<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kurpath-etal-2026-omni">
    <titleInfo>
        <title>Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Mohammed</namePart>
        <namePart type="given">Irfan</namePart>
        <namePart type="family">Kurpath</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jaseel</namePart>
        <namePart type="given">Muhammad</namePart>
        <namePart type="family">Kaithakkodan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sahal</namePart>
        <namePart type="given">Shaji</namePart>
        <namePart type="family">Mullappilly</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ivan</namePart>
        <namePart type="family">Laptev</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hisham</namePart>
        <namePart type="family">Cholakkal</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>Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters. We present Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into a single shared cosine space without updating any text-side parameter. Our key insight is that the teacher signal requires no separate model: each media sample is paired with a dense cascaded caption, and the teacher target is simply the frozen backbone’s own embedding of that caption. Because teacher and student share the same backbone weights, they inhabit byte-identical geometry, and lightweight projectors plus phased LoRA adapters on the modality encoders suffice for alignment. Training combines a Matryoshka SigLIP contrastive loss with an online hybrid hard-negative miner whose negatives sharpen as the encoder improves. The recipe scales unchanged to a 2.3B variant by swapping in a native vision-language backbone. Omni-Embed-Mini-0.9B keeps its text weights bit-identical to the backbone, so training cannot regress text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8), while extending it to five additional modalities, and is ~2.7× to 9.5× smaller than every open omni embedder we compare against. The 2.3B variant is competitive with closed API embedders, edging ahead of gemini-embedding-2 on the overall-modality average. Models, code, dataset, and evaluation harness are available on our project page: https://omniembed.cvmbzuai.com</abstract>
    <identifier type="citekey">kurpath-etal-2026-omni</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.2470/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>53150</start>
            <end>53175</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation
%A Kurpath, Mohammed Irfan
%A Kaithakkodan, Jaseel Muhammad
%A Mullappilly, Sahal Shaji
%A Laptev, Ivan
%A Cholakkal, Hisham
%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 kurpath-etal-2026-omni
%X Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters. We present Omni-Embed-Mini, a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into a single shared cosine space without updating any text-side parameter. Our key insight is that the teacher signal requires no separate model: each media sample is paired with a dense cascaded caption, and the teacher target is simply the frozen backbone’s own embedding of that caption. Because teacher and student share the same backbone weights, they inhabit byte-identical geometry, and lightweight projectors plus phased LoRA adapters on the modality encoders suffice for alignment. Training combines a Matryoshka SigLIP contrastive loss with an online hybrid hard-negative miner whose negatives sharpen as the encoder improves. The recipe scales unchanged to a 2.3B variant by swapping in a native vision-language backbone. Omni-Embed-Mini-0.9B keeps its text weights bit-identical to the backbone, so training cannot regress text retrieval (49.57 nDCG@10 on MTEB-v2 BEIR-8), while extending it to five additional modalities, and is ~2.7× to 9.5× smaller than every open omni embedder we compare against. The 2.3B variant is competitive with closed API embedders, edging ahead of gemini-embedding-2 on the overall-modality average. Models, code, dataset, and evaluation harness are available on our project page: https://omniembed.cvmbzuai.com
%U https://aclanthology.org/2026.findings-emnlp.2470/
%P 53150-53175

Markdown (Informal)

[Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation](https://aclanthology.org/2026.findings-emnlp.2470/) (Kurpath et al., Findings 2026)

• Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation (Kurpath et al., Findings 2026)

ACL

• Mohammed Irfan Kurpath, Jaseel Muhammad Kaithakkodan, Sahal Shaji Mullappilly, Ivan Laptev, and Hisham Cholakkal. 2026. Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 53150–53175, Budapest, Hungary. Association for Computational Linguistics.