Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

KGRAG-Ex: Explainable Retrieval-Augmented Generation with Knowledge Graph-based Perturbations

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read KGRAG-Ex converts knowledge-graph paths into text for RAG and uses node-, edge-, and sub-path deletions to trace which facts drive the model's answer, cutting explanation cost by roughly two-thirds.

desk verdict Novel graph-level perturbation framework whose factual-grounding claim is untested; the explainability results are self-referential. read the letter →

arxiv 2507.08443 v1 pith:YEYSGSAH submitted 2025-07-11 cs.LG

classification cs.LG
keywords knowledgegraphsretrieval-augmentedgenerationexplainabilityperturbation-basedexplanationsmultiple-choicequestionansweringmedicalQAlargelanguagemodelspseudo-paragraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes KGRAG-Ex, a retrieval-augmented generation system that builds a domain knowledge graph from medical text and uses it twice: first to retrieve structured reasoning paths that are turned into natural-language pseudo-paragraphs, and then to explain the model's answer by deleting individual nodes, edges, or whole sub-paths and watching whether the answer changes. The central claim is that this combination improves factual grounding and yields faithful, fine-grained explanations more cheaply than perturbing raw text, with RAG-Ex as the baseline. If true, high-stakes users could see not just an answer but a traceable reason: the graph element whose removal most often flips the answer, linked back to its source document.

What carries the argument

The load-bearing object is the retrieved shortest path in the knowledge graph, verbalized as pseudo-paragraphs. From that path, the system applies three perturbation levels—node removal, edge removal, and sub-path (triplet) removal—and counts how often each deletion changes the model's selected multiple-choice option. The component that appears in the most answer-changing deletions is presented as the most influential, with source filenames attached for traceability.

What would settle it

Take a set of queries where medical experts pre-identify which entities are genuinely needed to answer, delete a component that experts call irrelevant, and see whether the model's answer flips; any such flip shows the deletion signal is not a faithful measure of contribution. Alternatively, compare KGRAG-Ex's top-ranked component to a variant that deletes the same component but rephrases the remaining text to preserve meaning: if the answer still flips, the change is due to wording, not to the entity's role.

Watch

Extended reading notes

Core claim

The paper's discovery is that reasoning paths in a knowledge graph can serve as both the retrieval unit and the explanation unit for RAG. By extracting entity-relation-entity triples from chunks, connecting query entities by shortest paths, and verbalizing those paths into pseudo-paragraphs, KGRAG-Ex grounds responses in structured, source-annotated evidence. Deleting a node, edge, or triplet and observing an output change then attributes importance to that component; the most influential element is shown to the user. Experiments on MedMCQA and MMLU indicate this yields explanations at roughly a third of the LLM calls of text perturbation, while pointing to diseases, symptoms, and body parts as the most load-bearing node types.

Load-bearing premise

The whole explanation hinges on treating a change in the model's answer after deleting a graph component as proof that this component caused the original answer, with no check against the model's actual internal reasoning or human judgment.

Editorial extensions

If this is right

  • Explanations can be produced with roughly one-third the LLM calls and half the tokens of text-level perturbation alternatives on these benchmarks.
  • Removing a sub-path changes the model's answer more often than removing a single node or edge, suggesting triplet-level structure carries much of the reasoning.
  • Components near the start of the shortest path matter more than later ones, which could guide where to spend explanation effort.
  • Semantic node labels let users see whether the model leans on diseases, symptoms, or body parts, connecting graph structure to domain meaning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The binary flip/no-flip signal discards effect size and direction; measuring confidence changes or newly chosen alternatives could make the explanations more informative.
  • Because the KG is built from a subset of the corpus, the reported centrality findings may shift with graph coverage; scaling up is a direct stress test.
  • The same deletion logic could be run against human-judged relevance to check whether the explanations align with expert intuition, not just with answer flips.
  • The method's efficiency comes from perturbing few structural elements; if paths grow longer, the number of perturbations grows linearly, so cost comparisons may change.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces KGRAG-Ex, a retrieval-augmented generation system that builds a domain-specific knowledge graph from a medical corpus via LLM-based triple extraction, identifies shortest paths between query entities, converts those paths into natural-language pseudo-paragraphs used as RAG context, and then generates explanations by perturbing nodes, edges, and sub-paths of the graph and observing changes in the model's selected multiple-choice option. Experimental sections analyze perturbation sensitivity across datasets, positional effects along retrieved paths, semantic node-label importance, and correlations with graph centrality, along with a cost comparison against the text-level perturbation framework RAG-Ex. The central claim in the abstract is that KGRAG-Ex improves both factual grounding and explainability.

