Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

By perturbing a knowledge graph and fitting a weighted linear surrogate, KG-SMILE attributes each GraphRAG answer to the entities and relations that most influenced it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

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

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Interesting idea, but as written the attribution method is circular and the numbers don't hold together; not yet usable as evidence. the 4 major comments →

arxiv 2509.03626 v1 pith:M5FOC2FG submitted 2025-09-03 cs.AI

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

Zahra Zehtabi Sabeti Moghaddam , Zeinab Dehghani , Maneeha Rani , Koorosh Aslansefat , Bhupesh Kumar Mishra , Rameez Raja Kureshi , Dhavalkumar Thakker This is my paper
classification cs.AI
keywords Retrieval-Augmented GenerationKnowledge GraphExplainabilityPerturbation-based attributionLinear surrogate modelsGraphRAGWasserstein distanceBiomedical question answering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

RAG systems answer questions using external knowledge, but they do not say which retrieved facts shaped the answer. This paper argues that when the external knowledge is a knowledge graph, you can find out by deleting triples, measuring how much each deletion changes the answer, and fitting a simple regression that converts those changes into importance scores for nodes and edges. The resulting method, KG-SMILE, is model-agnostic and works on top of any GraphRAG pipeline. On a biomedical question-answering graph, its attributions are accurate, stable, and faithful at deterministic decoding, although higher decoding temperatures weaken all three properties. The point is not that the system answers better, but that it can show its work.

Core claim

The paper's central claim is that a perturbation-similarity-regression loop can expose which knowledge-graph components drive a GraphRAG response. Starting from a KG and a query, KG-SMILE generates a baseline answer, then repeatedly removes triples to create perturbed graphs and generates answers for each. Response similarities (inverse Wasserstein distance on text, cosine similarity on embeddings) convert each perturbation into a weight, and a weighted linear regression predicts those similarities from the perturbation pattern. The regression coefficients are read as the contribution of individual nodes and relations. Evaluation on ten biomedical questions shows near-perfect attribution fid

What carries the argument

KG-SMILE's load-bearing object is the weighted linear surrogate trained on perturbed graph vectors: features are triple-deletion indicators Pi, weights Wi come from kernel-smoothed similarity between original and perturbed responses, and the target Si is that same similarity. Coefficients beta_j are the explanation. The text-side similarity is inverse Wasserstein distance rather than cosine because it is more sensitive to distributional shifts and flags nodes whose removal opposes the original output; the paper accepts a negligible increase in loss for this interpretability gain.

Load-bearing premise

The load-bearing premise is that the regression coefficients are unbiased importance estimates even though the weights and the outcome scores both come from the same response-similarity measurements.

What would settle it

Build a small graph with a known generation rule—e.g., the answer changes if and only if one specific triple is removed—then run KG-SMILE on perturbations that include that triple. If the coefficient ranking does not put the rule's dependency at the top, or if random similarity scores produce equally high R2, the attribution is an artifact of the perturbation design rather than a genuine measure of influence.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Every GraphRAG query can return not only an answer but an importance map over the KG, color-coded by regression coefficient, showing which entities and relations carried the output.
  • At temperature 0, attributions are near-deterministic: accuracy against ground-truth relevant nodes averages about 0.88 AUC and Jaccard stability is 1.0 for most single-triple additions, so low-temperature GraphRAG can be audited.
  • At temperature 1, attribution accuracy falls to about 0.74 AUC and stability drops, so stochastic decoding should be avoided or separately qualified in high-stakes uses.
  • Inverse Wasserstein distance plus a linear surrogate is deliberately preferred over marginally better cosine fits because the coefficients stay transparent and interpretable.
  • Pre-prompting with rephrased queries and answer aggregation improves retrieval robustness but spreads attribution across more nodes, weakening pinpoint explainability.

Where Pith is reading between the lines

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

  • Because the weights and regression targets are derived from the same similarity scores, the near-perfect R2 values reported are partly self-consistency; a decisive test would use held-out perturbations or synthetic graphs with known true influence.
  • The surrogate is linear in triple-deletion indicators, so if the generator's dependence on graph structure is nonlinear or involves interactions among triples, coefficients can misrank components; a nonlinear or interaction-term surrogate is a natural extension.
  • The method should transfer to any structured retriever, not just biomedical QA; applying it to legal or financial GraphRAG with the same metrics would test whether the stability conclusions generalize.
  • A minimal external validation: compare KG-SMILE rankings to leave-one-out answer change on the same queries; agreement on the top few nodes would corroborate the attribution without relying on the method's own similarity metric.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes KG-SMILE, a perturbation-based explainability framework for GraphRAG. The method perturbs a knowledge graph by removing triples, computes similarity between original and perturbed generated responses using cosine similarity and inverse Wasserstein distance, assigns weights to perturbations, and trains a weighted linear regression surrogate whose coefficients are interpreted as the influence of KG nodes and edges. The framework is evaluated on a diabetes-related subset of PrimeKGQA with attribution fidelity, accuracy, faithfulness, stability, and consistency metrics, and the authors claim that KG-SMILE yields stable, human-aligned explanations and improves transparency of GraphRAG.

