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

GLM-RAG replaces the GNN retriever in graph-based retrieval-augmented generation with a Graph Language Model initialized from T5-large, allowing node labels, relation labels, queries, and graph structure to interact at the token level. Its graph encoder combines language-model self-attention with graph-aware relative positional encoding, then ranks entities in a 2-hop subgraph capped at 600 triplets before mapping them back to source documents. Across HotpotQA, MuSiQue, and 2WikiMultihopQA, training uses 273,830 question-document pairs; GLM-RAG remains competitive in-domain, reaching 46.4 Recall@2 on MuSiQue, while vanilla RAG is generally sufficient for single-hop questions. Its main advantage appears in zero-shot multi-hop transfer: on MultiHopRAG, GLM-RAG achieves 60.0 Recall@2 versus 39.0 for GFM-RAG+, and it reaches 76.9 accuracy on G-Bench Medical and 76.6 on G-Bench Computer Science, surpassing GFM-RAG, G-Reasoner, and the Microsoft GraphRAG comparison. Ablations show scaling benefits from T5-small through T5-large, with the larger encoder reaching 336M parameters, but the semantic gains carry a substantial systems cost: retrieval latency is 692.3 milliseconds versus 18.6 milliseconds for GFM-RAG*. Answers are generated with OpenAI’s gpt-4o-mini, and the results suggest GLM-RAG is best suited to compositional, out-of-domain reasoning where graph traversal and deep text semantics must be combined.