GLM-Based Retrievers Beat GNN and Vector Search Out of Domain, Hitting SOTA on Two Multi-Hop RAG Benchmarks

arXiv·medium signalXBlueskyLinkedInCopy link

GLM-RAG introduces a graph language model retriever for knowledge-graph RAG and benchmarks it head-to-head against GNN-based and traditional vector-search retrievers in single- and multi-hop settings, with an explicit focus on transfer to unseen domains. Finetuned GLM retrievers generalize best out of domain and achieve state of the art on two multi-hop benchmarks, while remaining merely comparable to prior work on in-domain multi-hop QA, with performance scaling as parameters and subgraph coverage increase. The honest split matters for architecture choices: GNN retrievers achieve higher graph coverage with a cheaper training setup, and plain vector search still wins on single-hop datasets.