Significance. If the proposed method worked as claimed, it would address a real need: making GraphRAG outputs more transparent by identifying which KG components drive generated answers. The paper also engages with a broad suite of attribution metrics and includes computational complexity comparisons, which is a useful framing. However, the central methodological construction appears to be circular, and the quantitative evidence is internally inconsistent and based on a very small evaluation. The contribution is therefore not established in its current form.

major comments (4)
  1. [§4, Eq. (11)] Eq. (11) is not a Wasserstein distance. It is an Lp norm between paired embedding coordinates, which presupposes that the two embedding vectors have the same dimension and coordinate ordering; it does not solve an optimal-transport problem. Since inverse Wasserstein distance is central to the similarity scores and to the perturbation weights, the validity of the fidelity and attribution results built on Eq. (11) is not established. The notation δ(Emb(T_j)) is also undefined and confusing.
  2. [§4, Steps 2–3, Eqs. (12)–(13)] The regression target S_i is the similarity score between original and perturbed responses, while the weights W_i are said to be derived from 'the calculated distances' — which include the inverse Wasserstein distance that constitutes S_i. If W_i = f(S_i), then Eq. (13) regresses S_i on f(S_i)·P_i, and the resulting coefficients and near-perfect R² are forced by construction rather than measuring causal influence. The text never defines W_i independently; π_i in Eq. (12) is cosine-based and is not connected to W_i. The abstract's central claim that KG-SMILE 'identifies the graph entities and relations most influential to generated outputs' therefore rests on an unvalidated, likely circular construction.
  3. [§5.3 Table 3 and §5.4 Table 4] There are direct internal contradictions. In §5.3 the text reports r=0.933 at T=0 and r=0.070 at T=1, but Table 3 reports r=0.975524 and r=0.845926; these are materially different. In §5.4 the text claims high Jaccard at T=0 and declines at T=1, but Table 4's first row shows Jaccard(T=1)=1.00 and Jaccard(T=0)=0.10, opposite to the narrative. Such inconsistencies undermine the empirical claims and must be corrected.
  4. [§5.2, §5.3, Table 7] The attribution-accuracy evaluation uses only 10 queries with no baseline or comparison method, and several per-question AUC values are low (e.g., 0.10 and 0.44). The faithfulness metric in §5.3 correlates ATT-AUC with 'externally reported benchmark accuracies,' but no such benchmark accuracies are specified for the 10 prompts, and correlation with benchmark accuracy is not evidence that explanations reflect the model's internal reasoning. Without a ground-truth importance benchmark or a comparison to existing explainability methods, the claim that KG-SMILE produces human-aligned explanations is not supported.
minor comments (4)
  1. [§5.1, Eq. (15)] The formula for R² is miswritten: the denominator uses f(Z_i) without explicitly indicating the mean, and the numerator/denominator are not the standard corrected sums of squares. Please revise for clarity.
  2. [Abstract and §5.7] The abstract says 'interoperability' where 'interpretability' is presumably intended. Also, §5.7 on Chain-of-Thought is more of a qualitative illustration than a numerical evaluation; consider moving it to a motivating example or supplement.
  3. [References] Several references are placeholder or incomplete entries, e.g., [42] 'DSEG-LIME' with unresolved contentReference markers, [43] 'SLICE', and [45] 'SS-LIME'. These must be completed before publication.
  4. [§5.4 Table 4] The table column order and the text should be reconciled; the current table appears to contradict the stability discussion. Please also clarify whether Jaccard is computed on sets of highlighted nodes, relations, or full explanations.

