GLM-RAG: Graph Language Models for Graph-Based Retrieval-Augmented Generation

2026-07-30

AI Summary

This work addresses the challenges of multi-hop reasoning and cross-domain generalization in retrieval-augmented generation (RAG) over knowledge graphs, which demand efficient retrievers that jointly leverage graph structure and semantic information. For the first time, we incorporate Graph Language Models (GLMs) into RAG retriever design and systematically compare GLMs against Graph Neural Networks (GNNs) and traditional vector-based retrieval across single-hop and multi-hop tasks. Experimental results demonstrate that a fine-tuned GLM retriever achieves state-of-the-art performance on two multi-hop question answering benchmarks and exhibits significantly stronger cross-domain generalization than baseline methods. While GNNs offer higher graph coverage and vector retrieval excels in single-hop settings, our analysis reveals fundamental trade-offs among graph coverage, semantic understanding, and scalability across these paradigms.

Abstract

Retrieval-augmented generation (RAG) over knowledge graphs requires retrievers that can effectively capture both graph structure and semantic information. Recent approaches have explored graph neural network (GNN)-based retrievers to model graph topology in multi-hop reasoning tasks. In parallel, graph language models (GLMs) have emerged as a promising paradigm that integrates graph reasoning and the semantic capabilities of language models. In this work, we introduce a GLM-based retriever and investigate the comparative strengths of GLM-based, GNN-based, and traditional vector-search-based retrievers in single- and multi-hop RAG settings, and with a particular focus on transferability to unseen domains. Our findings suggest that finetuned GLM retrievers generalize better out of domain, achieving SOTA on two multi-hop benchmarks. On in-domain multi-hop QA datasets they remain comparable to prior work, with promising scaling as parameters and subgraph coverage increase. GNN-based retrievers achieve higher graph coverage with an efficient training setup, whereas the vector-search baseline excels at single-hop datasets.

Problem

• Retrieval-Augmented Generation
• Knowledge Graph
• Graph Structure
• Semantic Information
• Out-of-Domain Generalization

Innovation

• Graph Language Models
• Retrieval-Augmented Generation
• Knowledge Graphs
• Multi-hop Reasoning
• Out-of-Domain Generalization