@inproceedings{sun-etal-2026-hierdiff,
title = "{H}ier{D}iff: Enhancing Large Language Models for Commit Message Generation with Hierarchical Diff Information",
author = "Sun, Jinrui and
Jia, Tong and
He, Minghua and
Li, Ying",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.451/",
pages = "9466--9480",
ISBN = "979-8-89176-462-0",
abstract = "Commit messages are vital artifacts that summarize changes to facilitate software maintenance, debugging, and evolution tracking. However, existing Commit Message Generation (CMG) methods rely solely on code diffs and overlook higher-level architectural dependency diff information. To address this limitation, we propose HierDiff, a framework using architectural diffs to complement code diffs for improved commit generation. Through this framework, we further construct HierCMD, a new dataset that incorporates architectural diffs alongside much longer code diffs. Building on this resource, we develop HierDiff-6.7B, a lightweight CMG model trained via multi-stage hierarchical diff learning. To enable more rigorous evaluation, we additionally extend the existing MCMD-New-Test benchmark by incorporating architectural diff information, resulting in MCMD-New-TestX. Experimental results on MCMD-New-TestX and HierCMD demonstrate that HierDiff-6.7B achieves state-of-the-art performance in CMG, even surpassing the renowned closed-source model GPT-4o by 2.1{\%} to 29.8{\%} across four evaluation metrics. We release the data and code at: https://anonymous.4open.science/r/HierDiff/."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sun-etal-2026-hierdiff">
<titleInfo>
<title>HierDiff: Enhancing Large Language Models for Commit Message Generation with Hierarchical Diff Information</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jinrui</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghua</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ying</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Commit messages are vital artifacts that summarize changes to facilitate software maintenance, debugging, and evolution tracking. However, existing Commit Message Generation (CMG) methods rely solely on code diffs and overlook higher-level architectural dependency diff information. To address this limitation, we propose HierDiff, a framework using architectural diffs to complement code diffs for improved commit generation. Through this framework, we further construct HierCMD, a new dataset that incorporates architectural diffs alongside much longer code diffs. Building on this resource, we develop HierDiff-6.7B, a lightweight CMG model trained via multi-stage hierarchical diff learning. To enable more rigorous evaluation, we additionally extend the existing MCMD-New-Test benchmark by incorporating architectural diff information, resulting in MCMD-New-TestX. Experimental results on MCMD-New-TestX and HierCMD demonstrate that HierDiff-6.7B achieves state-of-the-art performance in CMG, even surpassing the renowned closed-source model GPT-4o by 2.1% to 29.8% across four evaluation metrics. We release the data and code at: https://anonymous.4open.science/r/HierDiff/.</abstract>
<identifier type="citekey">sun-etal-2026-hierdiff</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.451/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>9466</start>
<end>9480</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T HierDiff: Enhancing Large Language Models for Commit Message Generation with Hierarchical Diff Information
%A Sun, Jinrui
%A Jia, Tong
%A He, Minghua
%A Li, Ying
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F sun-etal-2026-hierdiff
%X Commit messages are vital artifacts that summarize changes to facilitate software maintenance, debugging, and evolution tracking. However, existing Commit Message Generation (CMG) methods rely solely on code diffs and overlook higher-level architectural dependency diff information. To address this limitation, we propose HierDiff, a framework using architectural diffs to complement code diffs for improved commit generation. Through this framework, we further construct HierCMD, a new dataset that incorporates architectural diffs alongside much longer code diffs. Building on this resource, we develop HierDiff-6.7B, a lightweight CMG model trained via multi-stage hierarchical diff learning. To enable more rigorous evaluation, we additionally extend the existing MCMD-New-Test benchmark by incorporating architectural diff information, resulting in MCMD-New-TestX. Experimental results on MCMD-New-TestX and HierCMD demonstrate that HierDiff-6.7B achieves state-of-the-art performance in CMG, even surpassing the renowned closed-source model GPT-4o by 2.1% to 29.8% across four evaluation metrics. We release the data and code at: https://anonymous.4open.science/r/HierDiff/.
%U https://aclanthology.org/2026.findings-emnlp.451/
%P 9466-9480
Markdown (Informal)
[HierDiff: Enhancing Large Language Models for Commit Message Generation with Hierarchical Diff Information](https://aclanthology.org/2026.findings-emnlp.451/) (Sun et al., Findings 2026)
ACL