Reasoning Before Translation: Enhancing Legal Machine Translation with Structured Reasoning

This paper investigates how to improve neural machine translation (NMT) for the highly specialized and linguistically complex domain of Swiss legal texts. The authors explore whether integrating structured reasoning—where a model 'thinks' before translating—can enhance the performance of small language models (SLMs). They compare three distinct paradigms: prompting large reasoning models, supervised fine-tuning (SFT) with intermediate reasoning steps, and reinforcement learning (RL) using Group Relative Policy Optimization (GRPO).

The researchers utilized the SwissLawTranslations (CH-Law-Trans) dataset, creating a new subset of 40,000 translation pairs enriched with reasoning steps generated by the DeepSeek-R1 model. They evaluated three SLMs (Qwen3.5 4B, Qwen3.5 9B, and Gemma 3 12B) across different training setups. In the RL experiments, they implemented a reward system based on output formatting and translation quality (using the ChrF score) to guide the models toward more accurate and structured outputs.

The study reveals that while both SFT and RL improve translation quality, RL provides more substantial performance gains. Interestingly, 'Reasoning SFT'—fine-tuning on explicit reasoning steps—did not consistently outperform 'Simple SFT' (direct translation training). The researchers found that model size remains a critical factor, with the GRPO-trained Gemma 3 12B achieving the highest performance among the smaller models, coming close to the quality of frontier models like OpenAI's o4-mini. Finally, they observed that the benefits of these re-training paradigms diminish as the base model size increases.

Legal translation requires extreme precision and adherence to formal registers, which often makes it difficult for standard NMT systems. By demonstrating that RL can effectively 'teach' smaller, more cost-effective models to reason through translation challenges, this work provides a practical blueprint for deploying high-quality, domain-specific translation systems in environments where data privacy or computational costs preclude the use of massive, proprietary frontier models.