Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Mitigating LLM Hallucinations with Knowledge Graphs: A Case Study

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Forcing an LLM to query a knowledge graph beats plain GPT-4 on every question type tested.

desk verdict A modest but honest case study: LinkQ beats plain GPT-4 at KG query generation on a curated 120-question Mintaka subset, though the 'at least one of three' scoring rule makes the size of the advantage uncertain on harder types. read the letter →

arxiv 2504.12422 v1 pith:LAAHXNQT submitted 2025-04-16 cs.HC cs.AI

classification cs.HCcs.AI
keywords LLMhallucinationknowledgegraphquestionansweringLinkQMintakaBRONGPT-4human-in-the-looptrustworthyAI
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

This paper reports a case study of LinkQ, an interface that prevents a large language model from answering questions directly and instead requires it to build and run queries against a knowledge graph, so every answer is grounded in up-to-date, verified data. On 120 questions drawn from the Mintaka dataset, LinkQ outperformed a plain GPT-4 baseline on all five tested question types, with the largest gaps on multi-hop (75.0% versus 16.7%) and intersection (54.2% versus 12.5%) questions. The authors' central claim is that LinkQ's message-passing and prompting strategy increases the objective correctness of LLM question-to-query translation without fine-tuning. A qualitative session with two cybersecurity experts using the BRON knowledge graph showed the approach can answer domain-specific questions, though it sometimes returned empty results when data existed.

What carries the argument

The load-bearing mechanism is LinkQ's human-in-the-loop prompting protocol. When a user asks a question, the system intercepts it and has the LLM clarify the question, then lets the LLM fuzzy-search for entity and relation IDs in the knowledge graph; every data access is routed through the KG API, and the retrieved graph structure and ground-truth IDs are given back to the LLM in system messages. The LLM writes a graph query from a few-shot prompt, the user sees a query preview and can edit it, and the system executes the query and displays the tabular results with an LLM summary. This forces the LLM to derive answers from the KG's actual nodes and edges rather than from parametric memory, which is what suppresses hallucinations.

What would settle it

Re-run the same 120 Mintaka questions with LinkQ and plain GPT-4 scoring by majority or all-three agreement rather than at-least-one; if the gap on multi-hop and intersection questions disappears or reverses, the claim that LinkQ increases objective correctness without fine-tuning would be refuted.

Watch

Extended reading notes

Core claim

The paper's central discovery is that routing an LLM's knowledge through a knowledge-graph query protocol measurably reduces hallucination. The authors show that LinkQ, which uses GPT-4 as its underlying model but adds system-intercepted entity and relation lookups, few-shot query examples, and a user preview step, answered more questions correctly than off-the-shelf GPT-4 on every complexity type tested: comparative 91.7% versus 20.8%, yes/no 87.5% versus 54.2%, generic 79.2% versus 33.3%, multi-hop 75.0% versus 16.7%, and intersection 54.2% versus 12.5%. They interpret the result as evidence that message passing and prompting alone, without fine-tuning, can raise objective correctness of KG query generation. They also find that LinkQ still struggles on multi-hop and intersection questions, and that in the BRON domain study it produced empty results rather than fabricated facts when the KG structure was not fully explored.

Load-bearing premise

A question counts as correctly answered if at least one of three independently generated queries returns the right answer, and the paper reports accuracies without confidence intervals; if correctness instead required all three or a majority of attempts, both systems' scores would fall and LinkQ's margin could shrink.

Editorial extensions

If this is right

  • LinkQ-style grounding is sufficient for straightforward analytical questions such as comparative, yes/no, and generic single-hop queries, which are common in initial data exploration.
  • Multi-hop and intersection questions remain the bottleneck, so future systems should decompose complex questions into smaller steps, possibly through chain-of-thought prompting, or ask the user to refine the query step by step.
  • Because the improvement comes from prompting and message passing rather than fine-tuning, the approach can be transferred to new domain-specific knowledge graphs without retraining the model.
  • When a translation fails, the system should hand control back to the user rather than improvise an answer, preserving the no-hallucination guarantee.
  • The accuracy gain comes at a runtime cost, since LinkQ takes longer than plain GPT-4 to generate a query.

