Researchers have built a retriever that reads knowledge graphs like a language model reads text - and it generalizes better to new domains than vector search or graph neural networks, at least on multi-hop questions.
The new approach, called GLM-RAG, swaps a traditional GNN-based retriever for a graph language model (GLM) that processes graph structure and semantic text together. The researchers tested it against GNN-based and vector-search retrievers across single-hop and multi-hop retrieval-augmented generation tasks. On two multi-hop benchmarks built from domains the model had not seen during training, the finetuned GLM retriever set a new state of the art. On in-domain multi-hop tests it only matched prior methods, though the paper notes results should improve further as parameter count and subgraph coverage scale up.
Most RAG benchmarks test models on data similar to what they trained on, which hides how badly retrieval degrades once a system meets an unfamiliar knowledge base. This result matters because it isolates that failure mode and shows one retriever architecture holding up better than the alternatives specifically when the domain shifts, not just when the questions get harder.
Vector search still wins on simple, single-hop lookups, and GNN retrievers remain the more efficient choice for covering a graph during training - so this is not a wholesale replacement, just a result worth remembering the next time someone claims their retriever generalizes.