TY - GEN
T1 - Fine-Grained Repair Knowledge Organization for Knowledge-Guided Automated Program Repair
AU - Jiang, Jindong
AU - Cheng, Haoxuan
AU - Liu, Yi
AU - Lin, Nan
AU - Yu, Shihang
AU - Zhu, Xiaoyan
AU - Wang, Jiayin
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2027.
PY - 2027
Y1 - 2027
N2 - Template-based automated program repair (APR) exposes explicit repair knowledge, but existing template inventories are often too coarse to act as reusable knowledge units. We therefore study fine-grained repair knowledge organization, which starts from coarse template families, builds multi-view representations of historical bug-fix pairs, discovers finer repair groups, and abstracts them into structured knowledge entries. The resulting knowledge base supports retrieval, classifier-based selection, and retrieval-to-classifier reranking, and is grounded in localization-enhanced prompts for patch generation. On an internal Recoder workflow, the approach constructs a non-trivial inventory of 184 accepted fine-grained templates and improves repair over bug-only conditioning. On Defects4J, retrieval is stronger in the top-1 regime, while larger candidate budgets favor classifier or rerank selection. On the full Defects4J 1.2 benchmark, knowledge-guided repair substantially outperforms the template-free bug-only baseline, and the same trend remains visible across both Qwen and Llama.
AB - Template-based automated program repair (APR) exposes explicit repair knowledge, but existing template inventories are often too coarse to act as reusable knowledge units. We therefore study fine-grained repair knowledge organization, which starts from coarse template families, builds multi-view representations of historical bug-fix pairs, discovers finer repair groups, and abstracts them into structured knowledge entries. The resulting knowledge base supports retrieval, classifier-based selection, and retrieval-to-classifier reranking, and is grounded in localization-enhanced prompts for patch generation. On an internal Recoder workflow, the approach constructs a non-trivial inventory of 184 accepted fine-grained templates and improves repair over bug-only conditioning. On Defects4J, retrieval is stronger in the top-1 regime, while larger candidate budgets favor classifier or rerank selection. On the full Defects4J 1.2 benchmark, knowledge-guided repair substantially outperforms the template-free bug-only baseline, and the same trend remains visible across both Qwen and Llama.
KW - Automated Program Repair
KW - Defects4J
KW - Large Language Models
KW - Repair Knowledge Organization
KW - Template-Based Repair
UR - https://www.scopus.com/pages/publications/105047053766
U2 - 10.1007/978-981-92-3450-9_50
DO - 10.1007/978-981-92-3450-9_50
M3 - 会议稿件
AN - SCOPUS:105047053766
SN - 9789819234493
T3 - Lecture Notes in Computer Science
SP - 610
EP - 622
BT - Advanced Intelligent Computing Technology and Applications - 22nd International Conference on Intelligent Computing, ICIC 2026, Proceedings
A2 - Huang, De-Shuang
A2 - Pan, Yijie
A2 - Zhang, Chuanlei
A2 - Chen, Wei
A2 - Premaratne, Prashan
PB - Springer Science and Business Media Deutschland GmbH
T2 - 22nd International Conference on Intelligent Computing, ICIC 2026
Y2 - 22 July 2026 through 26 July 2026
ER -