@inproceedings{attimonelli-etal-2026-puma,
title = "{PUMA}: Post-Hoc Sparsification of Universal Multimodal Embeddings for Efficient Retrieval",
author = "Attimonelli, Matteo and
De Bellis, Alessandro and
Nardini, Franco Maria and
Pomo, Claudio and
Rulli, Cosimo and
Venturini, Rossano and
Di Noia, Tommaso",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1247/",
doi = "10.18653/2026.findings-emnlp.1247",
pages = "26388--26401",
ISBN = "979-8-89176-462-0",
abstract = "Universal multimodal embedders enable retrieval across text, image, and combined queries, but their dense representations incur high memory and inference costs. Post-hoc sparsification could reduce these costs but remains underexplored for multimodal retrieval. We introduce PUMA, a sparse autoencoder recipe that maps universal multimodal embeddings to compact sparse codes without retraining the backbone: a pretraining stage preserves dense dot-product geometry, after which the sparse encoder is fine-tuned for retrieval. We evaluate on five benchmarks covering text-to-image and composed image retrieval. On Qwen3-VL-Embedding-2B, PUMA is statistically indistinguishable from or improves over dense retrieval on four of five datasets. We further identify two failure modes of post-hoc sparsification: insufficient pre-TopK support and retrieval-misaligned active support. PUMA reduces per-vector storage by 8-16x (FP32) and runs up to 25x faster over exact dense scoring on larger candidate pools, offering a practical approach to storage- and latency-efficient multimodal retrieval."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="attimonelli-etal-2026-puma">
<titleInfo>
<title>PUMA: Post-Hoc Sparsification of Universal Multimodal Embeddings for Efficient Retrieval</title>
</titleInfo>
<name type="personal">
<namePart type="given">Matteo</namePart>
<namePart type="family">Attimonelli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">De Bellis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Franco</namePart>
<namePart type="given">Maria</namePart>
<namePart type="family">Nardini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Claudio</namePart>
<namePart type="family">Pomo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cosimo</namePart>
<namePart type="family">Rulli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rossano</namePart>
<namePart type="family">Venturini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tommaso</namePart>
<namePart type="family">Di Noia</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>Universal multimodal embedders enable retrieval across text, image, and combined queries, but their dense representations incur high memory and inference costs. Post-hoc sparsification could reduce these costs but remains underexplored for multimodal retrieval. We introduce PUMA, a sparse autoencoder recipe that maps universal multimodal embeddings to compact sparse codes without retraining the backbone: a pretraining stage preserves dense dot-product geometry, after which the sparse encoder is fine-tuned for retrieval. We evaluate on five benchmarks covering text-to-image and composed image retrieval. On Qwen3-VL-Embedding-2B, PUMA is statistically indistinguishable from or improves over dense retrieval on four of five datasets. We further identify two failure modes of post-hoc sparsification: insufficient pre-TopK support and retrieval-misaligned active support. PUMA reduces per-vector storage by 8-16x (FP32) and runs up to 25x faster over exact dense scoring on larger candidate pools, offering a practical approach to storage- and latency-efficient multimodal retrieval.</abstract>
<identifier type="citekey">attimonelli-etal-2026-puma</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1247</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1247/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>26388</start>
<end>26401</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T PUMA: Post-Hoc Sparsification of Universal Multimodal Embeddings for Efficient Retrieval
%A Attimonelli, Matteo
%A De Bellis, Alessandro
%A Nardini, Franco Maria
%A Pomo, Claudio
%A Rulli, Cosimo
%A Venturini, Rossano
%A Di Noia, Tommaso
%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 attimonelli-etal-2026-puma
%X Universal multimodal embedders enable retrieval across text, image, and combined queries, but their dense representations incur high memory and inference costs. Post-hoc sparsification could reduce these costs but remains underexplored for multimodal retrieval. We introduce PUMA, a sparse autoencoder recipe that maps universal multimodal embeddings to compact sparse codes without retraining the backbone: a pretraining stage preserves dense dot-product geometry, after which the sparse encoder is fine-tuned for retrieval. We evaluate on five benchmarks covering text-to-image and composed image retrieval. On Qwen3-VL-Embedding-2B, PUMA is statistically indistinguishable from or improves over dense retrieval on four of five datasets. We further identify two failure modes of post-hoc sparsification: insufficient pre-TopK support and retrieval-misaligned active support. PUMA reduces per-vector storage by 8-16x (FP32) and runs up to 25x faster over exact dense scoring on larger candidate pools, offering a practical approach to storage- and latency-efficient multimodal retrieval.
%R 10.18653/2026.findings-emnlp.1247
%U https://aclanthology.org/2026.findings-emnlp.1247/
%U https://doi.org/10.18653/2026.findings-emnlp.1247
%P 26388-26401
Markdown (Informal)
[PUMA: Post-Hoc Sparsification of Universal Multimodal Embeddings for Efficient Retrieval](https://aclanthology.org/2026.findings-emnlp.1247/) (Attimonelli et al., Findings 2026)
ACL