Reading between the lines

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

  • The reported accuracies use an at-least-one-of-three scoring rule; re-scoring the same outputs under a majority or all-three rule would lower both systems' scores and could shrink the gap on the hardest question types.
  • The qualitative finding that domain experts could correct empty-result queries suggests that LinkQ's benefit may depend on user familiarity with the KG schema; a non-expert may not know how to repair a query that returns no data.
  • The SME suggestion to search by node descriptions rather than properties points toward a hybrid that combines LinkQ's grounded queries with retrieval-augmented generation, a direction the paper does not implement.
  • Because only one baseline model was compared, the magnitude of LinkQ's advantage may vary with the underlying LLM; re-running the same protocol with a different model would test the generality of the claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents LinkQ, an open-source LLM-assisted visual interface that translates natural-language questions into knowledge-graph (KG) queries to suppress hallucinations. It reports a quantitative evaluation on 120 Mintaka questions spanning five complexity types (Multi-hop, Comparative, Yes/No, Generic, Intersection), comparing LinkQ (which uses GPT-4 under the hood) against plain GPT-4. The authors find that LinkQ outperforms GPT-4 on every question type, with the largest gaps on Multi-hop (75.0% vs 16.7%) and Intersection (54.2% vs 12.5%). A qualitative study with two cybersecurity experts on the BRON KG provides feedback and suggestions for future systems. The paper concludes that LinkQ can improve the objective correctness of LLM-to-KG query translation without fine-tuning, though some complex question types remain challenging.

Significance. If the reported results hold, the paper offers a practical, fine-tuning-free approach to grounding LLM answers in structured knowledge, which is relevant for high-stakes domains such as cyber operations. Its strengths include the use of an external benchmark dataset (Mintaka), a human-in-the-loop design that is clearly described, and an open-source implementation that supports reproducibility. The qualitative study adds real-world context, but the evidence for the central quantitative claim is weakened by the lenient 'correct at least once in three attempts' scoring rule and the absence of statistical inference on small per-type samples.

major comments (3)
  1. [Section 3.1, Correctness Criteria] The scoring rule, which counts a question as correct if at least one of three independently generated queries returns the right answer, measures peak capability rather than reliability. In a high-stakes setting, a system that succeeds only one time out of three still produces wrong or empty answers most of the time, so this rule is not clearly aligned with the paper's stated goal of mitigating hallucinations. Because Figure 3(a) already includes a 1/3, 2/3, 3/3 breakdown, the authors should report the exact per-attempt counts and present accuracy under stricter rules (e.g., at least two or all three attempts correct) and justify the chosen threshold.
  2. [Section 3.2 and Figure 3] With only 24 questions per type, the headline differences (e.g., Multi-hop 75.0% vs 16.7% and Intersection 54.2% vs 12.5%) are not accompanied by confidence intervals or significance tests. A Fisher's exact test or a bootstrap confidence interval would show whether the observed gaps are compatible with chance. Without such measures, the claim that LinkQ 'strongly outperforms' GPT-4 is not yet statistically supported, particularly for the categories with the largest reported advantage.
  3. [Section 3.1, Baseline] The prompt and parameters used for the plain GPT-4 baseline are described only as 'instructing it to generate an appropriate KG query.' This level of detail makes it difficult to assess whether the comparison is fair (e.g., whether the baseline received the same few-shot examples as LinkQ). The authors should include the baseline prompt, temperature, and other relevant settings in the repository or an appendix so that the comparison is reproducible.
minor comments (5)
  1. [Section 2, Figure 2] The flow diagram is information-dense; adding a concise bulleted list of the protocol steps in the text would help readers who do not parse the figure in detail.
  2. [Section 3.2] The phrase 'increases the objective correctness (and consequently the runtime)' is ambiguous about whether runtime increases or decreases and why correctness would affect runtime. Please clarify the intended relationship.
  3. [Section 4] There is a typo: 'BRON was too complex to for an LLM to navigate' should read 'too complex for an LLM to navigate.'
  4. [Section 4] The qualitative study relies on only two subject-matter experts; the paper should state this limitation explicitly when drawing broader conclusions about domain-expert usability.
  5. [Section 3.1] The procedure for selecting the 120 questions (random selection with criteria) is described, but the specific seed or code for the selection is not provided; including it would strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: evaluation is anchored to external benchmarks and fresh runs; the cited prior LinkQ design is normal attribution, not a load-bearing self-citation.

full rationale

The paper's central quantitative claim is an empirical comparison of LinkQ against plain GPT-4 on questions drawn from Mintaka, with answers grounded in Wikidata. The correctness rule ('a correct answer at least once from those three attempts') is a measurement choice applied identically to both systems; it may be lenient and could affect the strength of the reported advantage, but it is not a fitted parameter, a renamed output, or a derivation from the system's own claims. LinkQ's design is attributed to the authors' prior publication [15], yet the current case study runs fresh evaluations against external data and an independent baseline, so the self-citation is not load-bearing. The qualitative study with the BRON KG is likewise an external, user-grounded observation rather than a circular justification. No step in the paper reduces, by construction or by self-citation chain, to its own inputs, so the appropriate finding is no circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new entities or fitted constants are introduced. LinkQ is prior work [15]. The only hand-chosen numeric input is the evaluation parameter, primarily the 'at least one of three' correctness threshold, which is listed as a free parameter.

