<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-omnidenoise">
    <titleInfo>
        <title>OmniDenoise: Diffusion MultiModal LLMs for Omni-Modal Understanding and Prediction</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Xianfeng</namePart>
        <namePart type="family">Wu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yuhang</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jie</namePart>
        <namePart type="family">Peng</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jinhao</namePart>
        <namePart type="family">Duan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Tianlong</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Omni-modal LLMs aim to answer queries over heterogeneous inputs through a unified interface, but accepting these inputs does not guarantee using them. % The model’s confidence should change when an informative modality is removed. Existing autoregressive and diffusion multimodal LLMs typically optimize answer likelihood only under the full context, leaving multimodal shortcuts largely unconstrained. We study this problem through the lens of *intervention sensitive omni-modal prediction*, where confidence in the correct answer should decrease when protocol designated evidence is disrupted. We propose *OmniDenoise*, a diffusion MLLM framework built around the *Evidence Denoising Gap* (EDG), a diffusion native objective that compares gold token denoising losses under full and intervened contexts. EDG uses the token-by-timestep denoising trajectory as both a training signal and a diagnostic, without requiring token level evidence annotations. OmniDenoise further applies consistency distillation for efficient few step bidirectional denoising. Across tri-modal, clinical, and out-of-distribution omni-modal benchmarks, OmniDenoise substantially improves over diffusion baselines and remains competitive with strong autoregressive omni models. On OmniBench, the distilled model achieves 45.27% accuracy at 0.19 seconds per sample. On EHRXQA, EDG improves H-Acc from 36.9% to 43.9%, while the distilled student retains 42.6% with a 3.9\times speedup.</abstract>
    <identifier type="citekey">wu-etal-2026-omnidenoise</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.924/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>21273</start>
            <end>21289</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T OmniDenoise: Diffusion MultiModal LLMs for Omni-Modal Understanding and Prediction
%A Wu, Xianfeng
%A Chen, Yuhang
%A Peng, Jie
%A Duan, Jinhao
%A Chen, Tianlong
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F wu-etal-2026-omnidenoise
%X Omni-modal LLMs aim to answer queries over heterogeneous inputs through a unified interface, but accepting these inputs does not guarantee using them. % The model’s confidence should change when an informative modality is removed. Existing autoregressive and diffusion multimodal LLMs typically optimize answer likelihood only under the full context, leaving multimodal shortcuts largely unconstrained. We study this problem through the lens of *intervention sensitive omni-modal prediction*, where confidence in the correct answer should decrease when protocol designated evidence is disrupted. We propose *OmniDenoise*, a diffusion MLLM framework built around the *Evidence Denoising Gap* (EDG), a diffusion native objective that compares gold token denoising losses under full and intervened contexts. EDG uses the token-by-timestep denoising trajectory as both a training signal and a diagnostic, without requiring token level evidence annotations. OmniDenoise further applies consistency distillation for efficient few step bidirectional denoising. Across tri-modal, clinical, and out-of-distribution omni-modal benchmarks, OmniDenoise substantially improves over diffusion baselines and remains competitive with strong autoregressive omni models. On OmniBench, the distilled model achieves 45.27% accuracy at 0.19 seconds per sample. On EHRXQA, EDG improves H-Acc from 36.9% to 43.9%, while the distilled student retains 42.6% with a 3.9\times speedup.
%U https://aclanthology.org/2026.emnlp-main.924/
%P 21273-21289

Markdown (Informal)

[OmniDenoise: Diffusion MultiModal LLMs for Omni-Modal Understanding and Prediction](https://aclanthology.org/2026.emnlp-main.924/) (Wu et al., EMNLP 2026)

• OmniDenoise: Diffusion MultiModal LLMs for Omni-Modal Understanding and Prediction (Wu et al., EMNLP 2026)

ACL

• Xianfeng Wu, Yuhang Chen, Jie Peng, Jinhao Duan, and Tianlong Chen. 2026. OmniDenoise: Diffusion MultiModal LLMs for Omni-Modal Understanding and Prediction. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 21273–21289, Budapest, Hungary. Association for Computational Linguistics.