<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fan-etal-2026-refocus">
    <titleInfo>
        <title>REFOCUS: Training-Free Attention Sparsification via Uniform-Reference Filtering</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Fengyuan</namePart>
        <namePart type="family">Fan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Rena</namePart>
        <namePart type="given">Wei</namePart>
        <namePart type="family">Gao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yanbei</namePart>
        <namePart type="family">Jiang</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>Dense softmax attention assigns non-zero probability mass to many task-irrelevant keys, which can dilute useful signals in the attention output. We present REFOCUS, a training-free and plug-and-play inference-time sparsification method that keeps attention entries above a row-wise, context-length-aware threshold and renormalizes the remaining weights. This design is stable, simple to deploy, and requires no model retraining. REFOCUS is integrated in HuggingFace via attn_implementation=”refocus” without model-code modification. On Qwen3-4B under 0-shot greedy decoding, REFOCUS matches or outperforms dense eager attention on full-test GSM8K, HumanEval, WinoGrande, and AQuA-RAT, and improves RULER-4k by +0.52 percentage points overall, including +3.20 points on aggregation tasks. These results indicate that a single row-wise threshold can remove diffuse long-tail attention while preserving or improving accuracy.</abstract>
    <identifier type="citekey">fan-etal-2026-refocus</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2668/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>60717</start>
            <end>60730</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T REFOCUS: Training-Free Attention Sparsification via Uniform-Reference Filtering
%A Fan, Fengyuan
%A Gao, Rena Wei
%A Jiang, Yanbei
%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 fan-etal-2026-refocus
%X Dense softmax attention assigns non-zero probability mass to many task-irrelevant keys, which can dilute useful signals in the attention output. We present REFOCUS, a training-free and plug-and-play inference-time sparsification method that keeps attention entries above a row-wise, context-length-aware threshold and renormalizes the remaining weights. This design is stable, simple to deploy, and requires no model retraining. REFOCUS is integrated in HuggingFace via attn_implementation=”refocus” without model-code modification. On Qwen3-4B under 0-shot greedy decoding, REFOCUS matches or outperforms dense eager attention on full-test GSM8K, HumanEval, WinoGrande, and AQuA-RAT, and improves RULER-4k by +0.52 percentage points overall, including +3.20 points on aggregation tasks. These results indicate that a single row-wise threshold can remove diffuse long-tail attention while preserving or improving accuracy.
%U https://aclanthology.org/2026.emnlp-main.2668/
%P 60717-60730

Markdown (Informal)

[REFOCUS: Training-Free Attention Sparsification via Uniform-Reference Filtering](https://aclanthology.org/2026.emnlp-main.2668/) (Fan et al., EMNLP 2026)

• REFOCUS: Training-Free Attention Sparsification via Uniform-Reference Filtering (Fan et al., EMNLP 2026)

ACL

• Fengyuan Fan, Rena Wei Gao, and Yanbei Jiang. 2026. REFOCUS: Training-Free Attention Sparsification via Uniform-Reference Filtering. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 60717–60730, Budapest, Hungary. Association for Computational Linguistics.