Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective
Reinforcement learning fine-tuning has concentrated on decoder-only LLMs, leaving production encoder-decoder translation models largely unexplored. This work applies Group Relative Policy Optimization to Meta's NLLB-200 across 13 languages using reference-free rewards (LaBSE and COMET-Kiwi), eliminating the need for parallel data at fine-tuning time. Results show consistent gains up to 5.03 chrF++ on Traditional Chinese, matching supervised fine-tuning on morphologically complex languages without target-language data. The finding reshapes how practitioners can optimize deployed translation systems with minimal resource overhead.
Modelwire context
ExplainerThe practical implication buried in the results is that GRPO, a technique developed in the decoder-only LLM context, transfers to seq2seq architectures with minimal modification, which means the tooling and intuitions practitioners have built around RL fine-tuning for chat models are more portable than the field has assumed.
The multilingual angle here connects directly to the 'From Flat Language Labels to Typological Priors' coverage from the same day, which tackled a related problem: how to make translation systems work better across language families without proportionally scaling data requirements. Both papers are circling the same constraint, that parallel data is expensive and unevenly distributed, but from opposite ends of the stack. The S2ST-Omni 2 work attacks it through linguistic structure at conditioning time; this work attacks it by removing the reference requirement at training time entirely. Together they sketch a direction where low-resource translation improves without the traditional data collection bottleneck.
Watch whether Meta or any third party publishes GRPO fine-tuning results on NLLB-200 for languages below 1 million speakers in the next six months. If gains hold at that data tier, the reference-free framing becomes a genuine low-resource story rather than a compute efficiency story for already-resourced language pairs.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsNLLB-200 · Meta · Group Relative Policy Optimization · LaBSE · COMET-Kiwi
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.