@inproceedings{li-etal-2026-factual,
title = "Factual Decoding via Head Coordination Graphs in Large Language Models",
author = "Li, Shike and
Wang, Xiaokai and
Tong, Xin and
Liu, Xiaofeng and
Zhang, Hu",
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.1482/",
pages = "31569--31581",
ISBN = "979-8-89176-462-0",
abstract = "Hallucinations remain challenging for large language models (LLMs), especially in factual generation. Existing mitigation methods often rely on retrieval, supervision, representation editing, or output-level decoding, while paying limited attention to how factual support is internally coordinated across attention heads. In this work, we study factual decoding from a head-level coordination perspective and find that unreliable generations are often associated with fragmented and unstable use of factual anchors across heads. To improve such coordination, we propose Head Coordination Graphs (HCG), a plug-and-play inference-time framework that models attention heads as interacting nodes. HCG uses explicit factual anchors as observable support signals to construct sparse head graphs and applies risk-gated propagation to adaptively regulate cross-head influence without parameter updates, additional training, or external knowledge. Experiments across multiple factuality and hallucination benchmarks demonstrate broad improvements in factual consistency across open-source LLMs, while analyses and ablations demonstrate the importance of head-level coordination, sparse graph construction, and risk-gated propagation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-factual">
<titleInfo>
<title>Factual Decoding via Head Coordination Graphs in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shike</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaokai</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Tong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaofeng</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hu</namePart>
<namePart type="family">Zhang</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>Hallucinations remain challenging for large language models (LLMs), especially in factual generation. Existing mitigation methods often rely on retrieval, supervision, representation editing, or output-level decoding, while paying limited attention to how factual support is internally coordinated across attention heads. In this work, we study factual decoding from a head-level coordination perspective and find that unreliable generations are often associated with fragmented and unstable use of factual anchors across heads. To improve such coordination, we propose Head Coordination Graphs (HCG), a plug-and-play inference-time framework that models attention heads as interacting nodes. HCG uses explicit factual anchors as observable support signals to construct sparse head graphs and applies risk-gated propagation to adaptively regulate cross-head influence without parameter updates, additional training, or external knowledge. Experiments across multiple factuality and hallucination benchmarks demonstrate broad improvements in factual consistency across open-source LLMs, while analyses and ablations demonstrate the importance of head-level coordination, sparse graph construction, and risk-gated propagation.</abstract>
<identifier type="citekey">li-etal-2026-factual</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1482/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>31569</start>
<end>31581</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Factual Decoding via Head Coordination Graphs in Large Language Models
%A Li, Shike
%A Wang, Xiaokai
%A Tong, Xin
%A Liu, Xiaofeng
%A Zhang, Hu
%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 li-etal-2026-factual
%X Hallucinations remain challenging for large language models (LLMs), especially in factual generation. Existing mitigation methods often rely on retrieval, supervision, representation editing, or output-level decoding, while paying limited attention to how factual support is internally coordinated across attention heads. In this work, we study factual decoding from a head-level coordination perspective and find that unreliable generations are often associated with fragmented and unstable use of factual anchors across heads. To improve such coordination, we propose Head Coordination Graphs (HCG), a plug-and-play inference-time framework that models attention heads as interacting nodes. HCG uses explicit factual anchors as observable support signals to construct sparse head graphs and applies risk-gated propagation to adaptively regulate cross-head influence without parameter updates, additional training, or external knowledge. Experiments across multiple factuality and hallucination benchmarks demonstrate broad improvements in factual consistency across open-source LLMs, while analyses and ablations demonstrate the importance of head-level coordination, sparse graph construction, and risk-gated propagation.
%U https://aclanthology.org/2026.findings-emnlp.1482/
%P 31569-31581
Markdown (Informal)
[Factual Decoding via Head Coordination Graphs in Large Language Models](https://aclanthology.org/2026.findings-emnlp.1482/) (Li et al., Findings 2026)
ACL