先从整体上拆解修复任务，再检索与缺陷块相关的精细上下文，最后以局部到整体的方式生成补丁。初步实验表明，该方法能够有效缓解信息迷失，提升宽松定位场景下的修复效果。
Automated Program Repair (APR) is essential for reducing manual debugging effort and enhancing software reliability. However, traditional approaches rely on precise fault localization, which is costly and often unavailable. Given that many bugs exhibit fuzzy boundaries and wide fix spans, we conduct a preliminary study on APR under relaxed localization constraints. Our findings reveal that Large Language Models (LLMs) struggle in these settings due to chaotic or insufficient context—a challenge we define as Information Disorientation. To address this, we propose a novel APR approach that integrates logical decomposition, fine-grained context retrieval, and block-to-whole reasoning. Preliminary experiments demonstrate that our method effectively mitigates information disorientation.
@inproceedings{paperbot11641,
title = {Exploring the Information Disorientation Problem in Automated Program Repair under Relaxed Localization},
author = {Jinrui Sun and Tong Jia and Minghua He and Ying Li},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3774748.3795674}
}