Researchers have developed MiLMMT-46-v1.0, a multilingual machine translation model that applies reference-free post-training to open large language models. The team uses Group Relative Policy Optimization (GRPO) with a reward based on two reference-free quality estimation models, gated by language identification.

  • Starting from supervised-finetuned MiLMMT-46-v0.1 checkpoints, the authors apply reinforcement learning and linearly interpolate the results to create v1.0.
  • The model is evaluated across 46 languages and consistently improves translation quality over its supervised fine-tuning counterpart.
  • It outperforms strong open baselines including Seed-X, HY-MT2, and TranslateGemma.
  • Against proprietary systems like Google Translate, Gemini 3 Pro, and GPT-5, it achieves leading reference-free scores.
  • On-policy distillation was investigated but did not surpass the quality frontier achieved by RL with checkpoint interpolation.

The authors release the models and code to facilitate future research in this area.