Title: Enhancing Graph Foundation Models Through Hyperbolic Retrieval-Augmented Generation

Graph foundation models (GFMs) have become a leading approach in graph representation learning, relying on large-scale pre-training to facilitate inference across diverse domains. However, the parametric knowledge stored within these models often falls short when facing distribution shifts, which restricts their overall generalization potential. To overcome this bottleneck, retrieval-augmented generation (RAG) has been adopted to inject external knowledge during the inference phase. Despite this advancement, current RAG systems confined to Euclidean space encounter a critical geometric constraint: the polynomial volume expansion of Euclidean geometry is poorly suited to the tree-like structure of external knowledge bases. This structural misalignment causes a degradation in semantic granularity during retrieval and triggers the hubness phenomenon.

To resolve these challenges, we introduce Hyperbolic Retrieval-Augmented Generation (HyRAG), a framework aimed at bolstering the generalization performance of GFMs. The proposed Hyperbolic Knowledge Indexing module preserves the hierarchical, tree-like nature of external knowledge by representing it within hyperbolic space. Subsequently, the Multi-granularity Retrieval module equips GFMs with both global semantic anchors and local semantic details by executing coarse-grained and fine-grained knowledge retrieval, respectively. Finally, the Dual-path Fusion module ensures effective knowledge integration for graph tasks, operating at both the structural and feature levels. Empirical evaluations across various graph benchmarks reveal substantial gains in zero-shot scenarios, underscoring the method’s capacity to support robust GFM inference.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC