iCoSys is proud to announce the acceptance of the research paper “The Price of Reasoning: Cost-Quality Tradeoffs in Reinforcement Learning for Neural Machine Translation” at the WoLDE 2026 workshop, part of ACM DocEng’26.
Led by Aixiu An (iCoSys) and Michael Jungo (University of Fribourg), the research investigates the balance between performance and deployment costs when using Reinforcement Learning with Verifiable Rewards (RLVR) for Neural Machine Translation.

Key achievements

Alongside the publication, we are releasing two open-weight models fine-tuned specifically for Swiss legal translation:

• iCoSys-HEIA/Qwen3.5-4b-Swiss-Legal-Translation
• iCoSys-HEIA/Qwen3.5-9b-Swiss-Legal-Translation

Both models were post-trained using GRPO on SwiLTra-Bench, keeping reasoning traces in the training loop. These lightweight models achieve near-frontier quality at a fraction of the deployment cost of larger frontier models, offering an efficient, high-quality solution for Swiss LegalTech applications.

Links

Read the paper on arXiv
Access the models on Hugging Face: iCoSys-HEIA