English is not all you need: Rewarding better translation to inspire multilingual capability in LLMs

Despite the impressive performance of Large Language Models (LLMs) across a wide range of natural language processing (NLP) tasks in English, their multilingual capabilities remain limited due to the highly imbalanced training corpora. A common solution to this issue is the translate-test paradigm, where non-English queries are first translated into English before being processed by the LLM. This strategy aims to leverage the model's strong English abilities to improve performance on multilingual tasks. However, the quality of translation can become a critical bottleneck, with translation errors propagating to degrade downstream task performance. Furthermore, English may not always be the optimal language for reasoning. Simply translating inputs into English may fail to fully exploit the LLM's inherent multilingual capabilities, particularly for tasks involving cultural context or domain-specific knowledge. To tackle these challenges, we propose RBT (Rewarding Better Translation to Inspire Multilingual Capability in LLMs), a novel framework that trains a task-oriented translator via reinforcement learning, without modifying the parameters of the underlying LLM. The translator is continuously optimized using task accuracy and other multi-dimensional feedback signals, learning to generate intermediate representations that are more conducive to the LLM's understanding and reasoning. By adaptively preserving or introducing multilingual elements in the translation outputs, RBT effectively unlocks the LLM's latent multilingual capabilities. We evaluate RBT across a wide range of multilingual datasets, and the results demonstrate consistent and significant improvements. Our findings highlight the potential of task-driven translation as a general and model-agnostic approach for enhancing multilingual performance in LLMs.