SIRAG: Towards Stable and Interpretable RAG with A Process-Supervised Multi-Agent Framework

Abstract

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge retrieval, yet its overall reliability still hinges on how effectively the retriever and generator collaborate. Because these components are usually optimized separately, the retrieved information is often redundant or weakly relevant, which limits the quality and stability of the final output. To address this coordination issue, we present SIRAG, a process-supervised multi-agent framework designed to regulate retrieval and generation through lightweight cooperative agents. The framework introduces (1) a Decision Maker that adaptively determines whether to continue retrieving or to initiate answer generation, and (2) a Knowledge Selector that refines the retrieved set into concise and informative evidence. To enable fine-grained credit assignment, we incorporate an LLM-based process evaluator that assesses each intermediate decision rather than relying solely on the final accuracy. Both agents are trained end-to-end with Proximal Policy Optimization (PPO) under a tree-structured exploration scheme, ensuring stable policy learning and interpretable reasoning trajectories. Extensive experiments on single-hop and multi-hop QA benchmarks demonstrate that SIRAG achieves higher accuracy, faster convergence, and more transparent decision processes compared with standard RAG baselines. The proposed framework is modular and can be integrated into existing RAG pipelines without retraining the retriever or generator.

Ähnliche Arbeiten