Significance. The proposed idea of leveraging structured KG paths for perturbation-based explanation is timely and plausible, and it offers a concrete efficiency advantage: Table 5 reports roughly a threefold reduction in LLM calls and about 45% token savings compared with RAG-Ex. The pipeline is described in sufficient detail to be testable, and the authors state that code and data are publicly available. However, the headline claim of improved factual grounding is never tested, and the explanation faithfulness rests on an unvalidated deletion-based importance assumption; the small sample sizes and lack of statistical rigor further limit the strength of the empirical conclusions. If the missing accuracy and faithfulness evidence were provided, the work could constitute a useful contribution to RAG explainability evaluation, but as it stands the evidence only supports claims about sensitivity patterns, not about grounding improvement or faithful explanations.

major comments (4)
  1. [Abstract and Section 5.] The central claim that KGRAG-Ex 'improves both factual grounding and explainability' is not supported by the experiments, which contain no answer-accuracy measurement. Section 5 reports perturbation flip counts (Table 4), positional distributions (Figure 3), node-label distributions (Figure 4), centrality correlations (Figures 5-7), and LLM-call/token counts (Table 5), but never compares KGRAG-Ex accuracy against plain text RAG, retrieval without KG pseudo-paragraphs, or no-retrieval prompting. Without such a comparison, it is possible that the KG-derived context does not improve factual grounding at all, or even degrades it; at minimum, the abstract and conclusion should be qualified to claims about explainability and cost, not factual grounding. This is a load-bearing evidential gap because it affects the paper's primary contribution.
  2. [Section 4 and Table 2.] The explanation framework equates importance with the frequency with which deleting a node, edge, or sub-path changes the model's selected option, but this premise is never validated against an external standard. There is no comparison with human rationales, no faithfulness metric such as sufficiency/comprehensiveness, and no counterfactual verification linking the perturbed components to the internal inference process or to the source documents. The binary change/no-change signal also discards information about the magnitude and direction of the effect. Until such validation is provided, the user-facing explanation in Table 2 (e.g., 'The most important condition... is Persistent Pulmonary Hypertension in the Newborn') and the structural analyses in Figures 3-7 should be described as correlational observations about deletion sensitivity rather than faithful accounts of model reasoning.
  3. [Section 5.1, Table 4 and Figures 3-7.] The empirical conclusions are drawn from very small samples (128 examples for MedMCQA and 37 for MMLU) with no error bars, confidence intervals, or significance tests. For instance, Figure 3's claim that perturbations 'closer to the beginning of the path are more likely to cause significant changes' is based on histograms over these small counts in which the top/bottom panel ordering is not clearly labeled, and Figure 5's suggestion that both high- and low-degree nodes matter is asserted from the shapes of frequency distributions with no statistical assessment. The paper should either add significance testing (e.g., permutation tests or bootstrapped confidence intervals for the reported trends) or substantially weaken the wording of the claims about positional, label, and centrality patterns.
  4. [Section 5.2, Table 5.] The efficiency comparison with RAG-Ex depends on free parameters whose values may materially change the result. The paper states that RAG-Ex uses a window size of 5 tokens and that smaller windows 'often had little impact' while larger windows caused 'inconsistent effects', and the document chunk size is another parameter; however, no sensitivity analysis is reported. Without experiments varying these parameters, the claimed 'significant reduction' in LLM calls and tokens should be presented as a point estimate for a particular parameter setting rather than as a general property of the method.
minor comments (6)
  1. [Section 3, Table 1 caption.] The caption contains a duplicated word: 'the retrieved path, and and related sources' should read 'the retrieved path, and related sources'.
  2. [Section 4, Table 2.] In the sub-path perturbation row, the selected option after removal is reported as 'A. an association', which is the same as the original selected option in Table 1. This is confusing because the text above the table says the explanation is based on components that 'cause the most changes in the output', yet this example shows no change; please clarify whether this is intended or a typo.
  3. [Section 5.1, Figure 3.] The figure caption says 'across datasets' and the text below identifies 'Top: MMLU. Bottom: MedMCQA.', but the subplot labels (a), (b), (c) are applied to perturbation types, not datasets, which makes the figure hard to parse; consider labeling each panel with the dataset and perturbation type explicitly.
  4. [References.] Reference [8] (Lewis et al., NeurIPS 2020) and reference [9] are the same paper and are both cited in the reference list; this should be merged or deduplicated.
  5. [Throughout.] The naming is inconsistent between 'RAG-EX' and 'RAG-Ex' (e.g., Table 5 vs. Section 5.2); choose one convention. Also, the paper refers to the system as 'KGRAG-Ex' but sometimes says 'KGRAG-EX'; standardize the name.
  6. [Section 5, first paragraph.] The sentence 'we utilized the StatPearls corpus along with two established benchmarks: MedMCQA and MMLU' is misleading because MedMCQA and MMLU are datasets used as the evaluation, while StatPearls is the corpus used for KG construction; clarify the roles of these resources.

