Ex-GraphRAG: Interpretable Evidence Routing for Graph-Augmented LLMs
Researchers have developed Ex-GraphRAG, a novel method for interpreting how Large Language Models (LLMs) use information from knowledge graphs. This new approach replaces the standard Graph Neural Network encoder with a Multivariate Graph Neural Additive Network, allowing for an exact decomposition of the model's output across individual nodes and features. Auditing evidence routing with Ex-GraphRAG revealed a disconnect between semantic importance and structural connectivity in retrieved subgraphs, indicating that nodes dominating the model's output are often structurally disconnected within the graph. AI
IMPACT Provides a new auditable method for understanding how LLMs process graph-augmented information, aiding in debugging and improving retrieval strategies.