@inproceedings{xu-etal-2026-synthesizing,
title = "Synthesizing Code Completion Trajectories for Mid-training in Software Engineering",
author = "Xu, Siwei and
Liu, Qiyuan and
Li, Ruifan",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.234/",
pages = "4619--4630",
ISBN = "979-8-89176-444-6",
abstract = "Large language models (LLMs) have shown remarkable capabilities in interactive real-world software engineering (SWE) scenarios. However, the execution-aware training data are hard to scale and are mostly used for post-training, leaving agentic mid-training underexplored. To this end, we propose COMP-SYNTH, a novel automated pipeline for synthesizing reflective code completion trajectories. Specifically, based on executable repository environments, COMP-SYNTH identifies test-covered target functions, masks their implementations, and prompts an LLM to complete them under execution feedback. Then, each generated completion is verified by running the associated tests. Furthermore, failed attempts are augmented with diagnostic signals to extract useful supervision. Thus, we construct COMP-TRAJ, a mid-training dataset that exposes models to repository context, execution feedback, and iterative code revision. Extensive experiments on repository-level SWE benchmarks show that mid-training on COMP-TRAJ effectively enhances capabilities in SWE. Our code and data are publicly available at https://github.com/uuam5/comp-traj."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-synthesizing">
<titleInfo>
<title>Synthesizing Code Completion Trajectories for Mid-training in Software Engineering</title>
</titleInfo>
<name type="personal">
<namePart type="given">Siwei</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qiyuan</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifan</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Large language models (LLMs) have shown remarkable capabilities in interactive real-world software engineering (SWE) scenarios. However, the execution-aware training data are hard to scale and are mostly used for post-training, leaving agentic mid-training underexplored. To this end, we propose COMP-SYNTH, a novel automated pipeline for synthesizing reflective code completion trajectories. Specifically, based on executable repository environments, COMP-SYNTH identifies test-covered target functions, masks their implementations, and prompts an LLM to complete them under execution feedback. Then, each generated completion is verified by running the associated tests. Furthermore, failed attempts are augmented with diagnostic signals to extract useful supervision. Thus, we construct COMP-TRAJ, a mid-training dataset that exposes models to repository context, execution feedback, and iterative code revision. Extensive experiments on repository-level SWE benchmarks show that mid-training on COMP-TRAJ effectively enhances capabilities in SWE. Our code and data are publicly available at https://github.com/uuam5/comp-traj.</abstract>
<identifier type="citekey">xu-etal-2026-synthesizing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.234/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4619</start>
<end>4630</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Synthesizing Code Completion Trajectories for Mid-training in Software Engineering
%A Xu, Siwei
%A Liu, Qiyuan
%A Li, Ruifan
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F xu-etal-2026-synthesizing
%X Large language models (LLMs) have shown remarkable capabilities in interactive real-world software engineering (SWE) scenarios. However, the execution-aware training data are hard to scale and are mostly used for post-training, leaving agentic mid-training underexplored. To this end, we propose COMP-SYNTH, a novel automated pipeline for synthesizing reflective code completion trajectories. Specifically, based on executable repository environments, COMP-SYNTH identifies test-covered target functions, masks their implementations, and prompts an LLM to complete them under execution feedback. Then, each generated completion is verified by running the associated tests. Furthermore, failed attempts are augmented with diagnostic signals to extract useful supervision. Thus, we construct COMP-TRAJ, a mid-training dataset that exposes models to repository context, execution feedback, and iterative code revision. Extensive experiments on repository-level SWE benchmarks show that mid-training on COMP-TRAJ effectively enhances capabilities in SWE. Our code and data are publicly available at https://github.com/uuam5/comp-traj.
%U https://aclanthology.org/2026.findings-aacl.234/
%P 4619-4630
Markdown (Informal)
[Synthesizing Code Completion Trajectories for Mid-training in Software Engineering](https://aclanthology.org/2026.findings-aacl.234/) (Xu et al., Findings 2026)
ACL