@inproceedings{raya-rios-etal-2026-tiny,
title = "Tiny Titans at {WMT}26: Task-Calibrated Quantization and Activation-Aware Low-Rank Compression of Gemma 3 12{B}",
author = "Raya-Rios, Vania and
Klimchuk, Aleksandr and
Cuadrado Avila, Nicolas and
Gollini Navarrete, Ivo and
Horvath, Samuel and
Tak{\'a}c, Martin",
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.141/",
pages = "2030--2037",
ISBN = "979-8-89176-474-3",
abstract = "We present the TinyTitans submissions for the WMT26 Model Compression shared task, which focuses on Czech-to-German translation. Starting with the instruction-tuned Gemma 3 12B model, we compare two approaches: 4-bit GPTQ, which uses task-specific calibration examples, and a more aggressive low-rank factorization method that replaces attention and feed-forward projections with truncated Singular Value Decomposition (SVD) factors. On a 300-sentence development set, the calibrated GPTQ retains 99.7{\%} of the dense model{'}s COMET score while reducing GPU memory usage from 26.6 GiB to 17.6 GiB. The recovered SVD model reduces the number of parameters by 46.6{\%} and reaches 0.7804 COMET in a separate evaluation using 1,997 sentence pairs. In an external evaluation using 457 WMT25 paragraphs, the GPTQ, SVD, and SVD-plus-quantization submissions received COMET scores of 0.5751, 0.4263, and 0.3354, respectively. These findings suggest that while post-training quantization effectively reduces memory usage, combining low-rank truncation with low-bit quantization may increase approximation errors without clear efficiency gains."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="raya-rios-etal-2026-tiny">
<titleInfo>
<title>Tiny Titans at WMT26: Task-Calibrated Quantization and Activation-Aware Low-Rank Compression of Gemma 3 12B</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vania</namePart>
<namePart type="family">Raya-Rios</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleksandr</namePart>
<namePart type="family">Klimchuk</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nicolas</namePart>
<namePart type="family">Cuadrado Avila</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ivo</namePart>
<namePart type="family">Gollini Navarrete</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samuel</namePart>
<namePart type="family">Horvath</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Martin</namePart>
<namePart type="family">Takác</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 TinyTitans submissions for the WMT26 Model Compression shared task, which focuses on Czech-to-German translation. Starting with the instruction-tuned Gemma 3 12B model, we compare two approaches: 4-bit GPTQ, which uses task-specific calibration examples, and a more aggressive low-rank factorization method that replaces attention and feed-forward projections with truncated Singular Value Decomposition (SVD) factors. On a 300-sentence development set, the calibrated GPTQ retains 99.7% of the dense model’s COMET score while reducing GPU memory usage from 26.6 GiB to 17.6 GiB. The recovered SVD model reduces the number of parameters by 46.6% and reaches 0.7804 COMET in a separate evaluation using 1,997 sentence pairs. In an external evaluation using 457 WMT25 paragraphs, the GPTQ, SVD, and SVD-plus-quantization submissions received COMET scores of 0.5751, 0.4263, and 0.3354, respectively. These findings suggest that while post-training quantization effectively reduces memory usage, combining low-rank truncation with low-bit quantization may increase approximation errors without clear efficiency gains.</abstract>
<identifier type="citekey">raya-rios-etal-2026-tiny</identifier>
<location>
<url>https://aclanthology.org/2026.wmt-main.141/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>2030</start>
<end>2037</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Tiny Titans at WMT26: Task-Calibrated Quantization and Activation-Aware Low-Rank Compression of Gemma 3 12B
%A Raya-Rios, Vania
%A Klimchuk, Aleksandr
%A Cuadrado Avila, Nicolas
%A Gollini Navarrete, Ivo
%A Horvath, Samuel
%A Takác, Martin
%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 raya-rios-etal-2026-tiny
%X We present the TinyTitans submissions for the WMT26 Model Compression shared task, which focuses on Czech-to-German translation. Starting with the instruction-tuned Gemma 3 12B model, we compare two approaches: 4-bit GPTQ, which uses task-specific calibration examples, and a more aggressive low-rank factorization method that replaces attention and feed-forward projections with truncated Singular Value Decomposition (SVD) factors. On a 300-sentence development set, the calibrated GPTQ retains 99.7% of the dense model’s COMET score while reducing GPU memory usage from 26.6 GiB to 17.6 GiB. The recovered SVD model reduces the number of parameters by 46.6% and reaches 0.7804 COMET in a separate evaluation using 1,997 sentence pairs. In an external evaluation using 457 WMT25 paragraphs, the GPTQ, SVD, and SVD-plus-quantization submissions received COMET scores of 0.5751, 0.4263, and 0.3354, respectively. These findings suggest that while post-training quantization effectively reduces memory usage, combining low-rank truncation with low-bit quantization may increase approximation errors without clear efficiency gains.
%U https://aclanthology.org/2026.wmt-main.141/
%P 2030-2037
Markdown (Informal)
[Tiny Titans at WMT26: Task-Calibrated Quantization and Activation-Aware Low-Rank Compression of Gemma 3 12B](https://aclanthology.org/2026.wmt-main.141/) (Raya-Rios et al., WMT 2026)
ACL