CIIR@LiveRAG 2025: Optimizing Multi-Agent Retrieval Augmented Generation through Self-Training

2025-06-12

AI Summary

To address inefficient agent collaboration and poor adaptability to complex real-world tasks in multi-agent RAG, this paper proposes mRAG: a fine-grained, role-specialized architecture comprising Planning, Retrieval, Reasoning, and Coordination agents—the first such design tailored for RAG. It introduces a reward modeling and reinforcement learning–driven trajectory sampling mechanism to enable end-to-end learnable, self-optimized collaborative policy training. The method integrates modular prompt orchestration with multi-agent cooperative reasoning, significantly enhancing response quality and robustness. Evaluated on the DataMorgana benchmark from the SIGIR 2025 LiveRAG competition, mRAG substantially outperforms conventional single-agent RAG baselines, demonstrating strong effectiveness and generalization capability in realistic, complex question-answering scenarios.

Technology Category

• Multiagent Systems: Multiagent Learning
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Intelligent Robots: Multi-Robot Systems

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

This paper presents mRAG, a multi-agent retrieval-augmented generation (RAG) framework composed of specialized agents for subtasks such as planning, searching, reasoning, and coordination. Our system uses a self-training paradigm with reward-guided trajectory sampling to optimize inter-agent collaboration and enhance response generation. Evaluated on DataMorgana-derived datasets during the SIGIR 2025 LiveRAG competition, mRAG outperforms conventional RAG baselines. We further analyze competition outcomes and showcase the framework's strengths with case studies, demonstrating its efficacy for complex, real-world RAG tasks.

Problem

• Optimizing multi-agent RAG framework for subtasks
• Enhancing response generation via self-training paradigm
• Improving performance on complex real-world RAG tasks

Innovation

• Multi-agent RAG framework for subtasks
• Self-training with reward-guided sampling
• Optimized inter-agent collaboration for responses