Multi-agent System with Individual Optimized Expertise for Retrieval Augmented Generation

ICLR 2026 Conference Submission17946 Authors

Keywords: Retrieval Augmented Generation, RAG, LLMs, Multi-agent system

TL;DR: we propose a novel approach named Multi-agent System with Individual Optimized Expertise MAPS for RAG.

Abstract: This paper studies the problem of retrieval-augmented generation (RAG), which leverages external knowledge to increase the performance of language models. Despite the remarkable progress, current RAG approaches are still far from satisfactory due to inadequate retrieval and potential hallucination. Towards this end, we propose a novel approach named Multi-agent System with Individual Optimized Expertise (MAPS) for RAG. The core idea of our MAPS is to equip three well-designed agents with different specializations using individual optimization corpus. In particular, we first expand ambiguous queries with a query extension agent, and quantify the reward based on the accuracy, which can be utilized to refine our agent for higher expansion efficiency. To enhance the retrieval quality, we include a unanimity voting method to annotate the current query as insufficient or sufficient, and their generative outcomes are utilized as ground truth to supervised fine-tune the judge agent. To further mitigate potential hallucination, an answer agent is optimized with dynamic matching-based rewards with curriculum learning for final outputs. Extensive experiments across multiple benchmark datasets validate the effectiveness of the proposed MAPS in comparison with state-of-the-art approaches.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 17946