@inproceedings{zebaze-etal-2026-llm,
title = "{LLM} Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens",
author = "Zebaze, Armel Randy and
Bawden, Rachel and
Sagot, Beno{\^i}t",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1302/",
pages = "29644--29682",
ISBN = "979-8-89176-459-0",
abstract = "Large reasoning models (LRMs) have led to new possibilities in terms of problem-solving, through the devising of a natural language thought process prior to answering a query. While their capabilities are well known across mathematics and coding tasks, their impact on the task of machine translation (MT) remains underexplored. In this work, we explore the benefits of the generation of intermediate tokens when performing MT across multiple language pairs of different levels of resourcedness and multiple setups. We first show that the ``thinking tokens'' of general-purpose LRMs do not help them better perform MT, motivating the question of what kind of intermediate tokens could help when building a dedicated thinking MT model via fine-tuning. We then show that distilling chains of thought (CoT) from a teacher model using CoT templates inspired by human translators' step-by-step practices fails to outperform standard input-output fine-tuning. In contrast, constructing CoT traces with prompting strategies specifically designed for MT does yield improvements. A closer analysis reveals that these gains stem from the teacher{'}s translation attempts embedded in the traces, rather than from the surrounding reasoning. More broadly, our results suggest that leveraging a teacher to refine target translations or expand parallel corpora is more impactful than distilling its CoT explanations into ``thinking'' MT models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zebaze-etal-2026-llm">
<titleInfo>
<title>LLM Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens</title>
</titleInfo>
<name type="personal">
<namePart type="given">Armel</namePart>
<namePart type="given">Randy</namePart>
<namePart type="family">Zebaze</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rachel</namePart>
<namePart type="family">Bawden</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Benoît</namePart>
<namePart type="family">Sagot</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 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Large reasoning models (LRMs) have led to new possibilities in terms of problem-solving, through the devising of a natural language thought process prior to answering a query. While their capabilities are well known across mathematics and coding tasks, their impact on the task of machine translation (MT) remains underexplored. In this work, we explore the benefits of the generation of intermediate tokens when performing MT across multiple language pairs of different levels of resourcedness and multiple setups. We first show that the “thinking tokens” of general-purpose LRMs do not help them better perform MT, motivating the question of what kind of intermediate tokens could help when building a dedicated thinking MT model via fine-tuning. We then show that distilling chains of thought (CoT) from a teacher model using CoT templates inspired by human translators’ step-by-step practices fails to outperform standard input-output fine-tuning. In contrast, constructing CoT traces with prompting strategies specifically designed for MT does yield improvements. A closer analysis reveals that these gains stem from the teacher’s translation attempts embedded in the traces, rather than from the surrounding reasoning. More broadly, our results suggest that leveraging a teacher to refine target translations or expand parallel corpora is more impactful than distilling its CoT explanations into “thinking” MT models.</abstract>
<identifier type="citekey">zebaze-etal-2026-llm</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1302/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>29644</start>
<end>29682</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LLM Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens
%A Zebaze, Armel Randy
%A Bawden, Rachel
%A Sagot, Benoît
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F zebaze-etal-2026-llm
%X Large reasoning models (LRMs) have led to new possibilities in terms of problem-solving, through the devising of a natural language thought process prior to answering a query. While their capabilities are well known across mathematics and coding tasks, their impact on the task of machine translation (MT) remains underexplored. In this work, we explore the benefits of the generation of intermediate tokens when performing MT across multiple language pairs of different levels of resourcedness and multiple setups. We first show that the “thinking tokens” of general-purpose LRMs do not help them better perform MT, motivating the question of what kind of intermediate tokens could help when building a dedicated thinking MT model via fine-tuning. We then show that distilling chains of thought (CoT) from a teacher model using CoT templates inspired by human translators’ step-by-step practices fails to outperform standard input-output fine-tuning. In contrast, constructing CoT traces with prompting strategies specifically designed for MT does yield improvements. A closer analysis reveals that these gains stem from the teacher’s translation attempts embedded in the traces, rather than from the surrounding reasoning. More broadly, our results suggest that leveraging a teacher to refine target translations or expand parallel corpora is more impactful than distilling its CoT explanations into “thinking” MT models.
%U https://aclanthology.org/2026.emnlp-main.1302/
%P 29644-29682
Markdown (Informal)
[LLM Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens](https://aclanthology.org/2026.emnlp-main.1302/) (Zebaze et al., EMNLP 2026)
ACL