pith. sign in

arxiv: 2605.21994 · v1 · pith:ARXVBVSCnew · submitted 2026-05-21 · 💻 cs.LG · cs.AI

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

Yoav Kor Sade , Arvindh Arun , Rishi Puri , Steffen Staab , Maya Bechler-Speicher This is my paper

Pith reviewed 2026-05-22 07:05 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords GraphRAGinterpretable attributiongraph neural networksevidence routingmulti-hop QAknowledge graphsadditive modelsGNN interpretability
0
0 comments X

The pith

Replacing the GNN encoder in GraphRAG with an additive network yields exact per-node attributions that expose a mismatch between semantically dominant nodes and structural connectivity in retrieved subgraphs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper replaces standard message-passing GNN encoders with a Multivariate Graph Neural Additive Network that decomposes the encoder output into precise contributions from each individual node and feature group. This decomposition matches black-box performance on STaRK-Prime while enabling direct audits of which retrieved entities actually influence the language model. The audit reveals that nodes with the highest attribution are typically disconnected in the subgraph and depend on low-attribution intermediaries. Removing those intermediaries reduces multi-hop QA accuracy by up to 28 percent, showing that semantic importance and structural connectivity are controlled by disjoint sets of nodes.

Core claim

Substituting the opaque GNN with an M-GNAN produces a closed-form additive decomposition of the encoder output across nodes and feature groups without approximation. When this auditable encoder is used to inspect evidence routing on STaRK-Prime, the nodes that dominate the output turn out to be structurally disconnected within the retrieved subgraph, linked only through low-attribution bridge nodes. Removing those bridges degrades downstream multi-hop question answering by as much as 28 percent, demonstrating that semantic and structural signals are governed by separate node sets.

What carries the argument

The Multivariate Graph Neural Additive Network (M-GNAN), an extension of additive graph models to high-dimensional embeddings that supplies an exact, non-approximate decomposition of the encoder output over individual nodes and feature groups.

If this is right

  • Retrieval pruning can safely drop low-attribution intermediaries while preserving answer quality.
  • Context construction for LLMs should rank retrieved nodes by attribution rather than by graph distance alone.
  • Failure diagnosis can now isolate which parts of a subgraph actually reached the model.
  • Semantic importance and structural connectivity can be optimized separately in future graph-augmented systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decomposition could be folded into the retrieval stage itself to select subgraphs that maximize high-attribution coverage.
  • Current connectivity-based retrieval heuristics may systematically favor the wrong nodes for multi-hop reasoning tasks.
  • Repeating the audit across additional knowledge graphs would test whether the semantic-structural disconnect is a general phenomenon.

Load-bearing premise

The M-GNAN produces an exact decomposition of the encoder output into additive node and feature contributions without relying on post-hoc approximations.

What would settle it

Ablating the high-attribution nodes identified by M-GNAN and measuring whether the performance drop is larger than for random or structurally central nodes would confirm or refute the claimed mismatch and its impact on QA accuracy.

Figures

Figures reproduced from arXiv: 2605.21994 by Arvindh Arun, Maya Bechler-Speicher, Rishi Puri, Steffen Staab, Yoav Kor Sade.

Figure 1
Figure 1. Figure 1: Node importance scores for a representative query. M-GNAN importance scores for the query “Could you list the tablet or capsule medications that act on the MAPK1 gene/protein?” Three nodes (MAPK1, MAPK14, MAP3K4) account for approximately 60% of total importance; the remaining 180 nodes each contribute less than 0.05. Yet removing these low-scoring nodes causes 28% performance degradation in multi-hop sett… view at source ↗
Figure 2
Figure 2. Figure 2: Full subgraph vs. induced subgraph on high-importance nodes. Left: the full retrieved subgraph for the MAPK1 query, colored by M-GNAN importance (dark red = high, yellow = low). High-importance nodes sit within a dense network of low-importance intermediaries. Right: the induced subgraph on the top-10 important nodes and their im￾mediate neighbors. Removing bridge nodes fragments the graph into disconnecte… view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Ex-GraphRAG, which replaces the standard message-passing GNN encoder in GraphRAG with a Multivariate Graph Neural Additive Network (M-GNAN). This architecture is presented as an extension of additive graph models to high-dimensional embeddings that yields an exact (non-approximate) decomposition of the encoder output into per-node and per-feature contributions. On the STaRK-Prime benchmark the auditable encoder matches black-box performance; using the resulting attributions the authors identify a semantic-structural mismatch in which high-attribution nodes are structurally disconnected within the retrieved subgraph and are bridged only by low-attribution intermediaries whose removal degrades multi-hop QA performance by up to 28%.

