MoG: Mixture of Experts for Graph-based Retrieval-Augmented Generation

2026-05-29

AI Summary

Traditional retrieval-augmented generation (RAG) approaches often retrieve irrelevant information from a unified knowledge base, impeding complex reasoning. To address this limitation, this work proposes the Mixture-of-Graphs (MoG) framework, which introduces a mixture-of-experts mechanism into graph-based RAG for the first time. The framework organizes knowledge into a central graph—continuously accessible—and multiple expert graphs that are sparsely activated. A topology-aware routing mechanism dynamically selects relevant expert graphs based on the query context, enabling focused and sparse retrieval within task-specific knowledge subspaces. Evaluated on complex reasoning benchmarks such as MuSiQue, MoG significantly outperforms strong baselines, achieving relative performance improvements exceeding 20%.

Abstract

Retrieval-augmented generation is intensively studied to ground large language models on external evidence. However, retrieving from a unified knowledge base could inevitably introduce irrelevant information that may mislead generation for complex reasoning. Inspired by the conditional computation of mixture of experts (MoE), where a router sparsely selects specialized experts alongside shared ones for each input, we propose \textbf{M}ixture \textbf{o}f experts for \textbf{G}raph-based Retrieval-Augmented Generation, i.e., \textbf{MoG}. It organizes knowledge into two core components: (i) diverse, always-accessible hub graphs that encode semantically and structurally central knowledge and provide contextual clues for expert activation, and (ii) sparsely activated expert graphs that contain domain-specific evidence. MoG first accesses hub graphs to identify general evidence and derive contextual clues. Then, a topology-aware router dynamically activates a limited set of expert graphs conditioned on the query, thereby confining retrieval to a focused evidence subspace. Extensive experiments on challenging benchmarks show that MoG consistently outperforms strong baselines, with over 20\% relative improvement on MuSiQue. Our code is available in https://github.com/DEEP-PolyU/MoG.

Problem

• Retrieval-Augmented Generation
• irrelevant information
• complex reasoning
• knowledge retrieval
• large language models

Innovation

• Mixture of Experts
• Graph-based Retrieval
• Retrieval-Augmented Generation
• Topology-aware Routing
• Sparse Activation