@inproceedings{gilabert-etal-2025-investigating,
title = "Investigating the translation capabilities of Large Language Models trained on parallel data only",
author = "Gilabert, Javier Garc{\'i}a and
Escolano, Carlos and
Sant, Aleix and
Fornaciari, Francesca De Luca and
Mash, Audrey and
Liao, Xixian and
Melero, Maite",
editor = "Bouillon, Pierrette and
Gerlach, Johanna and
Girletti, Sabrina and
Volkart, Lise and
Rubino, Raphael and
Sennrich, Rico and
Farinha, Ana C. and
Gaido, Marco and
Daems, Joke and
Kenny, Dorothy and
Moniz, Helena and
Szoc, Sara",
booktitle = "Proceedings of Machine Translation Summit XX: Volume 1",
month = jun,
year = "2025",
address = "Geneva, Switzerland",
publisher = "European Association for Machine Translation",
url = "https://aclanthology.org/2025.mtsummit-1.4/",
pages = "24--53",
ISBN = "978-2-9701897-0-1",
abstract = "In recent years, Large Language Models (LLMs) have demonstrated exceptional proficiency across a broad spectrum of Natural Language Processing (NLP) tasks, including Machine Translation. However, previous methods predominantly relied on iterative processes such as instruction fine-tuning or continual pre-training, leaving unexplored the challenges of training LLMs solely on parallel data. In this work, we introduce Plume (Parallel Language Model), a collection of three 2B LLMs featuring varying vocabulary sizes (32k, 128k, and 256k) trained exclusively on Catalan-centric parallel examples. These models perform comparably to previous encoder-decoder architectures on 16 supervised translation directions and 56 zero-shot ones. Utilizing this set of models, we conduct a thorough investigation into the translation capabilities of LLMs, probing their performance, the role of vocabulary size, the impact of the different elements of the prompt, and their cross-lingual representation space. We find that larger vocabulary sizes improve zero-shot performance and that different layers specialize in distinct aspects of the prompt, such as language-specific tags. We further show that as the vocabulary size grows, a larger number of attention heads can be pruned with minimal loss in translation quality, achieving a reduction of over 64.7{\%} in attention heads."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gilabert-etal-2025-investigating">
<titleInfo>
<title>Investigating the translation capabilities of Large Language Models trained on parallel data only</title>
</titleInfo>
<name type="personal">
<namePart type="given">Javier</namePart>
<namePart type="given">García</namePart>
<namePart type="family">Gilabert</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Carlos</namePart>
<namePart type="family">Escolano</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleix</namePart>
<namePart type="family">Sant</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Francesca</namePart>
<namePart type="given">De</namePart>
<namePart type="given">Luca</namePart>
<namePart type="family">Fornaciari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Audrey</namePart>
<namePart type="family">Mash</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xixian</namePart>
<namePart type="family">Liao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maite</namePart>
<namePart type="family">Melero</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of Machine Translation Summit XX: Volume 1</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pierrette</namePart>
<namePart type="family">Bouillon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Johanna</namePart>
<namePart type="family">Gerlach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sabrina</namePart>
<namePart type="family">Girletti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lise</namePart>
<namePart type="family">Volkart</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raphael</namePart>
<namePart type="family">Rubino</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rico</namePart>
<namePart type="family">Sennrich</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ana</namePart>
<namePart type="given">C</namePart>
<namePart type="family">Farinha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Gaido</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joke</namePart>
<namePart type="family">Daems</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dorothy</namePart>
<namePart type="family">Kenny</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Helena</namePart>
<namePart type="family">Moniz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sara</namePart>
<namePart type="family">Szoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>European Association for Machine Translation</publisher>
<place>
<placeTerm type="text">Geneva, Switzerland</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">978-2-9701897-0-1</identifier>
</relatedItem>
<abstract>In recent years, Large Language Models (LLMs) have demonstrated exceptional proficiency across a broad spectrum of Natural Language Processing (NLP) tasks, including Machine Translation. However, previous methods predominantly relied on iterative processes such as instruction fine-tuning or continual pre-training, leaving unexplored the challenges of training LLMs solely on parallel data. In this work, we introduce Plume (Parallel Language Model), a collection of three 2B LLMs featuring varying vocabulary sizes (32k, 128k, and 256k) trained exclusively on Catalan-centric parallel examples. These models perform comparably to previous encoder-decoder architectures on 16 supervised translation directions and 56 zero-shot ones. Utilizing this set of models, we conduct a thorough investigation into the translation capabilities of LLMs, probing their performance, the role of vocabulary size, the impact of the different elements of the prompt, and their cross-lingual representation space. We find that larger vocabulary sizes improve zero-shot performance and that different layers specialize in distinct aspects of the prompt, such as language-specific tags. We further show that as the vocabulary size grows, a larger number of attention heads can be pruned with minimal loss in translation quality, achieving a reduction of over 64.7% in attention heads.</abstract>
<identifier type="citekey">gilabert-etal-2025-investigating</identifier>
<location>
<url>https://aclanthology.org/2025.mtsummit-1.4/</url>
</location>
<part>
<date>2025-06</date>
<extent unit="page">
<start>24</start>
<end>53</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Investigating the translation capabilities of Large Language Models trained on parallel data only
%A Gilabert, Javier García
%A Escolano, Carlos
%A Sant, Aleix
%A Fornaciari, Francesca De Luca
%A Mash, Audrey
%A Liao, Xixian
%A Melero, Maite
%Y Bouillon, Pierrette
%Y Gerlach, Johanna
%Y Girletti, Sabrina
%Y Volkart, Lise
%Y Rubino, Raphael
%Y Sennrich, Rico
%Y Farinha, Ana C.
%Y Gaido, Marco
%Y Daems, Joke
%Y Kenny, Dorothy
%Y Moniz, Helena
%Y Szoc, Sara
%S Proceedings of Machine Translation Summit XX: Volume 1
%D 2025
%8 June
%I European Association for Machine Translation
%C Geneva, Switzerland
%@ 978-2-9701897-0-1
%F gilabert-etal-2025-investigating
%X In recent years, Large Language Models (LLMs) have demonstrated exceptional proficiency across a broad spectrum of Natural Language Processing (NLP) tasks, including Machine Translation. However, previous methods predominantly relied on iterative processes such as instruction fine-tuning or continual pre-training, leaving unexplored the challenges of training LLMs solely on parallel data. In this work, we introduce Plume (Parallel Language Model), a collection of three 2B LLMs featuring varying vocabulary sizes (32k, 128k, and 256k) trained exclusively on Catalan-centric parallel examples. These models perform comparably to previous encoder-decoder architectures on 16 supervised translation directions and 56 zero-shot ones. Utilizing this set of models, we conduct a thorough investigation into the translation capabilities of LLMs, probing their performance, the role of vocabulary size, the impact of the different elements of the prompt, and their cross-lingual representation space. We find that larger vocabulary sizes improve zero-shot performance and that different layers specialize in distinct aspects of the prompt, such as language-specific tags. We further show that as the vocabulary size grows, a larger number of attention heads can be pruned with minimal loss in translation quality, achieving a reduction of over 64.7% in attention heads.
%U https://aclanthology.org/2025.mtsummit-1.4/
%P 24-53
Markdown (Informal)
[Investigating the translation capabilities of Large Language Models trained on parallel data only](https://aclanthology.org/2025.mtsummit-1.4/) (Gilabert et al., MTSummit 2025)
ACL