Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Matching Game Preferences Through Dialogical Large Language Models: A Perspective

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Coupling GRAPHYP's preference networks with LLMs could make personalized AI transparent and user-controllable, the paper argues.

desk verdict A perspective that names a real gap in graph-LLM personalization but overclaims empirical validation it never provides; the D-LLM idea is worth discussing, not citing as a result. read the letter →

arxiv 2507.20000 v1 pith:JW6NHET2 submitted 2025-07-26 cs.AI cs.DL

Renaud Fabre , Daniel Egret , Patrice Bellot This is my paper
classification cs.AIcs.DL
keywords LLMcustomizationdatapersonalizationvariationalinferencelanguagegamespreferencematchingdialogicalapproachpromptengineeringknowledgegraphs
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 perspective paper argues that today's large language models cannot represent the nuanced, evolving preferences of individual users, and that coupling them with the GRAPHYP knowledge-graph system would fix that. It proposes a "Dialogical LLM" (D-LLM) framework in which GRAPHYP encodes each user's preferences as structured networks of "cognitive communities," and the LLM navigates those networks through explicit graph actions while chatting. The paper claims this hybrid symbolic-neural design makes AI reasoning transparent and user-controllable, reduces hallucination by grounding answers in the graph, and can outperform standalone LLMs or knowledge graphs in personalization tasks. The intended payoff is a conversational AI that not only answers but shows its work and adapts in real time to how a user's preferences change.

What carries the argument

The machinery is the GRAPHYP-LLM coupling: GRAPHYP contributes preference graphs whose nodes and edges encode who likes, dislikes, visited, or was influenced by what, organized into cognitive communities; the LLM contributes natural-language dialogue and the ability to choose graph actions at each reasoning step. Three measured parameters—intensity, variety, attention—define the "language game" a user is playing, and the graph's structure supplies the rules the LLM should follow. Personalized PageRank sampling focuses the system on the user-relevant part of the graph, while variational inference and feedback loops let the preference model update without retraining. Together these pieces are meant to create interactive reasoning loops, dynamic context management, transparent reasoning pathways, and grounded inference.

What would settle it

A direct test would compare the proposed D-LLM against a plain LLM and a plain knowledge graph on the same personalization task, measuring how often the system's answers match stated user preferences and whether its reasoning traces are accurate; if the hybrid shows no measurable gain, or if the traces cannot be verified against the user's actual stated preferences, the central claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid architecture—GRAPHYP's search-experience networks coupled to an LLM—can deliver transparent, user-controlled personalization that neither component achieves alone. GRAPHYP builds "cognitive communities" from search behavior, measuring intensity, variety, and attention, so that the same query can be seen as hosting many valid "language games" tied to different user intentions. In the proposed D-LLM, the LLM drives dialogue by selecting discrete graph actions such as VisitNode or GetSharedNeighbours; every step leaves a visible reasoning trace, and the graph's structure grounds each inference. The paper asserts, as one of its three contributions, empirical validation that hybrid symbolic-neural approaches can outperform standalone systems in personalization tasks, while acknowledging that generalization, scalability, and evaluation frameworks for contested knowledge domains remain open challenges.

Load-bearing premise

The entire framework rests on the claim that GRAPHYP's three-parameter models of search behavior—intensity, variety, and attention—capture stable, meaningful user preferences that an LLM can exploit, a claim the paper takes from earlier GRAPHYP publications and does not revalidate with data here.

Editorial extensions

If this is right

  • A D-LLM would let users see and audit exactly which preference nodes and graph paths shaped a given answer, making personalization inspectable.
  • Grounding each reasoning step in the graph structure should cut hallucination rates on multi-hop or knowledge-intensive queries.
  • Because preferences live in the graph rather than in the model weights, profiles could be updated in real time without expensive retraining.
  • Community-based personalization would let one user's choices be informed by detected cognitive communities of similar users.
  • The same architecture could map scientific disputes by surfacing competing reasoning paths rather than a single consensus answer.

Reading between the lines

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

  • A natural test of the framework would be to run the proposed D-LLM against a plain LLM and a plain knowledge-graph recommender on a public conversational recommendation benchmark, measuring both preference alignment and whether the reasoning traces actually match the user's stated preferences.
  • The "language games" idea could be operationalized as instruction-tuning targets: one can imagine training an LLM to switch between game-specific response styles based on the three preference parameters, a concrete and testable extension the paper leaves implicit.
  • The transparency promise carries an implicit requirement the paper does not address: users must be able to read and verify a graph trace, so human-factors studies of trace comprehension would be needed before the auditability claim is credible.
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 / 5 minor

Summary. This perspective paper proposes a conceptual framework, D-LLM (Dialogical Large Language Models), that couples GRAPHYP's knowledge-graph-based preference modeling with LLMs to deliver transparent, personalized, multi-user conversational AI. The paper describes four core components (interactive reasoning loops, dynamic context management, transparent reasoning pathways, grounded inference) and argues that explicit graph actions and cognitive communities reduce hallucination, improve multi-hop reasoning, and increase user control. It claims three contributions: the D-LLM architecture, transparent personalization, and empirical validation of hybrid superiority, with the empirical claim repeated in Section 4.2 and the Conclusions.