Circularity Check

1 steps flagged · score 6.0 of 10

The user-facing 'most important condition' explanation is, by construction, the perturbation-frequency ranking, so the explainability claim is partially self-definitional; the factual-grounding half is untested but not circular.

  1. self definitional [Section 4, 'Explanations through Knowledge Graph Perturbations' (Table 2 and following paragraph)]
    "The explanation shown to the end user is the most influential graph element, identified as the one that causes the most changes in the output of the model across the different perturbations... 'Removing Persistent Pulmonary Hypertension in the Newborn led to a different answer 2 times, indicating it is a highly influential entity in the reasoning path.'"

    Section 4 defines importance as whether deletion changes the output: 'if removing a component significantly alters the outcome, it suggests that the component plays a critical role.' The end-user explanation then states that the selected entity 'had the biggest impact on the result,' but the selection criterion was precisely the number of deletions that changed the answer. The explanation is therefore the construction restated: 'most important' means 'most frequent flip inducer.' Since no external benchmark (human rationales, faithfulness metrics, counterfactual verification) is used, the explanation cannot confirm any property beyond the perturbation count that generated it.

full rationale

The only place where a result reduces to its own inputs is the explanation mechanism: KGRAG-Ex's 'importance' is defined as perturbation-induced option-change frequency, and that same frequency is converted directly into the user-facing 'most important condition' statement. This is a partial, self-definitional circularity in the explainability claim. The paper's other analyses (positional trends, node-label distributions, degree/betweenness correlations) are correlations between this importance score and independent structural quantities, so they are not circular. The 'improves factual grounding' half of the abstract claim is never tested with accuracy comparisons, but that is a missing-evidence and correctness concern rather than circularity. No load-bearing self-citations or imported uniqueness theorems were found; the KG-construction and efficiency comparisons are independent contributions.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on four unvalidated domain assumptions: the LLM-built graph is accurate, shortest paths are the right context, option flips measure importance, and pseudo-paragraphs preserve information. There are also two hand-picked parameters (RAG-Ex window size, unreported chunk size) that affect the quantitative comparisons. No new theoretical entities are introduced.

free parameters (2)
  • RAG-Ex window size = 5 tokens
    Hand-picked for the RAG-Ex baseline in Section 5.2; the paper notes that smaller windows had little impact and larger windows caused inconsistent effects, so the reported efficiency gain is sensitive to this choice.
  • document chunk size = not reported
    Documents are split into chunks using a 'predefined chunk size' in Section 3.1, but the value is never given; it controls the granularity of extracted triples and thus the KG structure and retrieval.
assumptions (4)
  • domain assumption LLM-based triplet extraction produces a factually accurate knowledge graph
    Section 3.1 builds the KG by prompting an LLM to output (entity, relation, entity) triples; there is no precision/recall evaluation of the extraction, yet all retrieval and explanations depend on the graph being correct.
  • domain assumption Shortest paths between query entities are semantically relevant reasoning chains
    Section 3.2 retrieves the shortest path between query entities as the reasoning context and falls back to regular retrieval when no path exists; the paper does not test whether shortest paths contain the information needed to answer.
  • domain assumption A change in the selected option under deletion measures a component's importance
    Section 4 defines importance as perturbation-induced output change; this assumes option flips reflect reasoning contributions rather than context sensitivity or phrasing artifacts, and no external validation is provided.
  • domain assumption Pseudo-paragraphs preserve the information content of the graph paths
    Section 3.3 converts graph paths into natural-language pseudo-paragraphs used as RAG context; the paper does not verify that this conversion loses no critical information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KGRAG-Ex: Explainable Retrieval-Augmented Generation with Knowledge Graph-based Perturbations." pith.science (2026). https://pith.science/paper/YEYSGSAH

