Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair

Oebele Lijzenga; Iman Hemati Moghadam; Vadim Zaytsev

doi:10.1145/3672608.3707774

Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair

Oebele Lijzenga
, Iman Hemati Moghadam^*
, Vadim Zaytsev

^*Corresponding author for this work

Formal Methods and Tools

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

1 Citation (Scopus)

4 Downloads (Pure)

Abstract

Background. Automated Program Repair (APR) techniques often face challenges in navigating vast search space of possible patches and often rely on redundancy-based assumptions, which can restrict the diversity of generated patches. Recently, Code Language Models (CLMs) have emerged as a method for dynamically generating patch ingredients, potentially enhancing patch quality.

Aim. This study aims to enhance APR by integrating search-based methods with CLMs to improve both the quality of generated patch ingredients and the efficiency of the search process.

Method. We propose ARJACLM, a novel APR technique that uses a genetic algorithm for search space navigation and dynamically generates patch ingredients with the CodeLLaMA-13B model, combining redundancy-based and CLM-derived patch ingredients.

Results. Testing on 176 bugs across 9 Java projects from Defect4J shows that CLM-generated patch ingredients significantly boost ARJACLM's performance, though at the cost of increased computation time. ARJACLM outperforms ARJA and GenProg, and CLM-generated patch ingredients are of higher quality than their redundancy-based counterparts. Additionally, ARJACLM performs best when redundancy-based patch ingredients are ignored.

Original language	English
Title of host publication	SAC '25
Subtitle of host publication	Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing
Editors	Jiman Hong, Sebastiano Battiato, Christian Esposito
Place of Publication	New York, NY
Publisher	Association for Computing Machinery (ACM)
Pages	1627-1636
Number of pages	10
ISBN (Electronic)	979-8-4007-0629-5
DOIs	https://doi.org/10.1145/3672608.3707774
Publication status	Published - 14 May 2025
Event	40th Annual ACM Symposium on Applied Computing, SAC 2025 - Catania International Airport Hotel, Catania, Italy Duration: 31 Mar 2025 → 4 Apr 2025 Conference number: 40 https://www.sigapp.org/sac/sac2025/

Conference

Conference	40th Annual ACM Symposium on Applied Computing, SAC 2025
Abbreviated title	SAC 2025
Country/Territory	Italy
City	Catania
Period	31/03/25 → 4/04/25
Internet address	https://www.sigapp.org/sac/sac2025/

Keywords

Code language model
Program repair
Search-based algorithm

Access to Document

10.1145/3672608.3707774Licence: CC BY

ArticleFinal published version, 1.43 MBLicence: CC BY

Cite this

Lijzenga, Oebele ; Hemati Moghadam, Iman ; Zaytsev, Vadim. / Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair. SAC '25: Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing. editor / Jiman Hong ; Sebastiano Battiato ; Christian Esposito. New York, NY : Association for Computing Machinery (ACM), 2025. pp. 1627-1636

@inproceedings{52b771c28e9f483ca9b18ef424fa58b3,

title = "Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair",

abstract = "Background. Automated Program Repair (APR) techniques often face challenges in navigating vast search space of possible patches and often rely on redundancy-based assumptions, which can restrict the diversity of generated patches. Recently, Code Language Models (CLMs) have emerged as a method for dynamically generating patch ingredients, potentially enhancing patch quality.Aim. This study aims to enhance APR by integrating search-based methods with CLMs to improve both the quality of generated patch ingredients and the efficiency of the search process.Method. We propose ARJACLM, a novel APR technique that uses a genetic algorithm for search space navigation and dynamically generates patch ingredients with the CodeLLaMA-13B model, combining redundancy-based and CLM-derived patch ingredients.Results. Testing on 176 bugs across 9 Java projects from Defect4J shows that CLM-generated patch ingredients significantly boost ARJACLM's performance, though at the cost of increased computation time. ARJACLM outperforms ARJA and GenProg, and CLM-generated patch ingredients are of higher quality than their redundancy-based counterparts. Additionally, ARJACLM performs best when redundancy-based patch ingredients are ignored.",

keywords = "Code language model, Program repair, Search-based algorithm",

author = "Oebele Lijzenga and \{Hemati Moghadam\}, Iman and Vadim Zaytsev",

note = "Publisher Copyright: Copyright {\textcopyright} 2025 held by the owner/author(s).; 40th Annual ACM Symposium on Applied Computing, SAC 2025, SAC 2025 ; Conference date: 31-03-2025 Through 04-04-2025",