Circularity Check

1 steps flagged

Central surrogate regression uses precomputed impact weights as features, forcing the β importance coefficients by construction.

specific steps
  1. fitted input called prediction [Section 4, Step 2–3, Eqs. 10–14]
    "The calculated distances are used to assign weights to the removed graph components, with a kernel function applied to adjust each component’s contribution. These weights highlight the most influential sections of the KG, identifying the parts critical for generating accurate responses [20]. ... [Step 3] Si = β0 + β1Wi1Pi1 + · · ·+ βkWikPik + ϵi ... Wi1, Wi2, . . . , Wik are their corresponding weights, reflecting impact on responses"

    Eq. 13's target S_i is defined as 'the similarity score between the original and perturbed graph responses' — i.e., how much a perturbation changed the output. The features W_i P_i are built from weights W_i that the paper says are assigned from 'the calculated distances' (cosine, WD, inverse WD) and that already 'highlight the most influential sections of the KG' and 'reflect impact on responses.' So the regression explains an impact measure by precomputed impact weights; β then merely re-labels the input. If W_i is a kernel of the same inverse-Wasserstein score later selected as the target S_i (§5.1), Eq. 13 reduces to S_i ≈ β0 + β1 f(S_i)P_i and the reported R²≈1 is forced by construction. No independent formula for W_i or ground-truth check of β is provided (ATT-accuracy uses only 10 q

full rationale

The claimed contribution of KG-SMILE is the weighted linear surrogate of Eq. 13: the β coefficients are supposed to reveal which KG nodes/edges drive output changes. That claim is circular in a load-bearing way: the input features are constructed from weights that the paper itself describes as already highlighting the most influential sections of the KG and reflecting impact on responses, while the regression target is also a similarity/impact measure (the paper later selects inverse Wasserstein distance for its linear surrogate). Thus the input and target are two codings of the same conceptual quantity, and the near-perfect R² and β values are consequences of this construction, not evidence about true node/edge influence. The paper never defines W_i independently of the computed distances/similarities, and the ATT-accuracy evaluation uses only 10 queries with no baseline, so the central importance-attribution claim is not independently established. The self-citations to SMILE ([11], [47]–[49]) are not themselves circular: SMILE is a published external framework, and the paper does not import a uniqueness theorem from these citations. Separately, §5.3's 'external benchmark accuracy of GPT-3.5-turbo' is never sourced, and several references (e.g., [43], [45]) are placeholders; these are support gaps that further weaken the validation but are distinct from the central circular construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method rests on several unspecified choices: sigma in Eq. 12, the number of perturbations, the embedding model, and the kernel function. The assumption that Eq. 11 is a Wasserstein distance is a mathematical claim that appears incorrect as written. The assumption that linear coefficients from a regression using target-derived weights are unbiased importance measures is unverified.

free parameters (4)
  • sigma (Eq. 12 kernel scaling) = not reported
    Controls the width of the weighting kernel; value not given in the paper.
  • number of perturbations = 20
    Chosen after trying 10, 20, 30, 60, 120; no principled criterion.
  • embedding model = not fully specified
    Embeddings drive all similarity scores; model choice is not stated.
  • top-10 connected components threshold = 10
    Selects the subgraph used for evaluation; may bias results.
axioms (4)
  • domain assumption The surrogate model in Eq. 13 can linearly predict response similarity from weighted perturbation indicators.
    The method assumes a linear relationship between weighted perturbations and response similarity, which is not derived or validated.
  • domain assumption Eq. 11 computes a Wasserstein distance between original and perturbed response embeddings.
    The formula as written is an Lp norm between paired coordinates, not the standard Wasserstein distance; the paper relies on it as WD without proof.
  • domain assumption The ground truth highlighted nodes used for AUC in Section 5.2 are correctly defined.
    No definition or source is given for which nodes are critical.
  • ad hoc to paper ATT-AUC correlated with external benchmark accuracy measures faithfulness.
    This operationalization is introduced without justification and conflates competence with faithfulness.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable Knowledge Graph Retrieval-Augmented Generation (KG-RAG) with KG-SMILE." pith.science (2026). https://pith.science/paper/M5FOC2FG

@misc{pith2026250903626,
  author       = {Pith},
  title        = {Pith review of: Explainable Knowledge Graph Retrieval-Augmented Generation (KG-RAG) with KG-SMILE},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5FOC2FG}},
  note         = {Machine review of arXiv:2509.03626}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generative AI, such as Large Language Models (LLMs), has achieved impressive progress but still produces hallucinations and unverifiable claims, limiting reliability in sensitive domains. Retrieval-Augmented Generation (RAG) improves accuracy by grounding outputs in external knowledge, especially in domains like healthcare, where precision is vital. However, RAG remains opaque and essentially a black box, heavily dependent on data quality. We developed a method-agnostic, perturbation-based framework that provides token and component-level interoperability for Graph RAG using SMILE and named it as Knowledge-Graph (KG)-SMILE. By applying controlled perturbations, computing similarities, and training weighted linear surrogates, KG-SMILE identifies the graph entities and relations most influential to generated outputs, thereby making RAG more transparent. We evaluate KG-SMILE using comprehensive attribution metrics, including fidelity, faithfulness, consistency, stability, and accuracy. Our findings show that KG-SMILE produces stable, human-aligned explanations, demonstrating its capacity to balance model effectiveness with interpretability and thereby fostering greater transparency and trust in machine learning technologies.

