@inproceedings{zhou-etal-2026-routing,
title = "Routing-Aware Head Amplification for Multimodal Long-Chain-of-Thought Reasoning",
author = "Zhou, Chuhao and
Wu, Zonghan and
Lu, Haolang and
Wang, Kun and
Yao, Weiran and
Pang, Linsey and
Mehrotra, Prakhar and
Wen, Qingsong and
Shao, Yilei",
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.2405/",
pages = "51665--51676",
ISBN = "979-8-89176-462-0",
abstract = "Multimodal long-chain-of-thought (CoT) generation requires models to dynamically route information between visual evidence and an expanding reasoning history. Existing hallucination mitigation methods often strengthen visual grounding or intervene in decoding, but typically treat grounding as a relatively static property of generation. We propose a training-free dynamic attention intervention framework that exploits the head-level sparsity, layer-level stratification, and trajectory-level dynamics of multimodal long-CoT decoding. Our method selects routing-sensitive heads using perturbation-only calibration, assigns visual-evidence and reasoning-history interventions to different layer regions, and continuously rescales selected head outputs according to their instantaneous token-group attention ratios. Across multiple MLLM architectures and multimodal hallucination benchmarks, our method consistently improves over vanilla decoding and visual contrastive decoding, achieving an average performance gain of 4.6{\%}."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhou-etal-2026-routing">
<titleInfo>
<title>Routing-Aware Head Amplification for Multimodal Long-Chain-of-Thought Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chuhao</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zonghan</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haolang</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kun</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weiran</namePart>
<namePart type="family">Yao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linsey</namePart>
<namePart type="family">Pang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Prakhar</namePart>
<namePart type="family">Mehrotra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qingsong</namePart>
<namePart type="family">Wen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yilei</namePart>
<namePart type="family">Shao</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>Multimodal long-chain-of-thought (CoT) generation requires models to dynamically route information between visual evidence and an expanding reasoning history. Existing hallucination mitigation methods often strengthen visual grounding or intervene in decoding, but typically treat grounding as a relatively static property of generation. We propose a training-free dynamic attention intervention framework that exploits the head-level sparsity, layer-level stratification, and trajectory-level dynamics of multimodal long-CoT decoding. Our method selects routing-sensitive heads using perturbation-only calibration, assigns visual-evidence and reasoning-history interventions to different layer regions, and continuously rescales selected head outputs according to their instantaneous token-group attention ratios. Across multiple MLLM architectures and multimodal hallucination benchmarks, our method consistently improves over vanilla decoding and visual contrastive decoding, achieving an average performance gain of 4.6%.</abstract>
<identifier type="citekey">zhou-etal-2026-routing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2405/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51665</start>
<end>51676</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Routing-Aware Head Amplification for Multimodal Long-Chain-of-Thought Reasoning
%A Zhou, Chuhao
%A Wu, Zonghan
%A Lu, Haolang
%A Wang, Kun
%A Yao, Weiran
%A Pang, Linsey
%A Mehrotra, Prakhar
%A Wen, Qingsong
%A Shao, Yilei
%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 zhou-etal-2026-routing
%X Multimodal long-chain-of-thought (CoT) generation requires models to dynamically route information between visual evidence and an expanding reasoning history. Existing hallucination mitigation methods often strengthen visual grounding or intervene in decoding, but typically treat grounding as a relatively static property of generation. We propose a training-free dynamic attention intervention framework that exploits the head-level sparsity, layer-level stratification, and trajectory-level dynamics of multimodal long-CoT decoding. Our method selects routing-sensitive heads using perturbation-only calibration, assigns visual-evidence and reasoning-history interventions to different layer regions, and continuously rescales selected head outputs according to their instantaneous token-group attention ratios. Across multiple MLLM architectures and multimodal hallucination benchmarks, our method consistently improves over vanilla decoding and visual contrastive decoding, achieving an average performance gain of 4.6%.
%U https://aclanthology.org/2026.findings-emnlp.2405/
%P 51665-51676
Markdown (Informal)
[Routing-Aware Head Amplification for Multimodal Long-Chain-of-Thought Reasoning](https://aclanthology.org/2026.findings-emnlp.2405/) (Zhou et al., Findings 2026)
ACL
- Chuhao Zhou, Zonghan Wu, Haolang Lu, Kun Wang, Weiran Yao, Linsey Pang, Prakhar Mehrotra, Qingsong Wen, and Yilei Shao. 2026. Routing-Aware Head Amplification for Multimodal Long-Chain-of-Thought Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 51665–51676, Budapest, Hungary. Association for Computational Linguistics.