Significance. If the exact-decomposition property is rigorously established, the work supplies a practical tool for faithful auditing of evidence routing in graph-augmented LLMs. The reported semantic-structural mismatch, if reproducible, would be a substantive finding with direct consequences for retrieval pruning, context construction, and failure analysis. Performance parity on STaRK-Prime is a necessary but not sufficient indicator that the interpretability benefit does not come at the cost of utility.

major comments (2)
  1. [§3] §3 (M-GNAN definition) and the accompanying derivation: the central claim that M-GNAN produces an exact, additive decomposition across nodes and feature groups without residual cross-terms or post-hoc approximation is load-bearing for both the 28% degradation result and the mismatch conclusion. The manuscript must supply the explicit construction (including how non-linear message-passing interactions and embedding normalizations are eliminated while preserving expressivity) and a proof or verification that additivity holds after graph embedding.
  2. [§5] §5 (Experiments) and Table reporting the 28% figure: the degradation result is obtained by removing low-attribution intermediaries identified via the M-GNAN attributions. Without the precise definition of the removal protocol, the subgraph variants tested, error bars, and statistical significance, it is impossible to determine whether the observed drop reflects a genuine structural mismatch or an artifact of the particular attribution threshold or subgraph construction.
minor comments (2)
  1. [Abstract] Abstract: the sentence describing the mismatch is information-dense; a brief parenthetical gloss on how 'structurally disconnected' is quantified would improve immediate readability.
  2. [Notation] Notation: ensure that the symbols for node attributions, feature-group contributions, and the final encoder output are introduced once and used consistently in all equations and figures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below, clarifying the technical foundations of M-GNAN and strengthening the experimental reporting. We will incorporate the requested details in the revised manuscript.

read point-by-point responses

  1. Referee: [§3] §3 (M-GNAN definition) and the accompanying derivation: the central claim that M-GNAN produces an exact, additive decomposition across nodes and feature groups without residual cross-terms or post-hoc approximation is load-bearing for both the 28% degradation result and the mismatch conclusion. The manuscript must supply the explicit construction (including how non-linear message-passing interactions and embedding normalizations are eliminated while preserving expressivity) and a proof or verification that additivity holds after graph embedding.

    Authors: We agree that the exact additivity property is foundational. Section 3 presents M-GNAN as a direct extension of additive graph models to multivariate embeddings, where the output is constructed as a sum of per-node, per-feature contributions by replacing iterative message-passing with an additive aggregation that factors out cross-terms at the embedding level. Non-linear interactions are eliminated by design through the additive structure while retaining expressivity via high-dimensional feature groups. We will expand §3 with the full explicit construction, the step-by-step elimination of normalization-induced residuals, and a formal proof (or empirical verification on synthetic graphs) that the decomposition remains exact after embedding. This revision will make the load-bearing claim fully rigorous. revision: yes

  2. Referee: [§5] §5 (Experiments) and Table reporting the 28% figure: the degradation result is obtained by removing low-attribution intermediaries identified via the M-GNAN attributions. Without the precise definition of the removal protocol, the subgraph variants tested, error bars, and statistical significance, it is impossible to determine whether the observed drop reflects a genuine structural mismatch or an artifact of the particular attribution threshold or subgraph construction.

    Authors: We acknowledge the need for greater experimental transparency. The 28% degradation is measured by ablating nodes whose M-GNAN attribution falls below a percentile threshold while preserving the original subgraph connectivity for the remaining nodes. In the revision we will: (i) state the exact attribution threshold and percentile used, (ii) enumerate the subgraph variants (full subgraph, high-attribution core only, and intermediary-removed variants), (iii) report standard error bars across five random seeds, and (iv) include paired t-test p-values confirming statistical significance of the performance drop. These additions will allow readers to reproduce and evaluate whether the drop is attributable to the identified semantic-structural mismatch. revision: yes

Circularity Check

0 steps flagged

M-GNAN exact decomposition presented as independent architectural extension with no reduction to fitted inputs or self-citations

full rationale

The provided abstract and context introduce M-GNAN as an extension of additive graph models to high-dimensional embeddings that yields exact decomposition without post-hoc approximation. No equations, derivations, or self-citations are shown that reduce this claim to a fitted parameter defined by the same data or to a prior result by the same authors. The central claim of faithful attribution enabling the semantic-structural mismatch finding is presented as a property of the new architecture rather than derived circularly from the target results. Per the rules, this is a self-contained architectural contribution with no load-bearing circular steps identifiable from the given text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities beyond the introduction of M-GNAN are stated. The central claim rests on the unshown mathematical property that M-GNAN provides exact additive decomposition in embedding space.

invented entities (1)
  • M-GNAN no independent evidence
    purpose: Extension of additive graph models to high-dimensional embeddings for exact output decomposition
    Introduced in abstract as the replacement encoder enabling faithful auditing without post-hoc approximation.

pith-pipeline@v0.9.0 · 5761 in / 1310 out tokens · 54622 ms · 2026-05-22T07:05:45.573964+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 2 internal anchors

  1. [1]

    and Laurent, Thomas and LeCun, Yann and Bresson, Xavier and Hooi, Bryan , title =

    He, Xiaoxin and Tian, Yijun and Sun, Yifei and Chawla, Nitesh V. and Laurent, Thomas and LeCun, Yann and Bresson, Xavier and Hooi, Bryan , title =. Proceedings of the 38th International Conference on Neural Information Processing Systems , articleno =. 2024 , isbn =

    work page 2024

  2. [2]

