@inproceedings{zou-etal-2026-saliency,
title = "From Saliency Ranking to Evidence Preservation: Visual Token Condensation for Efficient Multimodal Reasoning",
author = "Zou, Xin and
Lu, Di and
Yan, Yibo and
Liu, Shuliang and
Wang, Yizhou and
Liu, Xiang and
Zhang, Linfeng and
Hu, Xuming",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.771/",
pages = "17908--17935",
ISBN = "979-8-89176-459-0",
abstract = "Visual tokens dominate the inference cost of multimodal large language models (MLLMs), but aggressive token reduction often degrades reasoning. We argue that the fragility comes from a saliency-ranking view, i.e., existing training-free methods typically rank tokens by a scalar proxy, such as cross-attention, [CLS] attention, or feature similarity, and then prune or merge tokens independently. However, visual evidence is set-valued and structured: salient patches can be redundant, low-saliency tokens can be answer-critical, and feature-similar patches can be unsafe to merge when they are spatially incompatible. We propose CondenseVLM, a simple framework that formulates visual token compression as a budgeted evidence-preservation question. The formulation jointly optimizes salient evidence coverage, contextual coverage, and spatially constrained merge fidelity. It yields a monotone submodular coverage term, a marginal-gain interpretation of diversity-aware anchor selection, and a local distortion bound for spatial-semantic merging. Comprehensive experiments across six MLLM families and diverse general-purpose, grounding, text-dense, and video benchmarks demonstrate that improves the accuracy-efficiency trade-off, and achieves the SOTA performance."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zou-etal-2026-saliency">
<titleInfo>
<title>From Saliency Ranking to Evidence Preservation: Visual Token Condensation for Efficient Multimodal Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xin</namePart>
<namePart type="family">Zou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Di</namePart>
<namePart type="family">Lu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yibo</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuliang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yizhou</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Linfeng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuming</namePart>
<namePart type="family">Hu</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>Visual tokens dominate the inference cost of multimodal large language models (MLLMs), but aggressive token reduction often degrades reasoning. We argue that the fragility comes from a saliency-ranking view, i.e., existing training-free methods typically rank tokens by a scalar proxy, such as cross-attention, [CLS] attention, or feature similarity, and then prune or merge tokens independently. However, visual evidence is set-valued and structured: salient patches can be redundant, low-saliency tokens can be answer-critical, and feature-similar patches can be unsafe to merge when they are spatially incompatible. We propose CondenseVLM, a simple framework that formulates visual token compression as a budgeted evidence-preservation question. The formulation jointly optimizes salient evidence coverage, contextual coverage, and spatially constrained merge fidelity. It yields a monotone submodular coverage term, a marginal-gain interpretation of diversity-aware anchor selection, and a local distortion bound for spatial-semantic merging. Comprehensive experiments across six MLLM families and diverse general-purpose, grounding, text-dense, and video benchmarks demonstrate that improves the accuracy-efficiency trade-off, and achieves the SOTA performance.</abstract>
<identifier type="citekey">zou-etal-2026-saliency</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.771/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17908</start>
<end>17935</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Saliency Ranking to Evidence Preservation: Visual Token Condensation for Efficient Multimodal Reasoning
%A Zou, Xin
%A Lu, Di
%A Yan, Yibo
%A Liu, Shuliang
%A Wang, Yizhou
%A Liu, Xiang
%A Zhang, Linfeng
%A Hu, Xuming
%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 zou-etal-2026-saliency
%X Visual tokens dominate the inference cost of multimodal large language models (MLLMs), but aggressive token reduction often degrades reasoning. We argue that the fragility comes from a saliency-ranking view, i.e., existing training-free methods typically rank tokens by a scalar proxy, such as cross-attention, [CLS] attention, or feature similarity, and then prune or merge tokens independently. However, visual evidence is set-valued and structured: salient patches can be redundant, low-saliency tokens can be answer-critical, and feature-similar patches can be unsafe to merge when they are spatially incompatible. We propose CondenseVLM, a simple framework that formulates visual token compression as a budgeted evidence-preservation question. The formulation jointly optimizes salient evidence coverage, contextual coverage, and spatially constrained merge fidelity. It yields a monotone submodular coverage term, a marginal-gain interpretation of diversity-aware anchor selection, and a local distortion bound for spatial-semantic merging. Comprehensive experiments across six MLLM families and diverse general-purpose, grounding, text-dense, and video benchmarks demonstrate that improves the accuracy-efficiency trade-off, and achieves the SOTA performance.
%U https://aclanthology.org/2026.emnlp-main.771/
%P 17908-17935
Markdown (Informal)
[From Saliency Ranking to Evidence Preservation: Visual Token Condensation for Efficient Multimodal Reasoning](https://aclanthology.org/2026.emnlp-main.771/) (Zou et al., EMNLP 2026)
ACL
- Xin Zou, Di Lu, Yibo Yan, Shuliang Liu, Yizhou Wang, Xiang Liu, Linfeng Zhang, and Xuming Hu. 2026. From Saliency Ranking to Evidence Preservation: Visual Token Condensation for Efficient Multimodal Reasoning. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 17908–17935, Budapest, Hungary. Association for Computational Linguistics.