<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhang-etal-2026-shallow">
    <titleInfo>
        <title>Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shuo</namePart>
        <namePart type="family">Zhang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jintao</namePart>
        <namePart type="family">Tong</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yixiong</namePart>
        <namePart type="family">Zou</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yuhua</namePart>
        <namePart type="family">Li</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ruixuan</namePart>
        <namePart type="family">Li</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>Large Vision-Language Models (LVLMs) incur high computational costs from redundant visual tokens. Although training-free attention-based multi-layer pruning in the vision encoder has been explored as an effective strategy, we find that pruning in shallow layers consistently degrades performance. We aim to understand this problem and seek a solution. By analyzing attention patterns across network depth, we find that shallow layers primarily function as edge detectors with chaotic attention maps, while deeper layers transition through local subject recognition and unstable semantic aggregation. To address misalignment between pruning strategies and network stages, we propose STD, a hierarchical token pruning framework that adapts token selection mechanisms to the functional role of each stage. STD employs High-Frequency Spectral Analysis in shallow layers to deterministically preserve structural edges, uses Gaussian-Smoothed Attention in intermediate layers to maintain spatial coherence, and introduces a Stability-Adaptive Trigger in deep layers to execute pruning only during semantically stable phases. Extensive experiments show that STD outperforms state-of-the-art pruning methods by 1.1% on LLaVA-1.5-7B with 88.9% token reduction, while being plug-and-play and highly effective when combined with other methods, and by 2.1% on LLaVA-NeXT-7B with 94.4% reduction, delivering a 3.9× speed-up in the prefilling stage. Our code will be released at https://github.com/Twilight03/STD.</abstract>
    <identifier type="citekey">zhang-etal-2026-shallow</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.1226/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>27977</start>
            <end>27993</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs
%A Zhang, Shuo
%A Tong, Jintao
%A Zou, Yixiong
%A Li, Yuhua
%A Li, Ruixuan
%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 zhang-etal-2026-shallow
%X Large Vision-Language Models (LVLMs) incur high computational costs from redundant visual tokens. Although training-free attention-based multi-layer pruning in the vision encoder has been explored as an effective strategy, we find that pruning in shallow layers consistently degrades performance. We aim to understand this problem and seek a solution. By analyzing attention patterns across network depth, we find that shallow layers primarily function as edge detectors with chaotic attention maps, while deeper layers transition through local subject recognition and unstable semantic aggregation. To address misalignment between pruning strategies and network stages, we propose STD, a hierarchical token pruning framework that adapts token selection mechanisms to the functional role of each stage. STD employs High-Frequency Spectral Analysis in shallow layers to deterministically preserve structural edges, uses Gaussian-Smoothed Attention in intermediate layers to maintain spatial coherence, and introduces a Stability-Adaptive Trigger in deep layers to execute pruning only during semantically stable phases. Extensive experiments show that STD outperforms state-of-the-art pruning methods by 1.1% on LLaVA-1.5-7B with 88.9% token reduction, while being plug-and-play and highly effective when combined with other methods, and by 2.1% on LLaVA-NeXT-7B with 94.4% reduction, delivering a 3.9× speed-up in the prefilling stage. Our code will be released at https://github.com/Twilight03/STD.
%U https://aclanthology.org/2026.emnlp-main.1226/
%P 27977-27993

Markdown (Informal)

[Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs](https://aclanthology.org/2026.emnlp-main.1226/) (Zhang et al., EMNLP 2026)

• Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs (Zhang et al., EMNLP 2026)

ACL

• Shuo Zhang, Jintao Tong, Yixiong Zou, Yuhua Li, and Ruixuan Li. 2026. Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 27977–27993, Budapest, Hungary. Association for Computational Linguistics.