<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2026-blockbatch">
    <titleInfo>
        <title>BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Xiaoyou</namePart>
        <namePart type="family">Wu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Cheng-Jhih</namePart>
        <namePart type="family">Shih</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Binfei</namePart>
        <namePart type="family">Ji</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yong</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yingyan</namePart>
        <namePart type="given">Celine</namePart>
        <namePart type="family">Lin</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>Diffusion language models (dLLMs) generate text by iteratively denoising multiple token positions in parallel, offering an attractive alternative to strictly autoregressive decoding. In practice, however, block-wise dLLM inference exposes a difficult granularity trade-off: small blocks preserve local conditioning but require many denoising steps, whereas large blocks expose more parallelism but can make premature commitments. Existing block-wise acceleration methods either use a fixed block size or adapt block boundaries along a single decoding trajectory, leaving the complementarity among concurrently executed block-size trajectories unused. We show that block-size diversity provides a useful branching dimension. Different block sizes induce related but non-identical KV-cache trajectories: branches often share an initial prefix, bifurcate at semantically decisive positions, and later agree on syntactically lightweight tokens. Motivated by this structure, we propose BlockBatch, a training-free online inference framework that executes multiple block-size branches for the same request inside a batched forward pass. BlockBatch coordinates these branches through confidence-gated token merging, leader-based synchronization, and periodic full-sequence refreshes that re-anchor local block updates to a globally consistent KV state. Across three representative dLLMs and four datasets, BlockBatch reduces denoising NFEs by 26.6% on average and achieves a 1.33\times average end-to-end speedup, with comparable average accuracy.</abstract>
    <identifier type="citekey">wu-etal-2026-blockbatch</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1492/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>31747</start>
            <end>31774</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference
%A Wu, Xiaoyou
%A Shih, Cheng-Jhih
%A Ji, Binfei
%A Liu, Yong
%A Lin, Yingyan Celine
%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 wu-etal-2026-blockbatch
%X Diffusion language models (dLLMs) generate text by iteratively denoising multiple token positions in parallel, offering an attractive alternative to strictly autoregressive decoding. In practice, however, block-wise dLLM inference exposes a difficult granularity trade-off: small blocks preserve local conditioning but require many denoising steps, whereas large blocks expose more parallelism but can make premature commitments. Existing block-wise acceleration methods either use a fixed block size or adapt block boundaries along a single decoding trajectory, leaving the complementarity among concurrently executed block-size trajectories unused. We show that block-size diversity provides a useful branching dimension. Different block sizes induce related but non-identical KV-cache trajectories: branches often share an initial prefix, bifurcate at semantically decisive positions, and later agree on syntactically lightweight tokens. Motivated by this structure, we propose BlockBatch, a training-free online inference framework that executes multiple block-size branches for the same request inside a batched forward pass. BlockBatch coordinates these branches through confidence-gated token merging, leader-based synchronization, and periodic full-sequence refreshes that re-anchor local block updates to a globally consistent KV state. Across three representative dLLMs and four datasets, BlockBatch reduces denoising NFEs by 26.6% on average and achieves a 1.33\times average end-to-end speedup, with comparable average accuracy.
%U https://aclanthology.org/2026.findings-emnlp.1492/
%P 31747-31774

Markdown (Informal)

[BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference](https://aclanthology.org/2026.findings-emnlp.1492/) (Wu et al., Findings 2026)

• BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference (Wu et al., Findings 2026)

ACL

• Xiaoyou Wu, Cheng-Jhih Shih, Binfei Ji, Yong Liu, and Yingyan Celine Lin. 2026. BlockBatch: Multi-Scale Consensus Decoding for Efficient Diffusion Language Model Inference. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 31747–31774, Budapest, Hungary. Association for Computational Linguistics.