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Evaluating Retrieval-Augmented Generation for Explainable Malware Analysis

    cs.CR 2026-05 unverdicted novelty 5.0

    RAG frequently degrades LLM malware explanations when structured VirusTotal input is already available by introducing irrelevant context and narrative noise.

  2. ConceptSMILE: Auditing the Trustworthiness of Concept-Based Explainable AI

    cs.AI 2026-07 conditional novelty 4.0

    A perturbation-and-surrogate audit shows MedSAM and VLM retinal concept explanations have pathway- and concept-specific reliability, not automatic trustworthiness.

Reference graph

Works this paper leans on

132 extracted references · 29 canonical work pages · cited by 2 Pith papers · 9 internal anchors

  1. [1]

    Advances in neural information processing systems 33, 9459–9474 (2020)

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., Yih, W.-t., Rockt¨ aschel, T.,et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474 (2020)

    2020

  2. [2]

    arXiv preprint arXiv:2404.08189 (2024)

    B´ echard, P., Ayala, O.M.: Reducing hallucination in structured outputs via retrieval-augmented generation. arXiv preprint arXiv:2404.08189 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    arXiv preprint arXiv:2211.09110 (2022) 33

    Liang, P., Bommasani, R., Lee, T., Tsipras, D., Soylu, D., Yasunaga, M., Zhang, Y., Narayanan, D., Wu, Y., Kumar, A., et al.: Holistic evaluation of language models. arXiv preprint arXiv:2211.09110 (2022) 33

    Pith/arXiv arXiv 2022

  5. [5]

    ACM computing surveys 55(12), 1–38 (2023)

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A., Fung, P.: Survey of hallucination in natural language generation. ACM computing surveys 55(12), 1–38 (2023)

    2023

  6. [6]

    : A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. : A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems 43(2), 1–55 (2025)

    2025

  7. [7]

    arXiv preprint arXiv:2109.07958 (2021)

    Lin, S., Hilton, J., Evans, O.: Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958 (2021)

    Pith/arXiv arXiv 2021

  8. [8]

    : Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai

    Arrieta, A.B., D ´ ıaz-Rodr ´ ıguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., Garc ´ ıa, S., Gil-L´ opez, S., Molina, D., Benjamins, R.,et al. : Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai. Information fusion 58, 82–115 (2020)

    2020

  9. [9]

    arXiv preprint arXiv:2203.13224 (2022)

    Shuster, K., Komeili, M., Adolphs, L., Roller, S., Szlam, A., Weston, J.: Language models that seek for knowledge: Modular search & generation for dialogue and prompt completion. arXiv preprint arXiv:2203.13224 (2022)

    Pith/arXiv arXiv 2022

  10. [10]

    ACM Computing Surveys 56(4), 1–62 (2023)

    Zhong, L., Wu, J., Li, Q., Peng, H., Wu, X.: A comprehensive survey on automatic knowledge graph construction. ACM Computing Surveys 56(4), 1–62 (2023)

    2023

  11. [11]

    IEEE Software 41(1), 87–97 (2023)

    Aslansefat, K., Hashemian, M., Walker, M., Akram, M.N., Sorokos, I., Papadopou- los, Y.: Explaining black boxes with a smile: Statistical model-agnostic interpretability with local explanations. IEEE Software 41(1), 87–97 (2023)

    2023

  12. [12]

    arXiv preprint arXiv:2110.00976 (2021)

    Chalkidis, I., Jana, A., Hartung, D., Bommarito, M., Androutsopoulos, I., Katz, D.M., Aletras, N.: Lexglue: A benchmark dataset for legal language understanding in english. arXiv preprint arXiv:2110.00976 (2021)

    Pith/arXiv arXiv 2021

  13. [13]

    arXiv preprint arXiv:1908.10084 (2019)

    Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019)

    Pith/arXiv arXiv 1908

  14. [14]

    In: International Conference on Machine Learning, pp

    Kusner, M., Sun, Y., Kolkin, N., Weinberger, K.: From word embeddings to document distances. In: International Conference on Machine Learning, pp. 957–966 (2015). PMLR

    2015

  15. [15]

