CoSelector: Multi-Agent Collaboration for Multilingual Retrieval-Augmented Generation
Abstract
AbstractGeneral Retrieval-Augmented Generation (RAG) systems would suffer from performance degradation in multilingual scenarios due to the complexity of integrating different retrieval results from multilingual collections. In this paper, we propose CoSelector, a multi-agent collaborative framework for multilingual retrieval selection, which is optimized by the centralized training with decentralized execution. Under this paradigm, agents share the same goal but execute decisions using only local observations. CoSelector consists of a Selector agent and a Generator agent. The Selector dynamically filters and composes the most pertinent multilingual documents into a coherent context, while the Generator processes this context to understand and integrate information for answer generation. Across the 15 languages evaluated on MKQA and XOR-TyDi QA, CoSelector achieves competitive improvement in response language consistency, demonstrating its effectiveness in both linguistic alignment and the selection of retrieved results.