Skip to content
conftrace_
2026 EMNLP EMNLP 2026

LLM Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens

AbstractLarge 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.

More from EMNLP 2026

All EMNLP 2026 papers →