    Automatic Text Evaluation through the Lens of Wasserstein Barycenters

    Colombo, P., Staerman, G., Clavel, C., Piantanida, P.: Automatic text evaluation through the lens of wasserstein barycenters. arXiv preprint arXiv:2108.12463 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  16. [16]

    IEEE access 6, 52138–52160 (2018)

    Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial intelligence (xai). IEEE access 6, 52138–52160 (2018)

    2018

  17. [17]

    arXiv preprint arXiv:1702.08608 (2017)

    Doshi-Velez, F., Kim, B.: Towards a rigorous science of interpretable machine 34 learning. arXiv preprint arXiv:1702.08608 (2017)

    Pith/arXiv arXiv 2017

  18. [18]

    Christoph, M.: Interpretable machine learning: A guide for making black box models explainable (2020)

    2020

  19. [19]

    Nature machine intelligence 1(5), 206–215 (2019)

    Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206–215 (2019)

    2019

  20. [20]

    In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp

    Ribeiro, M.T., Singh, S., Guestrin, C.: ” why should i trust you?” explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1135– 1144 (2016)

    2016

  21. [21]

    Advances in neural information processing systems 30 (2017)

    Lundberg, S.M., Lee, S.-I.: A unified approach to interpreting model predictions. Advances in neural information processing systems 30 (2017)

    2017

  22. [22]

    ACM computing surveys (CSUR) 51(5), 1–42 (2018)

    Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., Pedreschi, D.: A survey of methods for explaining black box models. ACM computing surveys (CSUR) 51(5), 1–42 (2018)

    2018

  23. [23]

    Communications of the ACM 63(1), 68–77 (2019)

    Du, M., Liu, N., Hu, X.: Techniques for interpretable machine learning. Communications of the ACM 63(1), 68–77 (2019)

    2019

  24. [24]

    Shapley, L.S., et al.: A value for n-person games (1953)

    1953

  25. [25]

    arXiv preprint arXiv:1802.03888 (2018)

    Lundberg, S.M., Erion, G.G., Lee, S.-I.: Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888 (2018)

    Pith/arXiv arXiv 2018

  26. [26]

    In: International Conference on Machine Learning, pp

    Sundararajan, M., Najmi, A.: The many shapley values for model explanation. In: International Conference on Machine Learning, pp. 9269–9278 (2020). PMLR

    2020

  27. [27]

    Advances in neural information processing systems 31 (2018)

    Plumb, G., Molitor, D., Talwalkar, A.S.: Model agnostic supervised local explanations. Advances in neural information processing systems 31 (2018)

    2018

  28. [28]

    arXiv preprint arXiv:1812.04608 (2018)

    Hoffman, R.R., Mueller, S.T., Klein, G., Litman, J.: Metrics for explainable ai: Challenges and prospects. arXiv preprint arXiv:1812.04608 (2018)

    Pith/arXiv arXiv 2018

  29. [29]

    Defining Locality for Surrogates in Post-hoc Interpretablity

    Laugel, T., Renard, X., Lesot, M.-J., Marsala, C., Detyniecki, M.: Defining locality for surrogates in post-hoc interpretablity. arXiv preprint arXiv:1806.07498 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  30. [30]

    Zhao, X., Huang, W., Huang, X., Robu, V., Flynn, D.: Baylime: Bayesian local interpretable model-agnostic explanations (supplementary material)

  31. [31]

    In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp

    Zhou, Z., Hooker, G., Wang, F.: S-lime: Stabilized-lime for model explanation. In: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp. 2429–2438 (2021) 35

    2021

  32. [32]

    Extending LIME for Business Process Automation

    Upadhyay, S., Isahagian, V., Muthusamy, V., Rizk, Y.: Extending lime for business process automation. arXiv preprint arXiv:2108.04371 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  33. [33]

    OptiLIME: Optimized LIME Explanations for Diagnostic Computer Algorithms

    Visani, G., Bagli, E., Chesani, F.: Optilime: Optimized lime explanations for diagnostic computer algorithms. arXiv preprint arXiv:2006.05714 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  34. [34]

    In: Intelligent Data Engineering and Automated Learning– IDEAL 2019: 20th International Conference, Manchester, UK, November 14–16, 2019, Proceedings, Part I 20, pp

    Shankaranarayana, S.M., Runje, D.: Alime: Autoencoder based approach for local interpretability. In: Intelligent Data Engineering and Automated Learning– IDEAL 2019: 20th International Conference, Manchester, UK, November 14–16, 2019, Proceedings, Part I 20, pp. 454–463 (2019). Springer

    2019

  35. [35]

    Neurocomputing 597, 127969 (2024)

    Saadatfar, H., Kiani-Zadegan, Z., Ghahremani-Nezhad, B.: Us-lime: Increasing fidelity in lime using uncertainty sampling on tabular data. Neurocomputing 597, 127969 (2024)

    2024

  36. [36]

    Machine Learning and Knowledge Extraction 3(3), 525–541 (2021)

    Zafar, M.R., Khan, N.: Deterministic local interpretable model-agnostic expla- nations for stable explainability. Machine Learning and Knowledge Extraction 3(3), 525–541 (2021)

    2021

  37. [37]

    In: International Conference on Artificial Intelligence and Statistics, pp

    Garreau, D., Luxburg, U.: Explaining the explainer: A first theoretical analysis of lime. In: International Conference on Artificial Intelligence and Statistics, pp. 1287–1296 (2020). PMLR

    2020

  38. [38]

    In: ISMIR, vol

    Mishra, S., Sturm, B.L., Dixon, S.: Local interpretable model-agnostic explana- tions for music content analysis. In: ISMIR, vol. 53, pp. 537–543 (2017)

    2017

  39. [39]

    Artificial Intelligence 314, 103823 (2023)

    Li, X., Xiong, H., Li, X., Zhang, X., Liu, J., Jiang, H., Chen, Z., Dou, D.: G-lime: Statistical learning for local interpretations of deep neural networks using global priors. Artificial Intelligence 314, 103823 (2023)

    2023

  40. [40]

    IEEE Transactions on Knowledge and Data Engineering 35(7), 6968–6972 (2022)

    Huang, Q., Yamada, M., Tian, Y., Singh, D., Chang, Y.: Graphlime: Local interpretable model explanations for graph neural networks. IEEE Transactions on Knowledge and Data Engineering 35(7), 6968–6972 (2022)

    2022

  41. [41]

    Processes 11(2), 595 (2023)

    Abdullah, T.A., Zahid, M.S.M., Ali, W., Hassan, S.U.: B-lime: An improvement of lime for interpretable deep learning classification of cardiac arrhythmia from ecg signals. Processes 11(2), 595 (2023)

    2023

  42. [42]

    arXiv preprint arXiv:2403.07733 (2024)

    Knab, P., Marton, S., Bartelt, C.: Beyond pixels: Enhancing lime with hierarchical features and segmentation foundation models. arXiv preprint arXiv:2403.07733 (2024). DSEG-LIME: Data -Driven Segmentation LIME :con- tentReference[oaicite:1]index=1

    arXiv 2024

  43. [43]

    In: Proceedings of [venue] (2024)

    Bora, N., Others: SLICE: Structured sampling to improve lime image explanations. In: Proceedings of [venue] (2024). DOI or arXiv not found; placeholder entry; :contentReference[oaicite:2]index=2 36

    2024

  44. [44]

    Proceedings of the AAAI Conference on Artificial Intelligence 38(13), 14785–14792 (2024) https://doi.org/10.1609/aaai

    Rashid, M., Amparore, E.G., Ferrari, E., Verda, D.: Using stratified sampling to improve lime image explanations. Proceedings of the AAAI Conference on Artificial Intelligence 38(13), 14785–14792 (2024) https://doi.org/10.1609/aaai. v38i13.29397 . Stratified LIME :contentReference[oaicite:3]index=3

    doi:10.1609/aaai 2024

  45. [45]

    arXiv preprint arXiv:2501.xxxxx (2025)

    Lam, F., Others: Ss -lime: Self -supervised region detection for enhanced local explanations. arXiv preprint arXiv:2501.xxxxx (2025). SS -LIME placeholder; not found in search results

    2025

  46. [46]

    Survey placeholder; :contentReference[oaicite:4]index=4 (2025)

    Knab, P., Marton, S., Schlegel, U., Bartelt, C.: A Survey of LIME Extensions in Deep Learning. Survey placeholder; :contentReference[oaicite:4]index=4 (2025)

    2025

  47. [47]

    Mapping the Mind of an Instruction-based Image Editing using SMILE

    Dehghani, Z., Aslansefat, K., Khan, A., Rivera, A.R., George, F., Khalid, M.: Mapping the Mind of an Instruction-based Image Editing using SMILE (2024). https://arxiv.org/abs/2412.16277

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [48]

    https://arxiv.org/abs/2505

    Dehghani, Z., Akram, M.N., Aslansefat, K., Khan, A., Papadopoulos, Y.: Explain- ing Large Language Models with gSMILE (2025). https://arxiv.org/abs/2505. 21657

    2025

  49. [49]

    arXiv preprint arXiv:2410.15374 (2024)