@misc{pith2026250708443,
  author       = {Pith},
  title        = {Pith review of: KGRAG-Ex: Explainable Retrieval-Augmented Generation with Knowledge Graph-based Perturbations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YEYSGSAH}},
  note         = {Machine review of arXiv:2507.08443}
}
read the original abstract

Retrieval-Augmented Generation (RAG) enhances language models by grounding responses in external information, yet explainability remains a critical challenge, particularly when retrieval relies on unstructured text. Knowledge graphs (KGs) offer a solution by introducing structured, semantically rich representations of entities and their relationships, enabling transparent retrieval paths and interpretable reasoning. In this work, we present KGRAG-Ex, a RAG system that improves both factual grounding and explainability by leveraging a domain-specific KG constructed via prompt-based information extraction. Given a user query, KGRAG-Ex identifies relevant entities and semantic paths in the graph, which are then transformed into pseudo-paragraphs: natural language representations of graph substructures that guide corpus retrieval. To improve interpretability and support reasoning transparency, we incorporate perturbation-based explanation methods that assess the influence of specific KG-derived components on the generated answers. We conduct a series of experiments to analyze the sensitivity of the system to different perturbation methods, the relationship between graph component importance and their structural positions, the influence of semantic node types, and how graph metrics correspond to the influence of components within the explanations process.

Figures

Figures reproduced from arXiv: 2507.08443 by the authors.

Figure 1
Figure 1. Triplet extraction [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The proposed pipeline. Graph-level perturbations are applied at Step 3. 4 Explanations through Knowledge Graph Perturbations To generate explanations, we apply perturbations to the retrieved knowledge graph context by selectively removing elements at varying levels of granularity, this includes individual nodes, edges, or entire sub-paths. The motivation behind this approach is to identify which components of the kn… view at source ↗
Figure 3
Figure 3. Distribution of Critical Changes by Perturbation Position across datasets. Top: MMLU. Bottom: MedMCQA. Assessing node significance via label analysis. Our knowledge graph in￾cludes labels for each node, such as Disease, Drug, Symptom, reflecting the semantic roles of different concepts within the graph. To understand the impor￾tance of these node types, we further analyzed how critical perturbations–those that cause… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: , Disease, Symptom, and Body Part consistently emerged as the most impactful across both MMLU and MedMCQA, highlighting their key role in model reasoning. More general or peripheral labels appeared less often, indicating a weaker influence on output stability. By lever…
Figure 5
Figure 5. Figure 5: Relative rank distribution of important nodes across two datasets, based on each node degree within the derived path. 0.0 0.2 0.4 0.6 0.8 1.0 Relative Rank (0 = highest betweenness, 1 = lowest) 0 5 10 15 20 25 30 Frequency (a) MedMCQA 0.0 0.2 0.4 0.6 0.8 1.0 Relative R…
Figure 6
Figure 6. Figure 6: Relative rank distribution of important edges across two datasets, based on each edge betweenness score. critical conduits for information, lying on many of the shortest paths between pairs of nodes. In the context of our derived shortest paths, we examine the betweenn…
Figure 7
Figure 7. Figure 7: Relative rank distribution of important sub-paths across two datasets, based on sub-path score. Subpath Score = Betweenness of the Edge Degree of Node1 + Degree of Node2 (1) By normalizing edge betweenness by the degrees of the connected nodes, this score balances glob…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Features to Actions: Explainability in Traditional and Agentic AI Systems

    cs.AI 2026-02 conditional novelty 5.0 of 10

    Attribution explanations that work for static classifiers do not diagnose failures in multi-step AI agents; trace-grounded rubric evaluation does, with state-tracking inconsistency 2.7x more common in failed agent runs.

