Reasoning Before Translation: Enhancing Legal Machine Translation with Structured Reasoning

Neural machine translation (NMT) in the legal domain is a linguistically and conceptually demanding task, primarily due to the complexity of legal language and the high level of precision it requires. The recent emergence of reasoning-capable language models opens new possibilities for tackling such challenges. They complement a broader set of techniques previously proposed to improve translation quality, including supervised fine-tuning and reinforcement learning. In this work, we compare these different approaches. More specifically, we evaluate small language models such as Qwen3.5 4B, Qwen3.5 9B, and Gemma 3 12B enhanced with various re-training paradigms, and compare their performance against frontier reasoning models. We focus on the Swiss legal system, whose uniquely multilingual statutes provide a particularly challenging and realistic testbed for reasoning-augmented models. Our results show that the quality of small base models can be significantly improved, and that reinforcement learning with verifiable rewards can be effectively applied to legal-domain NMT, surpassing the translation quality achieved through supervised fine-tuning. Notably, these enhanced small models achieve competitive performance while using more than 100× fewer parameters than frontier reasoning models, although a performance gap still remains. We also observe that re-training paradigms yield diminishing returns as model size increases.