free parameters (1)
  • success threshold for correctness = at least 1 of 3 attempts
    The paper counts a question as correct if any of three query attempts succeeds (Section 3.1). This hand-chosen threshold inflates reported accuracy and is load-bearing for every percentage in Figure 3.
assumptions (4)
  • domain assumption A generated query that returns the gold answer is a valid proxy for hallucination-free question answering.
    The evaluation measures query correctness, not hallucination in free-form answers or summaries; the paper assumes query success equals grounded behavior.
  • ad hoc to paper The manually curated 120-question subset represents realistic natural language questions for KGQA.
    Section 3.1 selects 3 questions per type per topic under exclusion criteria; no evidence is given that this subset represents the Mintaka distribution or real user queries.
  • domain assumption GPT-4 has not been trained on the BRON knowledge graph.
    Section 4 states 'our tests suggest GPT-4 has not been trained on BRON', an unverifiable claim used to motivate the qualitative demonstration.
  • ad hoc to paper The 'at least once in three attempts' scoring rule is a fair accuracy measure.
    Section 3.1 defines correctness this way without justification; it is a methodological assumption that directly shapes all reported accuracies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating LLM Hallucinations with Knowledge Graphs: A Case Study." pith.science (2026). https://pith.science/paper/LAAHXNQT

@misc{pith2026250412422,
  author       = {Pith},
  title        = {Pith review of: Mitigating LLM Hallucinations with Knowledge Graphs: A Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAAHXNQT}},
  note         = {Machine review of arXiv:2504.12422}
}
read the original abstract

High-stakes domains like cyber operations need responsible and trustworthy AI methods. While large language models (LLMs) are becoming increasingly popular in these domains, they still suffer from hallucinations. This research paper provides learning outcomes from a case study with LinkQ, an open-source natural language interface that was developed to combat hallucinations by forcing an LLM to query a knowledge graph (KG) for ground-truth data during question-answering (QA). We conduct a quantitative evaluation of LinkQ using a well-known KGQA dataset, showing that the system outperforms GPT-4 but still struggles with certain question categories - suggesting that alternative query construction strategies will need to be investigated in future LLM querying systems. We discuss a qualitative study of LinkQ with two domain experts using a real-world cybersecurity KG, outlining these experts' feedback, suggestions, perceived limitations, and future opportunities for systems like LinkQ.

Figures

Figures reproduced from arXiv: 2504.12422 by the authors.

Figure 1
Figure 1. The LinkQ system [15], a natural language interface that combats an LLM’s tendency to hallucinate by instructing it to answer users’ questions by creating and executing knowledge graph (KG) queries. LinkQ ensures all data retrieved and summarized by the LLM comes from ground truth, up-to-date data in the KG. High-stakes domains like cyber operations need responsible and trustworthy AI methods. While large language m… view at source ↗
Figure 2
Figure 2. An overview of LinkQ’s [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Results from our quantitative evaluation (Section 3) comparing LinkQ (blue) and GPT-4 (orange). Left: LinkQ and GPT-4’s [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. From Static Bibliometrics to Dynamic Knowledge Graphs: An LLM-Powered Framework for Modernizing Science, Technology, and Innovation (STI) Analytics

    cs.DL 2026-07 conditional novelty 4.0 of 10

    A conceptual framework for STI analytics that uses LLMs only to propose semantic edges into a versioned knowledge graph, admitting them only after structural, evidentiary, comparative, and expert validation.

Reference graph

Works this paper leans on

24 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona Diab, and Marjan Ghazvininejad. 2022. A Review on Language Models as Knowledge Bases. arXiv preprint arXiv:2204.06031 (2022). https://doi.org/10.48550/arXiv.2204.06031 arXiv:2204.06031 [cs.CL]

  2. [2]

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. Concrete problems in AI safety. arXiv preprint arXiv:1606.06565 (2016)

  3. [3]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (Virtual Event, Canada) (FAccT ’21). Association for Computing Machinery, New York, NY, USA, 610–623. https:/...

