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

Researchers introduce RLSR, a reinforcement learning framework that trains smaller language models to rewrite source text for improved machine translation without manual prompt tuning. The approach achieves competitive performance with larger models across six MT systems and 16 language pairs, demonstrating that RL-optimized 4B parameter models can match capabilities of 235B parameter prompt-based systems.

The paper addresses a critical inefficiency in modern machine translation: the manual labor required to optimize prompts for different MT models when using LLMs for source rewriting. Rather than treating source rewriting as a static preprocessing task, the researchers framed it as a reinforcement learning problem where the reward signal comes directly from downstream translation quality improvements. This approach eliminates the need for hand-crafted prompts while enabling the use of much smaller, more efficient models.

Machine translation has long struggled with source ambiguity and linguistic phenomena that confuse MT systems. Prior work showed that rewriting source text to be clearer improves translation outputs, but scaling this technique across diverse MT models required extensive manual prompt engineering. RLSR automates this optimization by training a 4B parameter rewriting model that learns to generate rewrites specifically designed to help downstream translation systems, not necessarily to satisfy human readability preferences.

The implications extend beyond translation quality improvements. The research demonstrates that properly optimized smaller models can achieve competitive performance with massive foundation models in specialized tasks. For developers and translation service providers, RLSR reduces infrastructure costs while improving translation consistency across multiple language pairs and MT architectures. This has direct commercial value for localization services, content platforms, and international communication tools.

The competitive performance against 235B parameter prompt-based systems suggests a broader trend: task-specific reinforcement learning can close the capability gap between small and large models for domain-focused applications. Future work may explore whether similar approaches work for other text transformation tasks, from summarization to code generation.

• →RLSR trains smaller 4B parameter models to rewrite source text by optimizing directly for downstream MT quality improvements rather than manual prompt tuning.
• →The framework eliminates the need for per-model prompt engineering while achieving competitive performance with 235B parameter LLMs across 16 language pairs.
• →Extensive experiments across six MT models demonstrate significant improvements over no-rewriting baselines and same-scale prompt-based approaches.
• →Smaller RL-optimized models can match or exceed larger prompt-based systems when trained with task-specific reward signals.
• →The approach reduces computational costs and infrastructure requirements for organizations implementing source rewriting in translation pipelines.