The paper tests whether “thinking tokens” help translation and finds they mostly do not. The team tests big reasoning models that write hidden thoughts before the answer. They compare translations with and without those thoughts. Quality barely changes across many language pairs. They also train small models to copy step by step explanations before the final translation. Those models do worse than simple training that maps source text to target text. They try better intermediate traces that are specific to translation. These only help when the traces include actual draft translations. That means the useful signal is more parallel text, not explanations about how to translate. Using stronger teacher translations as targets beats teaching the student to think. Splitting sentences into smaller phrase pairs helps a bit, especially when data is scarce. Extra reinforcement learning does not change the ranking. Plain input to output training still gives the best quality for the compute used. ---- Paper – arxiv. org/abs/2510.11919 Paper Title: "LLM Reasoning for Machine Translation: Synthetic Data Generation over Thinking Tokens"

LLMs struggle with machine translation for low-resource languages, even with similar examples. This paper introduces Compositional Translation (CompTra). It decomposes sentences into phrases, translates each using retrieved examples, and recombines these translations for the final output. 📌 CompTra improves translation quality by processing simpler sentence components. It gains +1.8 XCOMET score with Command-R+. 📌 This method leverages LLMs' strength in translating short phrases for low-resource language machine translation. 📌 Compositional Translation uses self-generated phrase translations as in-context examples. This improves main translation in low resource scenarios. Methods Explored in this Paper 🔧: → CompTra decomposes sentences into simpler phrases using a divide prompt with examples. → Each phrase is independently translated by the LLM using similarity-based retrieved demonstrations. → Finally, the LLM recombines these phrase translations to generate the translation of the original sentence.