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

  5. [5]

    William N Caballero and Phillip R Jenkins. 2024. On large language models in national security applications. arXiv preprint arXiv:2407.03453 (2024)

  6. [6]

    Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608 (2017)

  7. [7]

    Lisa Ehrlinger and Wolfram Wöß. 2016. Towards a definition of knowledge graphs. Proc. ESWC Posters and Demos Track 48, 1-4 (2016), 2

  8. [8]

    Erik Hemberg, Jonathan Kelly, Michal Shlapentokh-Rothman, Bryn Reinstadler, Katherine Xu, Nick Rutar, and Una-May O’Reilly. 2020. Linking threat tactics, techniques, and patterns with defensive weaknesses, vulnerabilities and affected platform configurations for cyber hunting. arXiv preprint arXiv:2010.00533 (2020). https://doi.org/10.48550/arXiv.2010.00533

Show all 24 references
  1. [9]

    Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky, and Sharese King. 2024. Dialect prejudice predicts AI decisions about people’s character, employability, and criminality. arXiv preprint arXiv:2403.00742 (2024)

  2. [10]

    Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia D’amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel-Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequed...

  3. [11]

    Josua Krause, Adam Perer, and Kenney Ng. 2016. Interacting with predictions: Visual inspection of black-box machine learning models. InProceedings of the 2016 CHI conference on human factors in computing systems . 5686–5697

  4. [12]

    David Leslie. 2019. Understanding artificial intelligence ethics and safety. arXiv preprint arXiv:1906.05684 (2019)

  5. [13]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proc. NIPS (...

  6. [14]

    Harry Li, Gabriel Appleby, Camelia Daniela Brumar, Remco Chang, and Ashley Suh. 2024. Knowledge Graphs in Practice: Characterizing their Users, Challenges, and Visualization Opportunities. IEEE Trans. Vis. Comput. Graph. 30, 1 (2024), 584–594. https://doi.org/10.1109/TVCG.2023.3326904

  7. [15]

    Harry Li, Gabriel Appleby, and Ashley Suh. 2024. LinkQ: An LLM-Assisted Visual Interface for Knowledge Graph Question-Answering. In 2024 IEEE Visualization and Visual Analytics (VIS) . 116–120. https://doi.org/10.1109/VIS55277.2024.00031

  8. [16]

    Harry Li, Gabriel Appleby, and Ashley Suh. 2024. A Preliminary Roadmap for LLMs as Assistants in Exploring, Analyzing, and Visualizing Knowledge Graphs. arXiv:2404.01425 [cs.HC] https://arxiv.org/abs/2404.01425

  9. [17]

    Dmitry I Mikhailov. 2023. Optimizing national security strategies through llm-driven artificial intelligence integration.arXiv preprint arXiv:2305.13927 (2023)

  10. [18]

    Sadegh M Milajerdi, Birhanu Eshete, Rigel Gjomemo, and VN Venkatakrishnan. 2019. Poirot: Aligning attack behavior with kernel audit records for cyber threat hunting. In Proceedings of the 2019 ACM SIGSAC conference on computer and communications security . 1795–1812

  11. [19]

    Farzad Nourmohammadzadeh Motlagh, Mehrdad Hajizadeh, Mehryar Majd, Pejman Najafi, Feng Cheng, and Christoph Meinel. 2024. Large language models in cybersecurity: State-of-the-art. arXiv preprint arXiv:2402.00891 (2024)

  12. [20]

    Priyanka Sen, Alham Fikri Aji, and Amir Saffari. 2022. Mintaka: A complex, natural, and multilingual dataset for end-to-end question answering. arXiv preprint arXiv:2210.01613 (2022)

  13. [21]

    Yiming Tan, Dehai Min, Yu Li, Wenbo Li, Nan Hu, Yongrui Chen, and Guilin Qi. 2023. Can ChatGPT replace traditional KBQA models? An in-depth analysis of the question answering performance of the GPT LLM family. In International Semantic Web Conference. Springer, 348–367

  14. [22]

    Hao Wei. 2020. Combining knowledge graphs, quickly and accurately. Amazon Science (2020)

  15. [23]

    James Wexler, Mahima Pushkarna, Tolga Bolukbasi, Martin Wattenberg, Fernanda Viégas, and Jimbo Wilson. 2019. The what-if tool: Interactive probing of machine learning models. IEEE Trans. Vis. Comput. Graph. 26, 1 (2019), 56–65

  16. [24]

    Wikipedia. [n. d.]. Wikidata Statistics. https://www.wikidata.org/wiki/Wikidata:Statistics. Accessed: 2024-04-24. 7

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.