Reasoning makes translations worse

LLM reasoning has a simple appeal: give the model more room to think, and the answer should improve. That intuition works well enough in domains like math, coding, and planning that it is tempting to apply it everywhere. Machine translation seems like an obvious candidate. Translation is not just word substitution: a good translator resolves ambiguity, preserves tone, chooses between literal and idiomatic phrasing, and revises awkward drafts. Surely a model that “reasons” before translating should do better. The recent evidence says that's not always the case. Thinking damages the translation quality The most evident proof of this effect comes from Rajaee et al., Unlocking Reasoning Capability on Machine Translation in Large Language Models. They compare several reasoning-oriented LLMs on translation tasks with and without explicit reasoning. The result is counterintuitive: the models generally do better without reasoning. In their WMT24++ evaluation, every tested reasoning model has…