@inproceedings{zhang-etal-2026-distilling-multimodal,
title = "Distilling Multimodal Evolution with Adaptive Anchors for Cloud-Edge Collaborative {MLLM}s",
author = "Zhang, Chunyun and
Song, Qilong and
Cui, Chaoran and
Gong, Shuai and
Yang, Fujun and
Wang, Wenna and
Lin, Peiguang",
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.886/",
pages = "18732--18743",
ISBN = "979-8-89176-462-0",
abstract = "Cloud-edge collaboration provides a practical paradigm for deploying MLLMs on resource-constrained devices, where compact edge-side models are improved with supervision from stronger cloud-side models. However, existing cloud-edge distillation methods mainly rely on pseudo labels or static feature alignment, overlooking how the teacher progressively transforms multimodal inputs into answers. In this paper, we propose A$^2$-MED, Adaptive Anchor-Guided Multimodal Evolution Distillation for cloud-edge MLLMs. Our key idea is to organize the teacher{'}s layer-wise multimodal evolution into functional stages through adaptive teacher-side anchoring, and transfer the resulting stage-wise evolution as process-level supervision. Specifically, we adaptively identify functional anchors from the teacher according to layer-wise indicators of cross-modal binding, semantic integration, and answer prediction, and then project these anchors to the student to establish corresponding functional stages. Under these teacher-defined stages, we align teacher-student anchor states and distill inter-stage multimodal evolution relations, enabling the student to learn how the teacher progressively transforms multimodal inputs into answers. Experiments on A-OKVQA and VizWiz show that our method consistently improves over representative cloud-edge and distillation baselines, with further analyses validating the effectiveness of adaptive anchors and evolution distillation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-distilling-multimodal">
<titleInfo>
<title>Distilling Multimodal Evolution with Adaptive Anchors for Cloud-Edge Collaborative MLLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chunyun</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qilong</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chaoran</namePart>
<namePart type="family">Cui</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuai</namePart>
<namePart type="family">Gong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fujun</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenna</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peiguang</namePart>
<namePart type="family">Lin</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>Cloud-edge collaboration provides a practical paradigm for deploying MLLMs on resource-constrained devices, where compact edge-side models are improved with supervision from stronger cloud-side models. However, existing cloud-edge distillation methods mainly rely on pseudo labels or static feature alignment, overlooking how the teacher progressively transforms multimodal inputs into answers. In this paper, we propose A²-MED, Adaptive Anchor-Guided Multimodal Evolution Distillation for cloud-edge MLLMs. Our key idea is to organize the teacher’s layer-wise multimodal evolution into functional stages through adaptive teacher-side anchoring, and transfer the resulting stage-wise evolution as process-level supervision. Specifically, we adaptively identify functional anchors from the teacher according to layer-wise indicators of cross-modal binding, semantic integration, and answer prediction, and then project these anchors to the student to establish corresponding functional stages. Under these teacher-defined stages, we align teacher-student anchor states and distill inter-stage multimodal evolution relations, enabling the student to learn how the teacher progressively transforms multimodal inputs into answers. Experiments on A-OKVQA and VizWiz show that our method consistently improves over representative cloud-edge and distillation baselines, with further analyses validating the effectiveness of adaptive anchors and evolution distillation.</abstract>
<identifier type="citekey">zhang-etal-2026-distilling-multimodal</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.886/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>18732</start>
<end>18743</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Distilling Multimodal Evolution with Adaptive Anchors for Cloud-Edge Collaborative MLLMs
%A Zhang, Chunyun
%A Song, Qilong
%A Cui, Chaoran
%A Gong, Shuai
%A Yang, Fujun
%A Wang, Wenna
%A Lin, Peiguang
%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 zhang-etal-2026-distilling-multimodal
%X Cloud-edge collaboration provides a practical paradigm for deploying MLLMs on resource-constrained devices, where compact edge-side models are improved with supervision from stronger cloud-side models. However, existing cloud-edge distillation methods mainly rely on pseudo labels or static feature alignment, overlooking how the teacher progressively transforms multimodal inputs into answers. In this paper, we propose A²-MED, Adaptive Anchor-Guided Multimodal Evolution Distillation for cloud-edge MLLMs. Our key idea is to organize the teacher’s layer-wise multimodal evolution into functional stages through adaptive teacher-side anchoring, and transfer the resulting stage-wise evolution as process-level supervision. Specifically, we adaptively identify functional anchors from the teacher according to layer-wise indicators of cross-modal binding, semantic integration, and answer prediction, and then project these anchors to the student to establish corresponding functional stages. Under these teacher-defined stages, we align teacher-student anchor states and distill inter-stage multimodal evolution relations, enabling the student to learn how the teacher progressively transforms multimodal inputs into answers. Experiments on A-OKVQA and VizWiz show that our method consistently improves over representative cloud-edge and distillation baselines, with further analyses validating the effectiveness of adaptive anchors and evolution distillation.
%U https://aclanthology.org/2026.findings-emnlp.886/
%P 18732-18743
Markdown (Informal)
[Distilling Multimodal Evolution with Adaptive Anchors for Cloud-Edge Collaborative MLLMs](https://aclanthology.org/2026.findings-emnlp.886/) (Zhang et al., Findings 2026)
ACL