Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Language-Coupled Reinforcement Learning for Multilingual Retrieval-Augmented Generation

Rui Qi, Fengran Mo, Yufeng Chen, Xue Zhang, Shuo Wang, Hongliang Li, Jinan Xu, Meng Jiang, Jian-Yun Nie, Kaiyu Huang

TL;DR

This work tackles multilingual retrieval-augmented generation (MRAG) by addressing knowledge bias and knowledge conflicts that arise when semantically equivalent queries across languages interact with multilingual collections. It introduces Language-Coupled Reinforcement Learning (LcRL), which integrates a language-coupled rollout with Group Relative Policy Optimization (GRPO) and a language-aware reward model that includes an auxiliary anti-consistency penalty to stabilize training. Key innovations include multilingual group sampling in the rollout, a dense multilingual outcome reward based on Character 3-gram Recall, and a cross-language penalty to prevent collapse to incorrect reasoning paths. Empirical results on MKQA and XOR-TyDi QA demonstrate competitive performance and strong generalization to unseen languages, with robustness under limited training data and across large language counts; code is available at the project repository.

Abstract

Multilingual retrieval-augmented generation (MRAG) requires models to effectively acquire and integrate beneficial external knowledge from multilingual collections. However, most existing studies employ a unitive process where queries of equivalent semantics across different languages are processed through a single-turn retrieval and subsequent optimization. Such a ``one-size-fits-all'' strategy is often suboptimal in multilingual settings, as the models occur to knowledge bias and conflict during the interaction with the search engine. To alleviate the issues, we propose LcRL, a multilingual search-augmented reinforcement learning framework that integrates a language-coupled Group Relative Policy Optimization into the policy and reward models. We adopt the language-coupled group sampling in the rollout module to reduce knowledge bias, and regularize an auxiliary anti-consistency penalty in the reward models to mitigate the knowledge conflict. Experimental results demonstrate that LcRL not only achieves competitive performance but is also appropriate for various practical scenarios such as constrained training data and retrieval over collections encompassing a large number of languages. Our code is available at https://github.com/Cherry-qwq/LcRL-Open.

Language-Coupled Reinforcement Learning for Multilingual Retrieval-Augmented Generation

TL;DR

This work tackles multilingual retrieval-augmented generation (MRAG) by addressing knowledge bias and knowledge conflicts that arise when semantically equivalent queries across languages interact with multilingual collections. It introduces Language-Coupled Reinforcement Learning (LcRL), which integrates a language-coupled rollout with Group Relative Policy Optimization (GRPO) and a language-aware reward model that includes an auxiliary anti-consistency penalty to stabilize training. Key innovations include multilingual group sampling in the rollout, a dense multilingual outcome reward based on Character 3-gram Recall, and a cross-language penalty to prevent collapse to incorrect reasoning paths. Empirical results on MKQA and XOR-TyDi QA demonstrate competitive performance and strong generalization to unseen languages, with robustness under limited training data and across large language counts; code is available at the project repository.

Abstract

Multilingual retrieval-augmented generation (MRAG) requires models to effectively acquire and integrate beneficial external knowledge from multilingual collections. However, most existing studies employ a unitive process where queries of equivalent semantics across different languages are processed through a single-turn retrieval and subsequent optimization. Such a ``one-size-fits-all'' strategy is often suboptimal in multilingual settings, as the models occur to knowledge bias and conflict during the interaction with the search engine. To alleviate the issues, we propose LcRL, a multilingual search-augmented reinforcement learning framework that integrates a language-coupled Group Relative Policy Optimization into the policy and reward models. We adopt the language-coupled group sampling in the rollout module to reduce knowledge bias, and regularize an auxiliary anti-consistency penalty in the reward models to mitigate the knowledge conflict. Experimental results demonstrate that LcRL not only achieves competitive performance but is also appropriate for various practical scenarios such as constrained training data and retrieval over collections encompassing a large number of languages. Our code is available at https://github.com/Cherry-qwq/LcRL-Open.
Paper Structure (42 sections, 9 equations, 10 figures, 7 tables, 1 algorithm)

This paper contains 42 sections, 9 equations, 10 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multilingual RAG.
  4. Reinforcement Learning for RAG.
  5. Method
  6. Response Generation with Multi-Turn and Multilingual Retrieval
  7. Language-Coupled GRPO with the Search Engine Calling
  8. Multilingual Group Sampling.
  9. Objective Function.
  10. Language-Coupled Reward Modeling
  11. Multilingual Outcome Reward.
  12. Auxiliary Anti-Consistency Penalty.
  13. Reward Fusion and Normalization.
  14. Experimental Setup
  15. Datasets
  16. ...and 27 more sections

Figures (10)

  • Figure 1: Illustration of knowledge bias and knowledge conflict challenges in the MRAG scenario.
  • Figure 2: Illustration of our LcRL by integrating two modules (rollout and reward) within the GRPO for language-coupled (Lc) purposes. The Lc rollout is designed with sampling in multiple languages for multilingual scenarios. The Lc reward is incorporated with an anti-consistency mechanism to mitigate training collapse.
  • Figure 3: Performance with different retrieval options. Ul denotes the language of the user query, and High represents a randomly selected high-resource language.
  • Figure 4: Results of post-training methods with scaling utilized data.
  • Figure 5: Performance of post-training methods with different language coverages.
  • ...and 5 more figures