Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective
On the map: Deep reinforcement learning methods → RL training of language and agent models
Abstract (arXiv)
Production machine translation relies overwhelmingly on encoder-decoder Seq2Seq models, yet reinforcement learning approaches to MT fine-tuning have largely targeted decoder-only LLMs at ≥7B parameters, with limited systematic study of encoder-decoder architectures. We apply Group Relative Policy Optimization to NLLB-200 (600M and 1.3B) using a hybrid reference-free reward (LaBSE and COMET-Kiwi) that requires no parallel data at fine-tuning time, evaluating across 13 typologically diverse languages. GRPO yields consistent improvements on all 13 languages, up to +5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages . We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest, making this approach most effective precisely where parallel data is scarcest, and replicate this pattern across English and Spanish source languages.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in RL training of language and agent models, one of 15 directions in Deep reinforcement learning methods. Builds on 18 of its 26 references that are in the map. First posted on arXiv 15 May 2026.
Builds on
18 of 26 references.
Closest ancestors · 2
Same direction, newest first
- 2026Improving Low-Resource Machine Translation via Round-Trip Reinforcement Learning18 Jan
- 2025SSR-Zero: Simple Self-Rewarding Reinforcement Learning for Machine Translation22 May
Other references · 16
- 2022Training language models to follow instructions with human feedbackcited by 15,971
- 2017Proximal Policy Optimization Algorithmscited by 15,215
- 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
- 2019BERTScore: Evaluating Text Generation with BERTcited by 5,512
- 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,418
- 2018A Call for Clarity in Reporting BLEU Scorescited by 1,799
- 2022No Language Left Behind: Scaling Human-Centered Machine Translationcited by 1,312
- 2015Sequence Level Training with Recurrent Neural Networkscited by 576
- 2021The FLORES-101 Evaluation Benchmark for Low-Resource and Multilingual Machine Translationcited by 560
- 2017Six Challenges for Neural Machine Translationcited by 335
- 2022CometKiwi: IST-Unbabel 2022 Submission for the Quality Estimation Shared Taskcited by 207
- 2015Minimum Risk Training for Neural Machine Translationcited by 115
4 more
- 2023Scaling up COMETKIWI: Unbabel-IST 2023 Submission for the Quality Estimation Shared Taskcited by 88
- 2023An Open Dataset and Model for Language Identificationcited by 50
- 2017Bandit Structured Prediction for Neural Sequence-to-Sequence Learningcited by 31
- 2026GRRM: Group Relative Reward Modeling for Machine Translationcited by 2
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation
- 2026PEGRL: Improving Machine Translation by Post-Editing Guided Reinforcement Learning
- 2026Supervised Fine-Tuning versus Reinforcement Learning: A Study of Post-Training Methods for Large Language Models
- 2026Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning
- 2026Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax
- 2025MT-R1-Zero: Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning
- 2024Inference-Aware Fine-Tuning for Best-of-N Sampling in Large Language Models
- 2025Improving Low-Resource Translation with Dictionary-Guided Fine-Tuning and RL: A Spanish-to-Wayuunaiki Study