<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-satome">
    <titleInfo>
        <title>SaToMe: Salience-Driven Token Merging for Omni-Capable Diffusion MLLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Sixun</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Pengfei</namePart>
        <namePart type="family">Guo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jinghao</namePart>
        <namePart type="family">Zhang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Tingyu</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Xiaofeng</namePart>
        <namePart type="family">Zhang</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>Omni-capable Diffusion multimodal large language models (Diffusion MLLMs) unify text-to-image (T2I) generation and image-to-text (I2T) understanding within one backbone, but at substantial cost on both fronts: visual tokens far outnumber surrounding text, and for T2I this per-step cost compounds across many denoising steps. Token merging has been studied separately in adjacent fields, yet training-free acceleration spanning T2I+I2T in one omni-capable Diffusion MLLM remains limited. We present SaToMe, a training-free token-merging method built around a local-deviation operator that we interpret as measuring local representational uniqueness. The denoising trajectory exposes two natural inputs to this operator: the state h and the update direction \ensuremathΔh. Their spreads behave very differently along the trajectory: state-salience spread is dynamic and identifies the active merging window, while update-salience masks are window-stable and serve as a single-step protection signal for cell-level structural content. T2I uses both signals; I2T, with no temporal axis, instantiates the same uniqueness criterion on the state alone, producing a single-pass additive merge. The same I2T form transfers directly to a separate I2T-only Diffusion MLLM with a different visual representation, supporting its applicability across these visual representations. SaToMe attains +20.14% T2I wall-clock speedup on LLaDA-o at 68% visual-token retention while largely preserving reported generation quality, +35.5% I2T wall-clock speedup on LLaDA-o, and +27.5% I2T speedup on LaViDa.</abstract>
    <identifier type="citekey">liu-etal-2026-satome</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1954/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>41656</start>
            <end>41672</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T SaToMe: Salience-Driven Token Merging for Omni-Capable Diffusion MLLMs
%A Liu, Sixun
%A Guo, Pengfei
%A Zhang, Jinghao
%A Chen, Tingyu
%A Zhang, Xiaofeng
%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 liu-etal-2026-satome
%X Omni-capable Diffusion multimodal large language models (Diffusion MLLMs) unify text-to-image (T2I) generation and image-to-text (I2T) understanding within one backbone, but at substantial cost on both fronts: visual tokens far outnumber surrounding text, and for T2I this per-step cost compounds across many denoising steps. Token merging has been studied separately in adjacent fields, yet training-free acceleration spanning T2I+I2T in one omni-capable Diffusion MLLM remains limited. We present SaToMe, a training-free token-merging method built around a local-deviation operator that we interpret as measuring local representational uniqueness. The denoising trajectory exposes two natural inputs to this operator: the state h and the update direction \ensuremathΔh. Their spreads behave very differently along the trajectory: state-salience spread is dynamic and identifies the active merging window, while update-salience masks are window-stable and serve as a single-step protection signal for cell-level structural content. T2I uses both signals; I2T, with no temporal axis, instantiates the same uniqueness criterion on the state alone, producing a single-pass additive merge. The same I2T form transfers directly to a separate I2T-only Diffusion MLLM with a different visual representation, supporting its applicability across these visual representations. SaToMe attains +20.14% T2I wall-clock speedup on LLaDA-o at 68% visual-token retention while largely preserving reported generation quality, +35.5% I2T wall-clock speedup on LLaDA-o, and +27.5% I2T speedup on LaViDa.
%U https://aclanthology.org/2026.findings-emnlp.1954/
%P 41656-41672

Markdown (Informal)

[SaToMe: Salience-Driven Token Merging for Omni-Capable Diffusion MLLMs](https://aclanthology.org/2026.findings-emnlp.1954/) (Liu et al., Findings 2026)

• SaToMe: Salience-Driven Token Merging for Omni-Capable Diffusion MLLMs (Liu et al., Findings 2026)

ACL

• Sixun Liu, Pengfei Guo, Jinghao Zhang, Tingyu Chen, and Xiaofeng Zhang. 2026. SaToMe: Salience-Driven Token Merging for Omni-Capable Diffusion MLLMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 41656–41672, Budapest, Hungary. Association for Computational Linguistics.