    Ahmadi, S.M., Aslansefat, K., Valcarce-Dineiro, R., Barnfather, J.: Explain- ability of point cloud neural networks using smile: Statistical model-agnostic interpretability with local explanations. arXiv preprint arXiv:2410.15374 (2024)

    Pith/arXiv arXiv 2024

  50. [50]

    IEEE Transactions on Knowledge and Data Engineering (2024)

    Pan, S., Luo, L., Wang, Y., Chen, C., Wang, J., Wu, X.: Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  51. [51]

    arXiv preprint arXiv:2410.20724 (2024)

    Li, M., Miao, S., Li, P.: Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation. arXiv preprint arXiv:2410.20724 (2024)

    Pith/arXiv arXiv 2024

  52. [52]

    In: Proceedings of the 5th ACM International Conference on AI in Finance, pp

    Sarmah, B., Mehta, D., Hall, B., Rao, R., Patel, S., Pasquali, S.: Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In: Proceedings of the 5th ACM International Conference on AI in Finance, pp. 608–616 (2024)

    2024

  53. [53]

    arXiv preprint arXiv:2306.04136 (2023)

    Baek, J., Aji, A.F., Saffari, A.: Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. arXiv preprint arXiv:2306.04136 (2023)

    Pith/arXiv arXiv 2023

  54. [54]

    In: Proceedings of the ACM on Web Conference 2025, pp

    Li, Y., Zhang, X., Luo, L., Chang, H., Ren, Y., King, I., Li, J.: G-refer: Graph retrieval-augmented large language model for explainable recommendation. In: Proceedings of the ACM on Web Conference 2025, pp. 240–251 (2025)

    2025

  55. [55]

    arXiv 37 preprint arXiv:2507.08443 (2025)

    Balanos, G., Chasanis, E., Skianis, K., Pitoura, E.: Kgrag-ex: Explainable retrieval-augmented generation with knowledge graph-based perturbations. arXiv 37 preprint arXiv:2507.08443 (2025)

    Pith/arXiv arXiv 2025

  56. [56]

    arXiv preprint arXiv:2506.00783 (2025)

    Wu, R., Cai, P., Mei, J., Wen, L., Hu, T., Yang, X., Fu, D., Shi, B.: Kg-traces: Enhancing large language models with knowledge graph-constrained trajectory reasoning and attribution supervision. arXiv preprint arXiv:2506.00783 (2025)

    arXiv 2025

  57. [57]

    arXiv preprint arXiv:2508.07117 (2025)

    Baghershahi, P., Fournier, G., Nyati, P., Medya, S.: From nodes to narratives: Explaining graph neural networks with llms and graph context. arXiv preprint arXiv:2508.07117 (2025)

    Pith/arXiv arXiv 2025

  58. [58]

    arXiv preprint arXiv:2408.08921 (2024)

    Peng, B., Zhu, Y., Liu, Y., Bo, X., Shi, H., Hong, C., Zhang, Y., Tang, S.: Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921 (2024)

    Pith/arXiv arXiv 2024

  59. [59]

    arXiv preprint arXiv:2307.07697 (2023)

    Sun, J., Xu, C., Tang, L., Wang, S., Lin, C., Gong, Y., Shum, H.-Y., Guo, J.: Think-on-graph: Deep and responsible reasoning of large language model with knowledge graph. arXiv preprint arXiv:2307.07697 (2023)

    Pith/arXiv arXiv 2023

  60. [60]

    arXiv preprint arXiv:2508.12495 (2025)

    Li, Y., Shen, Y., Nian, Y., Gao, J., Wang, Z., Yu, C., Li, S., Wang, J., Hu, X., Zhao, Y.: Mitigating hallucinations in large language models via causal reasoning. arXiv preprint arXiv:2508.12495 (2025)

    arXiv 2025

  61. [61]

    arXiv preprint arXiv:2402.10612 (2024)

    Ding, H., Pang, L., Wei, Z., Shen, H., Cheng, X.: Retrieve only when it needs: Adaptive retrieval augmentation for hallucination mitigation in large language models. arXiv preprint arXiv:2402.10612 (2024)

    arXiv 2024

  62. [62]

    arXiv preprint arXiv:2407.00379 (2024)

    Tang, J., Zhang, Q., Li, Y., Li, J.: Grapharena: Benchmarking large language models on graph computational problems. arXiv preprint arXiv:2407.00379 (2024)

    Pith/arXiv arXiv 2024

  63. [63]