  2. Explainable Knowledge Graph Retrieval-Augmented Generation (KG-RAG) with KG-SMILE

    cs.AI 2025-09 reject novelty 4.0 of 10

    KG-SMILE applies perturbation and linear regression to a knowledge graph to attribute which entities and relations drive a GraphRAG system's answers.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Quantum critical Bose gas in the two-dimensional limit in the honeycomb antiferromagnet YbCl$_3$ under magnetic fields

    Amir Alkhamissi, Mahmoud Abdel-Aziz, Hamdy Mahgoub, Mohamed Hammad, et al. A review of large language models: Applications, challenges, and opportuni- ties. arXiv preprint arXiv:2207.02329 , 2022

  2. [2]

    Probing classifiers: Promises, shortcomings, and advances

    Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. In Proceedings of the 60th Annual Meeting of ACL (Volume 1: Long Papers) , pages 2735–2754. ACL, 2022

  3. [3]

    Walk&retrieve: Simple yet effective zero-shot retrieval-augmented generation via knowledge graph walks

    Martin Böckling, Heiko Paulheim, and Andreea Iana. Walk&retrieve: Simple yet effective zero-shot retrieval-augmented generation via knowledge graph walks. In Information Retrieval’s Role in RAG Systems (IR-RAG 2025) in conjunction with SIGIR, 2024

  4. [4]

    A survey of the state of KGRAG-Ex: Explainable RAG with KG-based Perturbations 15 explainable ai for natural language processing.arXiv preprint arXiv:2010.00711 , 2020

    Marina Danilevsky, Yannis Qian, Michael Aharon, Yannis Katsis, Ilia Kuznetsov, Pranav Sen, Anbang Singh, and Partha Pratim Talukdar. A survey of the state of KGRAG-Ex: Explainable RAG with KG-based Perturbations 15 explainable ai for natural language processing.arXiv preprint arXiv:2010.00711 , 2020

  5. [5]

    A survey on rag meeting llms: Towards retrieval- augmented large language models

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval- augmented large language models. InProceedings of the 30th ACM SIGKDD Con- ference, pages 6491–6501, 2024

  6. [6]

    Active retrieval augmented generation

    Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi- Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In Proceedings of EMNLP 2023, pages 7969–7992, 2023

  7. [7]

    Evaluating open-domain question answering in the era of large language models

    Ehsan Kamalloo, Nouha Dziri, Charles Clarke, and Davood Rafiei. Evaluating open-domain question answering in the era of large language models. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of ACL (Volume 1: Long Papers) , pages 5591–5606, Toronto, Canada, July 2023. ACL

  8. [8]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. NeurIPS, 33:9459–9474, 2020

Show all 17 references
  1. [9]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In H. Larochelle, M...

  2. [10]

    A comprehen- sive overview of large language models.ACM Transactions on Intelligent Systems and Technology, 2023

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehen- sive overview of large language models.ACM Transactions on Intelligent Systems and Technology, 2023

  3. [11]

    Rage against the machine: Retrieval-augmented llm explanations

    Joel Rorseth, Parke Godfrey, Lukasz Golab, Divesh Srivastava, and Jaroslaw Szlichta. Rage against the machine: Retrieval-augmented llm explanations. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) , 2024

  4. [12]

    Rag-ex: A generic framework for explaining retrieval augmented generation

    Viju Sudhi, Sinchana Ramakanth Bhat, Max Rudat, and Roman Teucher. Rag-ex: A generic framework for explaining retrieval augmented generation. SIGIR ’24. Association for Computing Machinery, 2024

  5. [13]

    Query2doc: Query expansion with large language models, 2023

    Liang Wang, Nan Yang, and Furu Wei. Query2doc: Query expansion with large language models, 2023. Available at: https://arxiv.org/abs/2303.07678

  6. [14]

    Benchmarking retrieval-augmented generation for medicine

    Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. Benchmarking retrieval-augmented generation for medicine. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of ACL 2024 , pages 6233–6251, Bangkok, Thai- land and virtual meeting, August 2024. ACL

  7. [15]

    Hashimoto

    Tianyi Zhang, Faisal Ladhak, Esin Durmus, Percy Liang, Kathleen McKeown, and Tatsunori B. Hashimoto. Benchmarking large language models for news summa- rization. Transactions of ACL, 12:39–57, 01 2024

  8. [16]

    Knowledge-grounded dialogue generation with pre-trained language models

    Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. Knowledge-grounded dialogue generation with pre-trained language models. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors,Proceedings of EMNLP 2020, pages 3377–3390, Online, November 2020. ACL

  9. [17]

    Knowledge graph- guided retrieval augmented generation, 2025

    Xiangrong Zhu, Yuexiang Xie, Yi Liu, Yaliang Li, and Wei Hu. Knowledge graph- guided retrieval augmented generation, 2025. arXiv:2502.06864

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.