Ex-GraphRAG Enables Auditable Evidence Routing in Graph-Augmented LLMs
Ex-GraphRAG replaces GNN encoders with a Multivariate Graph Neural Additive Network (M-GNAN) to provide exact decompositions of encoder outputs across nodes and feature groups. This allows faithful auditing of evidence routing in graph-augmented language models. On the STaRK-Prime benchmark, the auditable encoder matches black-box performance. The method reveals a semantic-structural mismatch where nodes dominating the encoder output are structurally disconnected from the query's semantic focus.
Key facts
- Ex-GraphRAG uses M-GNAN, an extension of additive graph models to high-dimensional embeddings.
- It yields exact output decomposition without post-hoc approximation.
- On STaRK-Prime, it matches black-box performance.
- It uncovers a semantic-structural mismatch in evidence routing.
- The approach enables auditing of which entities influence the encoder output.
- GraphRAG conditions language models on subgraphs from knowledge graphs.
- Standard GNN encoders entangle node contributions through neighborhood aggregation.
- There is no closed-form way to determine influence of each entity in standard GraphRAG.
Entities
Institutions
- arXiv