Rewrite to Translate, Translate to Reward: Reinforcement Learning for Source Rewriting in Machine Translation

Although directly prompting off-the-shelf Large Language Models (LLMs) to generate meaning-preserving source rewrites can effectively enhance Machine Translation (MT) quality, doing so requires manually tuning prompts for different MT models. In this work, we propose RLSR (Reinforcement Learning for Source Rewriting), a novel RL-based framework for training a source rewriting model without tuning prompts for each MT model. RLSR optimizes the rewriting model by directly using the improvement in downstream translation quality yielded by each rewritten source as the reward. Extensive experiments across six MT models and 16 language pairs demonstrate that our 4B rewriting models trained via RLSR significantly outperform the no-rewriting baseline and existing same-scale prompt-based rewriting baselines, while achieving competitive performance against prompt-based baselines based on the 235B LLM.