year = "2025",

month = may,

day = "14",

doi = "10.1145/3672608.3707774",

language = "English",

pages = "1627--1636",

editor = "Jiman Hong and Sebastiano Battiato and Christian Esposito",

booktitle = "SAC '25",

publisher = "Association for Computing Machinery (ACM)",

address = "United States",

url = "https://www.sigapp.org/sac/sac2025/",

}

Lijzenga, O, Hemati Moghadam, I & Zaytsev, V 2025, Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair. in J Hong, S Battiato & C Esposito (eds), SAC '25: Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing. Association for Computing Machinery (ACM), New York, NY, pp. 1627-1636, 40th Annual ACM Symposium on Applied Computing, SAC 2025, Catania, Italy, 31/03/25. https://doi.org/10.1145/3672608.3707774

Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair. / Lijzenga, Oebele; Hemati Moghadam, Iman; Zaytsev, Vadim.
SAC '25: Proceedings of the 40th ACM/SIGAPP Symposium on Applied Computing. ed. / Jiman Hong; Sebastiano Battiato; Christian Esposito. New York, NY: Association for Computing Machinery (ACM), 2025. p. 1627-1636.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Academic › peer-review

TY - GEN

T1 - Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair

AU - Lijzenga, Oebele

AU - Hemati Moghadam, Iman

AU - Zaytsev, Vadim

N1 - Conference code: 40

PY - 2025/5/14

Y1 - 2025/5/14

N2 - Background. Automated Program Repair (APR) techniques often face challenges in navigating vast search space of possible patches and often rely on redundancy-based assumptions, which can restrict the diversity of generated patches. Recently, Code Language Models (CLMs) have emerged as a method for dynamically generating patch ingredients, potentially enhancing patch quality.Aim. This study aims to enhance APR by integrating search-based methods with CLMs to improve both the quality of generated patch ingredients and the efficiency of the search process.Method. We propose ARJACLM, a novel APR technique that uses a genetic algorithm for search space navigation and dynamically generates patch ingredients with the CodeLLaMA-13B model, combining redundancy-based and CLM-derived patch ingredients.Results. Testing on 176 bugs across 9 Java projects from Defect4J shows that CLM-generated patch ingredients significantly boost ARJACLM's performance, though at the cost of increased computation time. ARJACLM outperforms ARJA and GenProg, and CLM-generated patch ingredients are of higher quality than their redundancy-based counterparts. Additionally, ARJACLM performs best when redundancy-based patch ingredients are ignored.

AB - Background. Automated Program Repair (APR) techniques often face challenges in navigating vast search space of possible patches and often rely on redundancy-based assumptions, which can restrict the diversity of generated patches. Recently, Code Language Models (CLMs) have emerged as a method for dynamically generating patch ingredients, potentially enhancing patch quality.Aim. This study aims to enhance APR by integrating search-based methods with CLMs to improve both the quality of generated patch ingredients and the efficiency of the search process.Method. We propose ARJACLM, a novel APR technique that uses a genetic algorithm for search space navigation and dynamically generates patch ingredients with the CodeLLaMA-13B model, combining redundancy-based and CLM-derived patch ingredients.Results. Testing on 176 bugs across 9 Java projects from Defect4J shows that CLM-generated patch ingredients significantly boost ARJACLM's performance, though at the cost of increased computation time. ARJACLM outperforms ARJA and GenProg, and CLM-generated patch ingredients are of higher quality than their redundancy-based counterparts. Additionally, ARJACLM performs best when redundancy-based patch ingredients are ignored.

KW - Code language model

KW - Program repair

KW - Search-based algorithm

UR - https://www.scopus.com/pages/publications/105006466520

U2 - 10.1145/3672608.3707774

DO - 10.1145/3672608.3707774

M3 - Conference contribution

AN - SCOPUS:105006466520

SP - 1627

EP - 1636

BT - SAC '25

A2 - Hong, Jiman

A2 - Battiato, Sebastiano

A2 - Esposito, Christian

PB - Association for Computing Machinery (ACM)

CY - New York, NY

T2 - 40th Annual ACM Symposium on Applied Computing, SAC 2025

Y2 - 31 March 2025 through 4 April 2025

ER -

Leveraging Search-Based and Pre-Trained Code Language Models for Automated Program Repair

Abstract

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this