@inproceedings{gowda-etal-2026-findings,
title = "Findings of the {WMT} 2026 Shared Task on Model Compression: No Free Lunch at Extreme Compression",
author = "Gowda, Thamme and
Gaido, Marco and
Grundkiewicz, Roman and
Negri, Matteo",
editor = "Haddow, Barry and
Kocmi, Tom and
Koehn, Philipp and
Monz, Christof",
booktitle = "Proceedings of the Eleventh Conference on Machine Translation",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.wmt-main.52/",
pages = "1080--1090",
ISBN = "979-8-89176-474-3",
abstract = "We present the second edition of the WMT 2026 Model Compression (WMT26MC) shared task, which studies how large language models can be made efficient for machine translation under model-size and latency constraints. WMT26MC offers a constrained track based on Gemma{\textasciitilde}3{\textasciitilde}12B and an unconstrained track for compressing models below 20B. Both tracks cover cs-de, en-zh, and en-ar. Participants submit complete runnable systems, which we run in a standardized environment on single H100 GPU. Evaluation jointly considers translation quality, model size on disk, and decoding speed, analyzing the resulting quality{--}size and quality{--}speed trade-offs. We received 41 submissions from 13 teams. The proposed systems span weight and activation quantization, structured and expert pruning, low-rank factorization, knowledge distillation, vocabulary and vision-component removal, mixed precision, and inference-time reranking. Evaluating with reference-free quality-estimation metrics on the blind test sets, we find that post-training quantization is the reliable lever: INT4 and FP8 systems shrink the constrained base to roughly a quarter of its on-disk size and run up to an order of magnitude faster while staying within quality-estimation noise of the uncompressed model, whereas aggressive low-rank and structural pruning collapse quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gowda-etal-2026-findings">
<titleInfo>
<title>Findings of the WMT 2026 Shared Task on Model Compression: No Free Lunch at Extreme Compression</title>
</titleInfo>
<name type="personal">
<namePart type="given">Thamme</namePart>
<namePart type="family">Gowda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Gaido</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Roman</namePart>
<namePart type="family">Grundkiewicz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matteo</namePart>
<namePart type="family">Negri</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 Eleventh Conference on Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Barry</namePart>
<namePart type="family">Haddow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tom</namePart>
<namePart type="family">Kocmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Philipp</namePart>
<namePart type="family">Koehn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christof</namePart>
<namePart type="family">Monz</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">979-8-89176-474-3</identifier>
</relatedItem>
<abstract>We present the second edition of the WMT 2026 Model Compression (WMT26MC) shared task, which studies how large language models can be made efficient for machine translation under model-size and latency constraints. WMT26MC offers a constrained track based on Gemma~3~12B and an unconstrained track for compressing models below 20B. Both tracks cover cs-de, en-zh, and en-ar. Participants submit complete runnable systems, which we run in a standardized environment on single H100 GPU. Evaluation jointly considers translation quality, model size on disk, and decoding speed, analyzing the resulting quality–size and quality–speed trade-offs. We received 41 submissions from 13 teams. The proposed systems span weight and activation quantization, structured and expert pruning, low-rank factorization, knowledge distillation, vocabulary and vision-component removal, mixed precision, and inference-time reranking. Evaluating with reference-free quality-estimation metrics on the blind test sets, we find that post-training quantization is the reliable lever: INT4 and FP8 systems shrink the constrained base to roughly a quarter of its on-disk size and run up to an order of magnitude faster while staying within quality-estimation noise of the uncompressed model, whereas aggressive low-rank and structural pruning collapse quality.</abstract>
<identifier type="citekey">gowda-etal-2026-findings</identifier>
<location>
<url>https://aclanthology.org/2026.wmt-main.52/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1080</start>
<end>1090</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Findings of the WMT 2026 Shared Task on Model Compression: No Free Lunch at Extreme Compression
%A Gowda, Thamme
%A Gaido, Marco
%A Grundkiewicz, Roman
%A Negri, Matteo
%Y Haddow, Barry
%Y Kocmi, Tom
%Y Koehn, Philipp
%Y Monz, Christof
%S Proceedings of the Eleventh Conference on Machine Translation
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-474-3
%F gowda-etal-2026-findings
%X We present the second edition of the WMT 2026 Model Compression (WMT26MC) shared task, which studies how large language models can be made efficient for machine translation under model-size and latency constraints. WMT26MC offers a constrained track based on Gemma~3~12B and an unconstrained track for compressing models below 20B. Both tracks cover cs-de, en-zh, and en-ar. Participants submit complete runnable systems, which we run in a standardized environment on single H100 GPU. Evaluation jointly considers translation quality, model size on disk, and decoding speed, analyzing the resulting quality–size and quality–speed trade-offs. We received 41 submissions from 13 teams. The proposed systems span weight and activation quantization, structured and expert pruning, low-rank factorization, knowledge distillation, vocabulary and vision-component removal, mixed precision, and inference-time reranking. Evaluating with reference-free quality-estimation metrics on the blind test sets, we find that post-training quantization is the reliable lever: INT4 and FP8 systems shrink the constrained base to roughly a quarter of its on-disk size and run up to an order of magnitude faster while staying within quality-estimation noise of the uncompressed model, whereas aggressive low-rank and structural pruning collapse quality.
%U https://aclanthology.org/2026.wmt-main.52/
%P 1080-1090
Markdown (Informal)
[Findings of the WMT 2026 Shared Task on Model Compression: No Free Lunch at Extreme Compression](https://aclanthology.org/2026.wmt-main.52/) (Gowda et al., WMT 2026)
ACL