Ex-GraphRAG: Interpretable Evidence Routing for Graph-Augmented LLMs

TL;DR

Ex-GraphRAG introduces an interpretable encoder for graph-augmented language models that precisely attributes influence to individual nodes, revealing semantic-structural mismatches affecting model performance.

Contribution

It replaces the traditional GNN encoder with M-GNAN, enabling exact attribution of encoder outputs to nodes and feature groups, facilitating faithful auditing.

Findings

Auditable encoder matches black-box performance on STaRK-Prime.

Discovered semantic-structural mismatch in evidence routing.

Removing low-attribution intermediaries degrades multi-hop QA by up to 28%.

Abstract

GraphRAG conditions language models on subgraphs retrieved from knowledge graphs, encoded via message-passing GNNs. Because these encoders entangle node contributions through iterated neighborhood aggregation, there is no closed-form way to determine how much each retrieved entity influenced the encoder's output, and therefore no way to faithfully audit what structural evidence actually reached the model. We introduce Ex-GraphRAG, which replaces the GNN encoder with a Multivariate Graph Neural Additive Network (M-GNAN), an extension of additive graph models to high-dimensional embedding spaces that yields an exact decomposition of the encoder's output across individual nodes and feature groups, without post-hoc approximation. On STaRK-Prime, this auditable encoder matches black-box performance. Using it to audit evidence routing, we uncover a semantic-structural mismatch: the nodes that dominate the encoder's output are structurally disconnected in the retrieved subgraph, held together by low-attribution intermediaries whose removal degrades multi-hop QA by up to 28%. This mismatch, invisible to any opaque encoder, reveals that semantic importance and structural connectivity are governed by disjoint sets of nodes, with direct implications for retrieval pruning, context construction, and failure diagnosis in graph-augmented LLMs.