Anchored SubgraphRAG: Structurally Gated Message Passing for Multi-Hop KG-RAG Retrieval

Answering multi-hop questions over knowledge graphs requires retrieving structurally coherent evidence, yet most Graph-RAG retrievers score triples in isolation or rely on semantic relevance alone. Adding a Graph Neural Network (GNN) to aggregate neighbourhood information is a natural fix, but recent work reports that doing so can actually degrade retrieval. We argue that this failure is not caused by message passing itself but by uncontrolled semantic aggregation that washes out the contrasts needed to identify relevant triples, an over-smoothing-like effect. We introduce Anchored SubgraphRAG, a lightweight retriever that uses query-dependent anchors to localise the relevant graph region and gates message passing with structural positional information. On WebQSP, Anchored SubgraphRAG improves triple recall@100 from $0.883$ to $0.905 \pm 0.0095$ over the SubgraphRAG baseline. Experiments show that both the GNN and the structural controls are required for this gain, identifying structural gating as a design principle for Graph-RAG retrieval.