    2025 , eprint=

    Retrieval-Augmented Generation with Graphs (GraphRAG) , author=. 2025 , eprint=

    work page 2025

  3. [3]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    The Intelligible and Effective Graph Neural Additive Network , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    work page

  4. [4]

    Ioannidis and Karthik Subbian and James Zou and Jure Leskovec , booktitle=

    Shirley Wu and Shiyu Zhao and Michihiro Yasunaga and Kexin Huang and Kaidi Cao and Qian Huang and Vassilis N. Ioannidis and Karthik Subbian and James Zou and Jure Leskovec , booktitle=. 2024 , url=

    work page 2024

  5. [5]

    2024 , eprint=

    Let Your Graph Do the Talking: Encoding Structured Data for LLMs , author=. 2024 , eprint=

    work page 2024

  6. [6]

    Proceedings of the 33rd International Conference on Neural Information Processing Systems , articleno =

    Ying, Rex and Bourgeois, Dylan and You, Jiaxuan and Zitnik, Marinka and Leskovec, Jure , title =. Proceedings of the 33rd International Conference on Neural Information Processing Systems , articleno =. 2019 , publisher =

    work page 2019

  7. [7]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    work page 2024

  8. [8]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Attention is not Explanation , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=. 2019 , publisher=

    work page 2019

  9. [9]

    International Conference on Learning Representations , year=

    How Powerful are Graph Neural Networks? , author=. International Conference on Learning Representations , year=

    work page

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Parameterized Explainer for Graph Neural Network , author=. Advances in Neural Information Processing Systems , volume=

    work page

  11. [11]

    1990 , publisher=

    Generalized Additive Models , author=. 1990 , publisher=

    work page 1990

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Neural Additive Models: Interpretable Machine Learning with Neural Nets , author=. Advances in Neural Information Processing Systems , volume=

    work page

  13. [13]

    From Local to Global: A Graph RAG Approach to Query-Focused Summarization

    Darren Edge and Ha Trinh and Newman Cheng and Joshua Bradley and Alex Chao and Apurva Mody and Steven Truitt and Jonathan Larson , year=. From Local to Global: A Graph. 2404.16130 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    International Conference on Learning Representations , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations , year=

    work page

  15. [15]

    International Conference on Learning Representations , year=

    Graph Attention Networks , author=. International Conference on Learning Representations , year=

    work page

  16. [16]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume=

    work page

  17. [17]

    Nature Machine Intelligence , volume=

    Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead , author=. Nature Machine Intelligence , volume=. 2019 , doi=

    work page 2019

  18. [18]

    Transactions of the Association for Computational Linguistics , volume=

    Lost in the Middle: How Language Models Use Long Contexts , author=. Transactions of the Association for Computational Linguistics , volume=

    work page

  19. [19]

    Attention is not not Explanation , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=. 2019 , publisher=

    work page 2019

  20. [20]

    International Conference on Learning Representations , year=

    How Attentive are Graph Attention Networks? , author=. International Conference on Learning Representations , year=

    work page

  21. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Explainability in Graph Neural Networks: A Taxonomic Survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    work page

  22. [22]

    Proceedings of the 38th International Conference on Machine Learning , pages=

    On Explainability of Graph Neural Networks via Subgraph Explorations , author=. Proceedings of the 38th International Conference on Machine Learning , pages=. 2021 , volume=

    work page 2021

  23. [23]

    Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

    Intelligible Models for Classification and Regression , author=. Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=. 2012 , publisher=

    work page 2012

  24. [24]

    Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

    Accurate Intelligible Models with Pairwise Interactions , author=. Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=. 2013 , publisher=

    work page 2013

  25. [25]

    Advances in Neural Information Processing Systems , pages=

    A Unified Approach to Interpreting Model Predictions , author=. Advances in Neural Information Processing Systems , pages=

    work page

  26. [26]

    Computational Linguistics , volume=

    Measuring Attribution in Natural Language Generation Models , author=. Computational Linguistics , volume=. 2023 , publisher=

    work page 2023

  27. [27]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Retrieval-Augmented Generation for Large Language Models: A Survey , author=. arXiv preprint arXiv:2312.10997 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Unifying Large Language Models and Knowledge Graphs: A Roadmap , author=. IEEE Transactions on Knowledge and Data Engineering , volume=

    work page

  29. [29]

    2021 , publisher=

    Yasunaga, Michihiro and Ren, Hongyu and Bosselut, Antoine and Liang, Percy and Leskovec, Jure , booktitle=. 2021 , publisher=

    work page 2021

  30. [30]

    2025 , publisher=

    Mavromatis, Costas and Karypis, George , booktitle=. 2025 , publisher=

    work page 2025

  31. [31]

    International Conference on Learning Representations , year=

    Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning , author=. International Conference on Learning Representations , year=

    work page

  32. [32]

    Edward J Hu and yelong shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    work page 2022