Retrieving Minimal and Sufficient Reasoning Subgraphs with Graph Foundation Models for Path-aware GraphRAG

A team of researchers has introduced GFM-Retriever, a novel approach to Graph-based Retrieval-Augmented Generation (GraphRAG) that fundamentally rethinks how AI systems retrieve and reason with structured knowledge. The core innovation is treating a pre-trained Graph Foundation Model (GFM) not just as a ranking tool, but as a generalized, cross-domain retriever. This directly addresses a major weakness in existing methods: their failure in 'cold-start' scenarios where target domain data is scarce, which often leads to incomplete or redundant reasoning contexts. By responding to user queries with a precisely selected subgraph, the system moves beyond treating graphs as mere intermediate artifacts.

Building on this retrieval, the method employs a principled Information Bottleneck objective to act as a label-free subgraph selector. This identifies a query-conditioned 'core set'—a subgraph that is both informationally sufficient and structurally minimal, containing the golden evidence needed for reasoning. To bridge the gap between structured graph data and text generation, the system explicitly extracts and reorganizes the relational paths within this subgraph into in-context prompts. This path-aware structuring makes the model's reasoning process more interpretable.

Extensive experiments on multi-hop question answering benchmarks demonstrate that GFM-Retriever achieves state-of-the-art performance. It excels in both the quality of the retrieved evidence and the accuracy of the final generated answers, all while maintaining computational efficiency. The work, detailed in the arXiv preprint 'Retrieving Minimal and Sufficient Reasoning Subgraphs with Graph Foundation Models for Path-aware GraphRAG,' represents a significant step toward more robust and explainable knowledge-intensive AI systems.