    : Exploring the potential of large language models (llms) in learning on graphs

    Chen, Z., Mao, H., Li, H., Jin, W., Wen, H., Wei, X., Wang, S., Yin, D., Fan, W., Liu, H., et al. : Exploring the potential of large language models (llms) in learning on graphs. ACM SIGKDD Explorations Newsletter 25(2), 42–61 (2024)

    2024

  64. [64]

    GraphXAIN: Narratives to Explain Graph Neural Networks

    Cedro, M., Martens, D.: Graphxain: Narratives to explain graph neural networks. arXiv preprint arXiv:2411.02540 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  65. [65]

    arXiv preprint arXiv:2507.22915 (2025)

    Gumaan, E.: Theoretical foundations and mitigation of hallucination in large language models. arXiv preprint arXiv:2507.22915 (2025)

    Pith/arXiv arXiv 2025

  66. [66]

    arXiv preprint arXiv:2507.21509 (2025)

    Chen, R., Arditi, A., Sleight, H., Evans, O., Lindsey, J.: Persona vectors: Mon- itoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509 (2025)

    Pith/arXiv arXiv 2025

  67. [67]

    Advances in Neural Information Processing Systems 33, 1877–1901 (2020) 38

    Brown, T.B., Mann, B., Ryder, N., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877–1901 (2020) 38

    1901

  68. [68]

    arXiv preprint arXiv:2208.03299 1(2), 4 (2022)

    Izacard, G., Lewis, P., Lomeli, M., Hosseini, L., Petroni, F., Schick, T., Dwivedi- Yu, J., Joulin, A., Riedel, S., Grave, E.: Few-shot learning with retrieval augmented language models. arXiv preprint arXiv:2208.03299 1(2), 4 (2022)

    Pith/arXiv arXiv 2022

  69. [69]

    Research handbook on big data law, 171–184 (2021)

    Surden, H.: Machine learning and law: An overview. Research handbook on big data law, 171–184 (2021)

    2021

  70. [70]

    arXiv preprint arXiv:2112.09332 (2021)

    Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021)

    Pith/arXiv arXiv 2021

  71. [71]

    arXiv preprint arXiv:2303.08896 (2023)

    Manakul, P., Liusie, A., Gales, M.J.: Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896 (2023)

    Pith/arXiv arXiv 2023

  72. [72]

    arXiv preprint arXiv:2108.07258 (2021)

    Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    Pith/arXiv arXiv 2021

  73. [73]

    In: ICLR Workshop on Representation Learning on Graphs and Manifolds (2019)

    Wang, M.Y.: Deep graph library: Towards efficient and scalable deep learning on graphs. In: ICLR Workshop on Representation Learning on Graphs and Manifolds (2019)

    2019

  74. [74]

    arXiv preprint arXiv:2404.16130 (2024)

    Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., Metropolitansky, D., Ness, R.O., Larson, J.: From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

    Pith/arXiv arXiv 2024

  75. [75]

    In: 31st International Conference on Computational Linguistics Workshop Knowledge Graph & GenAI (2025)

    Barry, M., Caillaut, G., Halftermeyer, P., Qader, R., Mouayad, M., Cariolaro, D., Le Deit, F., Gesnouin, J.: Graphrag: Leveraging graph-based efficiency to minimize hallucinations in llm-driven rag for finance data. In: 31st International Conference on Computational Linguistics Workshop Knowledge Graph & GenAI (2025)

    2025

  76. [76]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Li, L., Gan, Z., Cheng, Y., Liu, J.: Relation-aware graph attention network for visual question answering. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10313–10322 (2019)

    2019

  77. [77]

    arXiv preprint arXiv:2005.00631 (2020)

    Bhatt, U., Weller, A., Moura, J.M.: Evaluating and aggregating feature-based model explanations. arXiv preprint arXiv:2005.00631 (2020)

    Pith/arXiv arXiv 2005

  78. [78]

    ACM Transactions on Information Systems (TOIS) 40(4), 1–24 (2022)

    Li, Y., Li, W., Nie, L.: Dynamic graph reasoning for conversational open-domain question answering. ACM Transactions on Information Systems (TOIS) 40(4), 1–24 (2022)

    2022

  79. [79]

    arXiv preprint 39 arXiv:1904.09537 (2019)

    Sun, H., Bedrax-Weiss, T., Cohen, W.W.: Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text. arXiv preprint 39 arXiv:1904.09537 (2019)

    Pith/arXiv arXiv 1904

  80. [80]

    Reasoning Over Paths via Knowledge Base Completion

    Sudhahar, S., Roberts, I., Pierleoni, A.: Reasoning over paths via knowledge base completion. arXiv preprint arXiv:1911.00492 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1911

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.