@inproceedings{nguyen-etal-2026-beyond-caption,
title = "Beyond Caption Loss: Generative Distillation for Modality Alignment in Vision-Language Models",
author = "Nguyen, Nam Kha and
Zhang, Wei Emma and
Dai, Xiang and
Sim, Mong Yuan and
Chen, Weitong",
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.1876/",
pages = "40021--40034",
ISBN = "979-8-89176-462-0",
abstract = "Vision-language models align frozen vision encoders to language models using only a next-token prediction loss, an asymmetric signal that discards rich visual signals not captured in the captions. We propose GenDA (Generative Distillation Alignment), a novel training framework that augments the captioning loss with a generative distillation objective from a frozen text-to-image diffusion model. GenDA concatenates selected visual tokens with the encoded caption to form a joint conditioning signal, encouraging these tokens to encode visual information not already captured by the caption. Across five vision encoders and four language models, GenDA gains 3-9 CIDEr points over LLaVA-1.5 on five captioning benchmarks. These gains also transfer to downstream VLM benchmarks while preserving the encoders' discriminative performance. Visual preservation and modality-gap analyses suggest that visual information is retained deeper within the LLM, while visual features become more closely aligned with the language representation space."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nguyen-etal-2026-beyond-caption">
<titleInfo>
<title>Beyond Caption Loss: Generative Distillation for Modality Alignment in Vision-Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nam</namePart>
<namePart type="given">Kha</namePart>
<namePart type="family">Nguyen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="given">Emma</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiang</namePart>
<namePart type="family">Dai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mong</namePart>
<namePart type="given">Yuan</namePart>
<namePart type="family">Sim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Weitong</namePart>
<namePart type="family">Chen</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>Vision-language models align frozen vision encoders to language models using only a next-token prediction loss, an asymmetric signal that discards rich visual signals not captured in the captions. We propose GenDA (Generative Distillation Alignment), a novel training framework that augments the captioning loss with a generative distillation objective from a frozen text-to-image diffusion model. GenDA concatenates selected visual tokens with the encoded caption to form a joint conditioning signal, encouraging these tokens to encode visual information not already captured by the caption. Across five vision encoders and four language models, GenDA gains 3-9 CIDEr points over LLaVA-1.5 on five captioning benchmarks. These gains also transfer to downstream VLM benchmarks while preserving the encoders’ discriminative performance. Visual preservation and modality-gap analyses suggest that visual information is retained deeper within the LLM, while visual features become more closely aligned with the language representation space.</abstract>
<identifier type="citekey">nguyen-etal-2026-beyond-caption</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1876/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>40021</start>
<end>40034</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Beyond Caption Loss: Generative Distillation for Modality Alignment in Vision-Language Models
%A Nguyen, Nam Kha
%A Zhang, Wei Emma
%A Dai, Xiang
%A Sim, Mong Yuan
%A Chen, Weitong
%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 nguyen-etal-2026-beyond-caption
%X Vision-language models align frozen vision encoders to language models using only a next-token prediction loss, an asymmetric signal that discards rich visual signals not captured in the captions. We propose GenDA (Generative Distillation Alignment), a novel training framework that augments the captioning loss with a generative distillation objective from a frozen text-to-image diffusion model. GenDA concatenates selected visual tokens with the encoded caption to form a joint conditioning signal, encouraging these tokens to encode visual information not already captured by the caption. Across five vision encoders and four language models, GenDA gains 3-9 CIDEr points over LLaVA-1.5 on five captioning benchmarks. These gains also transfer to downstream VLM benchmarks while preserving the encoders’ discriminative performance. Visual preservation and modality-gap analyses suggest that visual information is retained deeper within the LLM, while visual features become more closely aligned with the language representation space.
%U https://aclanthology.org/2026.findings-emnlp.1876/
%P 40021-40034
Markdown (Informal)
[Beyond Caption Loss: Generative Distillation for Modality Alignment in Vision-Language Models](https://aclanthology.org/2026.findings-emnlp.1876/) (Nguyen et al., Findings 2026)
ACL