Adaptive multilingual query rewriting improves reasoning task results
Rewrite What Matters: Adaptive Multilingual Query Rewriting for Reasoning via Agentic Reinforcement Learning
Computation and Language
Summary
When people ask questions in different languages, even if the meaning is the same, computers might think differently and make more mistakes. The authors created a new system that changes questions step-by-step in smart ways depending on the language and task. This helps computers reason better across many languages and types of questions. Their system works well with different large AI models and adapts its approach based on the question’s details.
What this means in practice
- •For multilingual chatbot developers: Improve chatbot understanding by automatically rewriting queries to suit different languages and reasoning tasks.
- •For cross-lingual search engine teams: Enhance search accuracy by adaptively rewriting user queries in multiple languages for better reasoning over documents.
Authors
Rui Qi, Yufeng Chen, Yunlong Liang, Chuan Meng, Sijin Lu, Ge Shi, Jinan Xu, Fandong Meng, Kaiyu Huang
Abstract
In multilingual scenarios, queries with equivalent semantics but in different languages could guide the model into different reasoning trajectories, leading to performance disparities. To mitigate this gap, previous studies typically apply a one-size-fits-all query rewriting strategy, such as translation, which overlooks the fact that different scenarios require diverse types of semantic transformations. In this paper, we propose mRewriter-R1, an agentic multilingual query rewriting framework with reinforcement learning. Unlike single-turn rewriting, mRewriter-R1 formulates multilingual query rewriting as a multi-turn sequential decision-making process, where the model dynamically performs multi-aspect optimization through adaptive operator selection. Experimental results demonstrate that mRewriter-R1 outperforms all strong multilingual rewriting baselines on different large reasoning backbones. Further analyses show that the learned policy can adaptively decide on rewriting operators according to query characteristics, exhibiting strong generalization ability across diverse reasoning tasks, and plug-and-play compatibility with heterogeneous reasoning language models.