无参考后训练提升开源大模型多语言机器翻译质量
原标题：Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation
AI 摘要
该研究提出了一种针对多语言机器翻译的无参考后训练方法，从监督微调的MiLMMT-46-v0.1模型出发，应用GRPO算法结合两个无参考质量估计模型的平均奖励，并通过语言识别门控，最后对SFT和RL检查点进行线性插值得到MiLMMT-46-v1.0。在46种语言上，该模型持续优于SFT版本，并超越Seed-X、HY-MT2、TranslateGemma等开源基线，在无参考评分上领先Google Translate、Gemini 3 Pro和GPT-5等专有系统。研究还发现，在线策略蒸馏能达到但未超越RL加检查点插值的质量前沿，并已发布模型和代码。
正文节选
Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation Abstract Open multilingual translation models are improved via group relative policy optimization with reference-free quality rewards and checkpoint interpolation, surpassing strong open and proprietary baselines. We study reference-free post-training for multilingual machine translation with open large language models. Starting from the supervised-finetuned MiLMMT-46-v0.1 models, we apply Group Relat