💬 NLP

Reasoning Before Translation: Enhancing Legal Machine Translation with Structured Reasoning

This paper evaluates and compares various re-training paradigms, including supervised fine-tuning and reinforcement learning with verifiable rewards, for enhancing small language models in Swiss legal machine translation, finding that while reinforcement learning outperforms fine-tuning and narrows the gap with frontier reasoning models, it still falls short of their performance and yields diminishing returns as model size increases.

Original authors: Aixiu An, Michael Jungo, Eloi Eynard, Mark Drenhaus, Andreas Fischer, Jean Hennebert, Sébastien Rumley

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aixiu An, Michael Jungo, Eloi Eynard, Mark Drenhaus, Andreas Fischer, Jean Hennebert, Sébastien Rumley

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to translate a complex legal contract from one language to another. This isn't like translating a recipe for cookies; it's more like translating a rulebook for a high-stakes video game where every single word changes the rules of the game. In the world of Artificial Intelligence, this is the domain of Neural Machine Translation (NMT). Think of NMT as a super-smart robot brain that has read millions of books and learned to guess the next word in a sentence. Recently, scientists discovered that if you give these robots a "thinking" step before they speak—like asking them to pause and reason through a problem—they get much better at tricky tasks. This is the era of Reasoning Models. But here's the catch: the biggest, smartest robots are incredibly expensive to run, like trying to power a city with a single battery. So, researchers are asking: Can we make smaller, cheaper robots just as smart by teaching them how to think, rather than just giving them more memory?

This paper dives into that exact question, specifically looking at the Swiss legal system, which is a nightmare for translators because it has four official languages and laws that are written in very strict, complicated ways. The researchers wanted to see if they could take small, affordable AI models and boost their legal translation skills using two different training methods: Supervised Fine-Tuning (SFT), which is like showing a student the correct answers and the step-by-step notes on how to get there, and Reinforcement Learning (RL), which is more like a video game where the robot gets points for good translations and loses points for bad ones, learning through trial and error. They also tested if forcing the robot to "think out loud" (generate reasoning steps) actually helped, or if it just confused the poor thing.

The team set up a massive experiment using a dataset of 40,000 legal sentences. They first tried the "think out loud" method by training small models (like the Qwen3.5 4B, Qwen3.5 9B, and Gemma 3 12B) to copy the reasoning steps of a giant, super-smart AI. Surprisingly, this didn't work as hoped. In fact, for the smallest models, forcing them to write out their reasoning steps actually made their translations worse than if they had just been trained on the final answers alone. It's as if telling a nervous student to "write down every thought you have before answering" made them overthink and mess up the test.

However, when they switched to the "video game" approach using Reinforcement Learning (specifically a method called GRPO), the results were fantastic. By rewarding the models for high-quality translations and penalizing bad ones, the small models learned to translate Swiss laws with incredible precision. The best performer, a 12-billion-parameter model called Gemma 3, got so good that its translation quality was close to the most expensive, frontier models that cost thousands of dollars to use, yet it still remained slightly inferior. The researchers found that the "reward-based" learning was consistently better than the "copy-the-notes" method. They also discovered that as the models got bigger, the extra boost from these fancy training methods got smaller; the tiny models gained the most, while the bigger ones only improved a little bit.

In the end, the paper suggests that while the giant, expensive AI models are still the champions, we don't necessarily need them for everything. By using smart training tricks like Reinforcement Learning, we can make small, open-source models that are cheap to run and surprisingly good at handling complex legal translations. The study explicitly rules out the idea that simply copying "reasoning steps" from a big AI helps small models; instead, it suggests that letting them learn through rewards is the secret sauce. While the small models didn't quite beat the giants, they closed the gap significantly, proving that with the right training, a small robot can do a big job in the high-stakes world of law.

tags) preceding the final translation. * *Training:* Utilized LoRA (Rank 16, Alpha 32) with 4-bit quantization on NVIDIA A40 GPUs. Sequence lengths were capped at 512 tokens for Simple SFT and 1024 for Reasoning SFT to prevent repetition. 3. **Experiment III (Reinforcement Learning - RL):** * **Algorithm:** Group Relative Policy Optimization (GRPO). * **Reward Mechanism:** A dual-reward system was employed: * *Format Reward (0.5–1.5):* Based on the correct presence and ordering of(reasoning) and` tags.
* Quality Reward: The ChrF score of the generated translation compared to the ground truth, used directly as a continuous reward signal.
* Configuration: Group size $G=4$ , single epoch training with a reduced learning rate ( $2e^{-5}$ ) to ensure stability.

Evaluation Metrics

Performance was measured using lexical (ChrF, METEOR), semantic (COMET), and LLM-based (MetricX) metrics. Human evaluation was also conducted on a 240-sample subset by a legal expert to validate automatic metrics.

3. Key Contributions

Systematic Comparison: A detailed benchmark comparing prompting frontier models, SFT with intermediate reasoning, and RL-based optimization across different model scales (4B–12B) in a high-stakes legal domain.
New Dataset: The release of a dataset containing 40,000 legal translation pairs enriched with automatically generated intermediate reasoning steps, filtered for quality.
RL Training Recipe: A proposed GRPO training framework with a simple, verifiable reward formulation (combining format compliance and ChrF-based quality) tailored for legal translation.

4. Results

Frontier Models: The OpenAI o3 model achieved the highest performance (COMET ~85.92) but at a significantly higher cost. The Mistral-Small model offered a cost-effective alternative with comparable performance.
RL vs. SFT: Reinforcement Learning consistently outperformed both Simple SFT and Reasoning SFT across all model sizes and metrics.
- RL Performance: The RL-enhanced Gemma 3 12B model achieved a COMET score of 83.08, approaching the performance of the frontier o4-mini model (84.35).
- SFT Performance: Contrary to the hypothesis that explicit reasoning steps improve performance, Reasoning SFT degraded translation quality compared to Simple SFT, particularly in smaller models (e.g., Qwen3.5 4B). Simple SFT itself performed marginally worse than the unaltered base models.
Model Scale and Diminishing Returns: The performance gain from RL was most pronounced in smaller models. The COMET score improvement for Qwen3.5 4B was +2.42 points, whereas Gemma 3 12B improved by only +0.59 points, suggesting diminishing returns as model size increases.
Metric Correlation: Human evaluation revealed that COMET had the strongest correlation ( $r=0.24$ ) with expert legal judgments, outperforming ChrF, METEOR, and MetricX.

5. Significance and Claims

The paper claims that while frontier models remain the gold standard for translation quality, Reinforcement Learning with verifiable rewards is a highly effective strategy for enhancing smaller, open-source models in the legal domain.

Efficiency: RL-enhanced small models can achieve competitive translation quality using over 100 times fewer parameters than frontier models, offering a viable path for cost-effective, on-premise legal translation systems.
Reasoning Mechanism: The study suggests that explicitly training models to output reasoning steps via SFT may be counterproductive, potentially because these models already possess intrinsic reasoning capabilities from pre-training. Instead, RL appears to be a more effective mechanism for internalizing reasoning strategies and optimizing translation quality through reward signals.
Practicality: The findings support the use of RL-augmented small models as a practical alternative to expensive proprietary APIs, provided that the trade-off between training costs and inference savings is favorable.

The authors conclude that their work provides empirical evidence that explicit reasoning integration via RL, rather than simple imitation of reasoning steps, significantly advances domain-specific NMT.

Evaluation Metrics

3. Key Contributions

4. Results

5. Significance and Claims

More like this