@inproceedings{qian-scherrer-2026-large,
title = "Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation",
author = "Qian, Shenbin and
Scherrer, Yves",
editor = "Shterionov, Dimitar and
Vanmassenhove, Eva and
De Sisto, Mirella and
Blain, Fred and
Pourmostafa Roshan Sharami, Javad and
Lepp, Lisa and
Manna, Chiara and
Rescigno, Argentina Anna and
Karakanta, Alina and
Rigouts Terryn, Ayla and
Lardelli, Manuel and
Resende, Natalia and
Murgolo, Elena and
Hackenbuchner, Jani{\c{c}}a and
Zaretskaya, Anna and
Espl{\`a}-Gomis, Miquel and
Etchegoyhen, Thierry and
Gromann, Dagmar and
Bawden, Rachel and
Haddow, Barry and
Szoc, Sara and
Forcada, Mikel and
Moniz, Helena",
booktitle = "Proceedings of the 26th Annual Conference of the {E}uropean Association for Machine Translation (Volume 1)",
month = jun,
year = "2026",
address = "Tilburg, The Netherlands",
publisher = "European Association for Machine Translation",
url = "https://aclanthology.org/2026.eamt-1.4/",
pages = "22--47",
ISBN = "9789403901411",
abstract = "Large Language Models (LLMs) have recently demonstrated strong performance in machine translation (MT). However, most prior work focuses on improving or benchmarking translation quality, offering limited insight into when and why LLM-based translation fails. In this work, we systematically analyze failure modes of LLMs in MT by evaluating 15 models, including four reasoning LLMs, across 22 language pairs (LPs) with varying resource levels. We find that non-English-centric LPs consistently yield lower COMET scores than English-centric pairs. To investigate the underlying causes, we introduce Token Activation Rate (TAR), a metric that captures how effectively a model utilizes language-specific tokens in its vocabulary during generation. We validate TAR as a proxy for language representation using models with known language distributions in the training data, and show that lower TAR is strongly associated with poorer translation performance. Furthermore, reasoning LLMs tend to generate more tokens when translating into low-TAR languages, suggesting a compensatory mechanism, although its impact on translation quality varies across models. Overall, our findings emphasize the importance of token-level dynamics in understanding MT performance of LLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="qian-scherrer-2026-large">
<titleInfo>
<title>Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shenbin</namePart>
<namePart type="family">Qian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yves</namePart>
<namePart type="family">Scherrer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dimitar</namePart>
<namePart type="family">Shterionov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Eva</namePart>
<namePart type="family">Vanmassenhove</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mirella</namePart>
<namePart type="family">De Sisto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fred</namePart>
<namePart type="family">Blain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Javad</namePart>
<namePart type="family">Pourmostafa Roshan Sharami</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lisa</namePart>
<namePart type="family">Lepp</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chiara</namePart>
<namePart type="family">Manna</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Argentina</namePart>
<namePart type="given">Anna</namePart>
<namePart type="family">Rescigno</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alina</namePart>
<namePart type="family">Karakanta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayla</namePart>
<namePart type="family">Rigouts Terryn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Manuel</namePart>
<namePart type="family">Lardelli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Natalia</namePart>
<namePart type="family">Resende</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elena</namePart>
<namePart type="family">Murgolo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Janiça</namePart>
<namePart type="family">Hackenbuchner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Zaretskaya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Miquel</namePart>
<namePart type="family">Esplà-Gomis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thierry</namePart>
<namePart type="family">Etchegoyhen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dagmar</namePart>
<namePart type="family">Gromann</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rachel</namePart>
<namePart type="family">Bawden</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<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">Sara</namePart>
<namePart type="family">Szoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mikel</namePart>
<namePart type="family">Forcada</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>
<originInfo>
<publisher>European Association for Machine Translation</publisher>
<place>
<placeTerm type="text">Tilburg, The Netherlands</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">9789403901411</identifier>
</relatedItem>
<abstract>Large Language Models (LLMs) have recently demonstrated strong performance in machine translation (MT). However, most prior work focuses on improving or benchmarking translation quality, offering limited insight into when and why LLM-based translation fails. In this work, we systematically analyze failure modes of LLMs in MT by evaluating 15 models, including four reasoning LLMs, across 22 language pairs (LPs) with varying resource levels. We find that non-English-centric LPs consistently yield lower COMET scores than English-centric pairs. To investigate the underlying causes, we introduce Token Activation Rate (TAR), a metric that captures how effectively a model utilizes language-specific tokens in its vocabulary during generation. We validate TAR as a proxy for language representation using models with known language distributions in the training data, and show that lower TAR is strongly associated with poorer translation performance. Furthermore, reasoning LLMs tend to generate more tokens when translating into low-TAR languages, suggesting a compensatory mechanism, although its impact on translation quality varies across models. Overall, our findings emphasize the importance of token-level dynamics in understanding MT performance of LLMs.</abstract>
<identifier type="citekey">qian-scherrer-2026-large</identifier>
<location>
<url>https://aclanthology.org/2026.eamt-1.4/</url>
</location>
<part>
<date>2026-06</date>
<extent unit="page">
<start>22</start>
<end>47</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation
%A Qian, Shenbin
%A Scherrer, Yves
%Y Shterionov, Dimitar
%Y Vanmassenhove, Eva
%Y De Sisto, Mirella
%Y Blain, Fred
%Y Pourmostafa Roshan Sharami, Javad
%Y Lepp, Lisa
%Y Manna, Chiara
%Y Rescigno, Argentina Anna
%Y Karakanta, Alina
%Y Rigouts Terryn, Ayla
%Y Lardelli, Manuel
%Y Resende, Natalia
%Y Murgolo, Elena
%Y Hackenbuchner, Janiça
%Y Zaretskaya, Anna
%Y Esplà-Gomis, Miquel
%Y Etchegoyhen, Thierry
%Y Gromann, Dagmar
%Y Bawden, Rachel
%Y Haddow, Barry
%Y Szoc, Sara
%Y Forcada, Mikel
%Y Moniz, Helena
%S Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 1)
%D 2026
%8 June
%I European Association for Machine Translation
%C Tilburg, The Netherlands
%@ 9789403901411
%F qian-scherrer-2026-large
%X Large Language Models (LLMs) have recently demonstrated strong performance in machine translation (MT). However, most prior work focuses on improving or benchmarking translation quality, offering limited insight into when and why LLM-based translation fails. In this work, we systematically analyze failure modes of LLMs in MT by evaluating 15 models, including four reasoning LLMs, across 22 language pairs (LPs) with varying resource levels. We find that non-English-centric LPs consistently yield lower COMET scores than English-centric pairs. To investigate the underlying causes, we introduce Token Activation Rate (TAR), a metric that captures how effectively a model utilizes language-specific tokens in its vocabulary during generation. We validate TAR as a proxy for language representation using models with known language distributions in the training data, and show that lower TAR is strongly associated with poorer translation performance. Furthermore, reasoning LLMs tend to generate more tokens when translating into low-TAR languages, suggesting a compensatory mechanism, although its impact on translation quality varies across models. Overall, our findings emphasize the importance of token-level dynamics in understanding MT performance of LLMs.
%U https://aclanthology.org/2026.eamt-1.4/
%P 22-47
Markdown (Informal)
[Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation](https://aclanthology.org/2026.eamt-1.4/) (Qian & Scherrer, EAMT 2026)
ACL