@inproceedings{li-etal-2026-sieveserve,
title = "{S}ieve{S}erve: Adaptive Visual Token Pruning for Reliable and Efficient Multimodal {LLM} Inference",
author = "Li, Xiao and
Shi, Xiao and
Zhang, Cheng and
Wang, Sa and
Zhang, Zhining and
Zhao, Xiaofang",
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.648/",
pages = "14918--14934",
ISBN = "979-8-89176-459-0",
abstract = "Efficient serving of multimodal large language models (MLLMs) is increasingly bottlenecked by high-resolution visual inputs, which can generate thousands of visual tokens per request and severely increase latency, memory pressure, and serving cost. Visual token pruning is a natural solution, but existing methods are difficult to deploy in production MLLM serving: many operate inside LLM layers and conflict with engine invariants such as KV-cache granularity, fused kernels, prefix caching, and batching, while fixed-ratio pruning can collapse accuracy on dense-information tasks such as document and chart understanding.\par We present SieveServe, a serving-compatible framework for adaptive visual token pruning. SieveServe establishes pre-LLM vision-centric pruning as an engine-compatible paradigm and introduces a request-aware dual-guard policy: a Query Complexity Router (QCR) estimates query-side visual evidence demand, and a Dense Information Guard (DIG) detects image-side pruning risk using lightweight channel-wise embedding variance. Together, they allocate visual-token budgets per request, aggressively compressing pruning-robust inputs while conservatively preserving dense/high-risk ones.\par Experiments with Qwen2.5-VL-7B and InternVL3-8B across eight benchmarks show that SieveServe retains at least 95{\%} of baseline accuracy on all benchmarks and at least 97.8{\%} on dense-information tasks, while compressing pruning-robust tasks by 3.2x and recovering over 20 accuracy points lost by fixed-ratio pruning. On high-resolution serving workloads, SieveServe reduces P90 TTFT by 5.4x and sustains 2.2-2.7x higher throughput under latency SLO constraints compared with vanilla vLLM."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-sieveserve">
<titleInfo>
<title>SieveServe: Adaptive Visual Token Pruning for Reliable and Efficient Multimodal LLM Inference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xiao</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiao</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cheng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sa</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhining</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaofang</namePart>
<namePart type="family">Zhao</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>Efficient serving of multimodal large language models (MLLMs) is increasingly bottlenecked by high-resolution visual inputs, which can generate thousands of visual tokens per request and severely increase latency, memory pressure, and serving cost. Visual token pruning is a natural solution, but existing methods are difficult to deploy in production MLLM serving: many operate inside LLM layers and conflict with engine invariants such as KV-cache granularity, fused kernels, prefix caching, and batching, while fixed-ratio pruning can collapse accuracy on dense-information tasks such as document and chart understanding.\par We present SieveServe, a serving-compatible framework for adaptive visual token pruning. SieveServe establishes pre-LLM vision-centric pruning as an engine-compatible paradigm and introduces a request-aware dual-guard policy: a Query Complexity Router (QCR) estimates query-side visual evidence demand, and a Dense Information Guard (DIG) detects image-side pruning risk using lightweight channel-wise embedding variance. Together, they allocate visual-token budgets per request, aggressively compressing pruning-robust inputs while conservatively preserving dense/high-risk ones.\par Experiments with Qwen2.5-VL-7B and InternVL3-8B across eight benchmarks show that SieveServe retains at least 95% of baseline accuracy on all benchmarks and at least 97.8% on dense-information tasks, while compressing pruning-robust tasks by 3.2x and recovering over 20 accuracy points lost by fixed-ratio pruning. On high-resolution serving workloads, SieveServe reduces P90 TTFT by 5.4x and sustains 2.2-2.7x higher throughput under latency SLO constraints compared with vanilla vLLM.</abstract>
<identifier type="citekey">li-etal-2026-sieveserve</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.648/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14918</start>
<end>14934</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SieveServe: Adaptive Visual Token Pruning for Reliable and Efficient Multimodal LLM Inference
%A Li, Xiao
%A Shi, Xiao
%A Zhang, Cheng
%A Wang, Sa
%A Zhang, Zhining
%A Zhao, Xiaofang
%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 li-etal-2026-sieveserve
%X Efficient serving of multimodal large language models (MLLMs) is increasingly bottlenecked by high-resolution visual inputs, which can generate thousands of visual tokens per request and severely increase latency, memory pressure, and serving cost. Visual token pruning is a natural solution, but existing methods are difficult to deploy in production MLLM serving: many operate inside LLM layers and conflict with engine invariants such as KV-cache granularity, fused kernels, prefix caching, and batching, while fixed-ratio pruning can collapse accuracy on dense-information tasks such as document and chart understanding.\par We present SieveServe, a serving-compatible framework for adaptive visual token pruning. SieveServe establishes pre-LLM vision-centric pruning as an engine-compatible paradigm and introduces a request-aware dual-guard policy: a Query Complexity Router (QCR) estimates query-side visual evidence demand, and a Dense Information Guard (DIG) detects image-side pruning risk using lightweight channel-wise embedding variance. Together, they allocate visual-token budgets per request, aggressively compressing pruning-robust inputs while conservatively preserving dense/high-risk ones.\par Experiments with Qwen2.5-VL-7B and InternVL3-8B across eight benchmarks show that SieveServe retains at least 95% of baseline accuracy on all benchmarks and at least 97.8% on dense-information tasks, while compressing pruning-robust tasks by 3.2x and recovering over 20 accuracy points lost by fixed-ratio pruning. On high-resolution serving workloads, SieveServe reduces P90 TTFT by 5.4x and sustains 2.2-2.7x higher throughput under latency SLO constraints compared with vanilla vLLM.
%U https://aclanthology.org/2026.emnlp-main.648/
%P 14918-14934
Markdown (Informal)
[SieveServe: Adaptive Visual Token Pruning for Reliable and Efficient Multimodal LLM Inference](https://aclanthology.org/2026.emnlp-main.648/) (Li et al., EMNLP 2026)
ACL