@inproceedings{ding-etal-2026-sustaining,
title = "Sustaining Cross-Modal Dependency for Generalizable Vision-Language Model Tuning",
author = "Ding, Chenhao and
Han, Jizhou and
Dong, SongLin and
Gao, Xinyuan and
Wang, Qiang and
Xiu, Lixin and
Ke, Wei and
Gong, Yihong and
Nakayama, Hideki",
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.192/",
pages = "4067--4095",
ISBN = "979-8-89176-462-0",
abstract = "Recent parameter-efficient fine-tuning (PEFT) methods effectively adapt Vision-Language Models (VLMs) to seen tasks, yet often exhibit limited generalization to unseen novel classes. To better understand this phenomenon, we analyze adaptation dynamics from the perspective of cross-modal mutual information (MI). Our empirical analyses reveal a consistent pattern across existing methods: during the adaptation process, cross-modal MI for novel classes frequently encounters early plateaus or declines, indicating weakened cross-modal semantic dependency. Motivated by this observation, we propose Retrieval-Augmented Dynamic Adaptation (RADA), a retrieval-enhanced adaptation framework that continuously injects rich instance-specific semantic priors through an LLM-enhanced dual-modality memory bank and a lightweight modality-agnostic token generator. By facilitating fine-grained cross-modal alignment while preserving pretrained zero-shot ability, RADA achieves more transferable adaptation across diverse scenarios. Extensive experiments across multiple widely used benchmarks demonstrate that RADA consistently achieves state-of-the-art performance while maintaining strong zero-shot capability. In addition, our empirical analyses show that RADA preserves more stable cross-modal dependency trajectories throughout adaptation, offering further insights into the relationship between adaptation dynamics and generalization in VLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ding-etal-2026-sustaining">
<titleInfo>
<title>Sustaining Cross-Modal Dependency for Generalizable Vision-Language Model Tuning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chenhao</namePart>
<namePart type="family">Ding</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jizhou</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">SongLin</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinyuan</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qiang</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lixin</namePart>
<namePart type="family">Xiu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Ke</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yihong</namePart>
<namePart type="family">Gong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hideki</namePart>
<namePart type="family">Nakayama</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>Recent parameter-efficient fine-tuning (PEFT) methods effectively adapt Vision-Language Models (VLMs) to seen tasks, yet often exhibit limited generalization to unseen novel classes. To better understand this phenomenon, we analyze adaptation dynamics from the perspective of cross-modal mutual information (MI). Our empirical analyses reveal a consistent pattern across existing methods: during the adaptation process, cross-modal MI for novel classes frequently encounters early plateaus or declines, indicating weakened cross-modal semantic dependency. Motivated by this observation, we propose Retrieval-Augmented Dynamic Adaptation (RADA), a retrieval-enhanced adaptation framework that continuously injects rich instance-specific semantic priors through an LLM-enhanced dual-modality memory bank and a lightweight modality-agnostic token generator. By facilitating fine-grained cross-modal alignment while preserving pretrained zero-shot ability, RADA achieves more transferable adaptation across diverse scenarios. Extensive experiments across multiple widely used benchmarks demonstrate that RADA consistently achieves state-of-the-art performance while maintaining strong zero-shot capability. In addition, our empirical analyses show that RADA preserves more stable cross-modal dependency trajectories throughout adaptation, offering further insights into the relationship between adaptation dynamics and generalization in VLMs.</abstract>
<identifier type="citekey">ding-etal-2026-sustaining</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.192/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>4067</start>
<end>4095</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Sustaining Cross-Modal Dependency for Generalizable Vision-Language Model Tuning
%A Ding, Chenhao
%A Han, Jizhou
%A Dong, SongLin
%A Gao, Xinyuan
%A Wang, Qiang
%A Xiu, Lixin
%A Ke, Wei
%A Gong, Yihong
%A Nakayama, Hideki
%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 ding-etal-2026-sustaining
%X Recent parameter-efficient fine-tuning (PEFT) methods effectively adapt Vision-Language Models (VLMs) to seen tasks, yet often exhibit limited generalization to unseen novel classes. To better understand this phenomenon, we analyze adaptation dynamics from the perspective of cross-modal mutual information (MI). Our empirical analyses reveal a consistent pattern across existing methods: during the adaptation process, cross-modal MI for novel classes frequently encounters early plateaus or declines, indicating weakened cross-modal semantic dependency. Motivated by this observation, we propose Retrieval-Augmented Dynamic Adaptation (RADA), a retrieval-enhanced adaptation framework that continuously injects rich instance-specific semantic priors through an LLM-enhanced dual-modality memory bank and a lightweight modality-agnostic token generator. By facilitating fine-grained cross-modal alignment while preserving pretrained zero-shot ability, RADA achieves more transferable adaptation across diverse scenarios. Extensive experiments across multiple widely used benchmarks demonstrate that RADA consistently achieves state-of-the-art performance while maintaining strong zero-shot capability. In addition, our empirical analyses show that RADA preserves more stable cross-modal dependency trajectories throughout adaptation, offering further insights into the relationship between adaptation dynamics and generalization in VLMs.
%U https://aclanthology.org/2026.findings-emnlp.192/
%P 4067-4095
Markdown (Informal)
[Sustaining Cross-Modal Dependency for Generalizable Vision-Language Model Tuning](https://aclanthology.org/2026.findings-emnlp.192/) (Ding et al., Findings 2026)
ACL
- Chenhao Ding, Jizhou Han, SongLin Dong, Xinyuan Gao, Qiang Wang, Lixin Xiu, Wei Ke, Yihong Gong, and Hideki Nakayama. 2026. Sustaining Cross-Modal Dependency for Generalizable Vision-Language Model Tuning. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 4067–4095, Budapest, Hungary. Association for Computational Linguistics.