@inproceedings{nguyen-etal-2026-fast,
title = "Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval",
author = "Nguyen, Long S. T. and
Huynh, Dat and
Vo, Quynh and
Hai, LE Dang Nguyen and
Nguyen, Cong-Duy T and
Luu, Anh Tuan and
Nguyen-Duc, Anh and
Quan, Tho",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.81/",
pages = "1856--1874",
ISBN = "979-8-89176-459-0",
abstract = "Large language models are effective at code completion when the required evidence is local, but repository-level fill-in-the-middle completion remains challenging when it depends on code and structural relationships across files. We present ACAR, a framework that combines lexical and AST/LSP-based structural retrieval through line-level reciprocal rank fusion, then applies syntax-aware post-processing to align generated completions with the surrounding code. We evaluate ACAR on RepoEval, ReccEval, and CrossCodeEval across Python, Java, TypeScript, and C{\#}, covering more than 3,000 real-world repositories. In the main Qwen2.5-Coder experiments, ACAR outperforms strong retrieval baselines in most settings, reaching 42.98 EM on ReccEval and 32.86 on CrossCodeEval at 3B, versus 36.25 and 28.40 for the strongest retrieval baseline, respectively. With the largest tested Qwen and DeepSeek backbones, ACAR improves EM over their corresponding no-retrieval baselines across all five evaluation settings. Ablations show that lexical and structural evidence are complementary, while removing syntax-aware refinement causes the largest drop in edit similarity. These gains come with 87{--}95 ms of non-decoding overhead under warm-cache conditions across the three benchmarks, supporting latency-sensitive code completion. To support reproducibility, we make our implementation and evaluation scripts publicly available at https://github.com/longstnguyen/ACAR."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nguyen-etal-2026-fast">
<titleInfo>
<title>Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval</title>
</titleInfo>
<name type="personal">
<namePart type="given">Long</namePart>
<namePart type="given">S</namePart>
<namePart type="given">T</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dat</namePart>
<namePart type="family">Huynh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Quynh</namePart>
<namePart type="family">Vo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">LE</namePart>
<namePart type="given">Dang</namePart>
<namePart type="given">Nguyen</namePart>
<namePart type="family">Hai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cong-Duy</namePart>
<namePart type="given">T</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anh</namePart>
<namePart type="given">Tuan</namePart>
<namePart type="family">Luu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anh</namePart>
<namePart type="family">Nguyen-Duc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tho</namePart>
<namePart type="family">Quan</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Large language models are effective at code completion when the required evidence is local, but repository-level fill-in-the-middle completion remains challenging when it depends on code and structural relationships across files. We present ACAR, a framework that combines lexical and AST/LSP-based structural retrieval through line-level reciprocal rank fusion, then applies syntax-aware post-processing to align generated completions with the surrounding code. We evaluate ACAR on RepoEval, ReccEval, and CrossCodeEval across Python, Java, TypeScript, and C#, covering more than 3,000 real-world repositories. In the main Qwen2.5-Coder experiments, ACAR outperforms strong retrieval baselines in most settings, reaching 42.98 EM on ReccEval and 32.86 on CrossCodeEval at 3B, versus 36.25 and 28.40 for the strongest retrieval baseline, respectively. With the largest tested Qwen and DeepSeek backbones, ACAR improves EM over their corresponding no-retrieval baselines across all five evaluation settings. Ablations show that lexical and structural evidence are complementary, while removing syntax-aware refinement causes the largest drop in edit similarity. These gains come with 87–95 ms of non-decoding overhead under warm-cache conditions across the three benchmarks, supporting latency-sensitive code completion. To support reproducibility, we make our implementation and evaluation scripts publicly available at https://github.com/longstnguyen/ACAR.</abstract>
<identifier type="citekey">nguyen-etal-2026-fast</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.81/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1856</start>
<end>1874</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval
%A Nguyen, Long S. T.
%A Huynh, Dat
%A Vo, Quynh
%A Hai, LE Dang Nguyen
%A Nguyen, Cong-Duy T.
%A Luu, Anh Tuan
%A Nguyen-Duc, Anh
%A Quan, Tho
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F nguyen-etal-2026-fast
%X Large language models are effective at code completion when the required evidence is local, but repository-level fill-in-the-middle completion remains challenging when it depends on code and structural relationships across files. We present ACAR, a framework that combines lexical and AST/LSP-based structural retrieval through line-level reciprocal rank fusion, then applies syntax-aware post-processing to align generated completions with the surrounding code. We evaluate ACAR on RepoEval, ReccEval, and CrossCodeEval across Python, Java, TypeScript, and C#, covering more than 3,000 real-world repositories. In the main Qwen2.5-Coder experiments, ACAR outperforms strong retrieval baselines in most settings, reaching 42.98 EM on ReccEval and 32.86 on CrossCodeEval at 3B, versus 36.25 and 28.40 for the strongest retrieval baseline, respectively. With the largest tested Qwen and DeepSeek backbones, ACAR improves EM over their corresponding no-retrieval baselines across all five evaluation settings. Ablations show that lexical and structural evidence are complementary, while removing syntax-aware refinement causes the largest drop in edit similarity. These gains come with 87–95 ms of non-decoding overhead under warm-cache conditions across the three benchmarks, supporting latency-sensitive code completion. To support reproducibility, we make our implementation and evaluation scripts publicly available at https://github.com/longstnguyen/ACAR.
%U https://aclanthology.org/2026.emnlp-main.81/
%P 1856-1874
Markdown (Informal)
[Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval](https://aclanthology.org/2026.emnlp-main.81/) (Nguyen et al., EMNLP 2026)
ACL
- Long S. T. Nguyen, Dat Huynh, Quynh Vo, LE Dang Nguyen Hai, Cong-Duy T Nguyen, Anh Tuan Luu, Anh Nguyen-Duc, and Tho Quan. 2026. Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 1856–1874, Budapest, Hungary. Association for Computational Linguistics.