Significance. If the D-LLM framework were validated as promised, it could offer a meaningful step toward human-in-the-loop personalization by making preference signals explicit and auditable. The conceptual synthesis of GRAPHYP's cognitive communities with LLM dialogue is interesting and ties to active research on graph-based grounding, soft prompts, and profile-centric agents. The paper gives credit to prior work on Graphologue, SaBART, and conversational recommendation systems, and its emphasis on dispute modeling and multi-perspective reasoning is a useful perspective. However, the paper's central 'empirical validation' contribution is unsupported: no datasets, baselines, metrics, or protocols appear anywhere in the manuscript. The contribution is therefore currently a hypothesis rather than a demonstrated result, and the architecture's technical components (variational inference, fractal analysis, PPR sampling) are described only at a conceptual level. The significance is conditional on future validation and on the reliability of GRAPHYP's preference models, neither of which is established here.

major comments (4)
  1. [§1.4, §4.2, Tables 3/6/7] The contribution list in Section 1.4 promises 'Empirical Validation: Demonstration that hybrid symbolic-neural approaches can outperform standalone systems in personalization tasks,' but the manuscript reports no data, no baselines, no error bars, and no evaluation protocol. Section 4.2 asserts 'empirical validation across three domains demonstrates encouraging results' without identifying the domains' datasets, tasks, or metrics. Tables 3, 6, and 7 are qualitative capability matrices, not empirical results, and the appendices present hypothetical scenarios. This unsupported claim is load-bearing: it converts a plausible hypothesis into an asserted result, and the Conclusions repeat the same assertion ('we demonstrate empirical validation'). The claim should either be substantiated with real experiments or reframed as a future objective.
  2. [§2.3, §3.1.2] The framework's foundation is GRAPHYP's ability to model stable, meaningful user preferences through cognitive communities, computed from search logs with three parameters (intensity, variety, attention). The paper takes this capability as given from the authors' prior publications (refs 15-17), without presenting any validation data, error analysis, or independent evidence within this manuscript. Several claimed benefits of D-LLM—reduced hallucination, superior multi-hop reasoning, better personalization—rest on the as-yet-unverified premise that GRAPHYP's subgraphs genuinely encode actionable preferences. Section 2.3 states that GRAPHYP 'can model these differences computationally' and 'demonstrated effective preference modeling,' but no results are shown here. Since this premise is central to the entire architecture, the paper should either summarize the supporting evidence from refs 15-17 in sufficient detail for the reader to judge, or present new validation.
  3. [§3.3.2, §3.4.5] The 'variational personalization framework' and 'fractal geometric applications' are named as key components, but they are never formally defined. No equations, loss functions, algorithms, or integration steps are provided for how variational inference updates preference models, how fractal dimensions are computed from LLM embeddings, or how these quantities guide dialogue. Similarly, PPR sampling is described only by its general benefits (Table 5) without a formal definition of the teleportation set or the graph on which it operates. For a paper that claims a 'technical architecture' (Section 3.2), the level of specification is too low to support the claimed capabilities or to enable replication. The authors should either provide formal definitions and pseudocode or clearly label these as open research directions.
  4. [§3.1.3, §3.5.3] The claim that explicit graph actions (VisitNode, GetSharedNeighbours, AnswerQuestion) and reasoning traces are sufficient to make reasoning transparent and to reduce hallucinations is asserted repeatedly, but no evidence is given that users actually understand these traces or that grounded inference statistically lowers hallucination rates in the proposed hybrid. Transparency is a user-centered property that cannot be established by architectural design alone; a user study or an evaluation of trace comprehensibility is needed. Similarly, the factual-consistency advantages in Table 3 ('Enhanced across domains') are presented as inherent by design. These are empirical claims that require experimental support.
minor comments (5)
  1. [§2.3.1 vs. Appendix A.2] The three GRAPHYP preference parameters are inconsistent between sections: Section 2.3.1 lists 'intensity, variety, attention,' while Appendix A.2 lists 'mass (volume of engagement), intensity (depth), and variety (diversity).' Please unify the terminology.
  2. [§3.3.2] The reference for Hausdorff dimension is a non-archival blog URL (numberanalytics.com). Please replace it with a standard textbook or peer-reviewed source, or remove the citation.
  3. [§4.1] The sentence listing 'seven core capabilities' does not enumerate the seven items; consider an explicit list to improve readability.
  4. [§2.3.2] The phrase 'diversity from within' is introduced without a definition or a pointer to where it is formally defined in the GRAPHYP papers; adding a brief explanation would help readers not familiar with refs 15-17.
  5. [Appendix A.3] The climate-change and CRISPR scenarios are useful illustrations, but they are presented as 'use cases' without any data. Consider labeling them as 'hypothetical scenarios' in the heading to avoid confusion with empirical case studies.

Circularity Check

2 steps flagged · score 4.0 of 10

The framework's symbolic side depends on GRAPHYP capabilities that are supported only by the authors' own prior papers; the promised 'empirical validation' is asserted without data, yet the coupling idea itself has independent grounding in the broader graph-LLM literature.

  1. self citation load bearing [Section 1.2 (Motivations), supported by Section 2.3.1 (GRAPHYP Architecture)]
    "The GRAPHYP system (see Section 2.3 below for a short description) demonstrated effective preference modeling for individual users through interpretable subgraph representations [15–17]."

    The D-LLM framework's symbolic component is GRAPHYP, and the paper's claimed contribution 'Empirical Validation: Demonstration that hybrid symbolic-neural approaches can outperform standalone systems' presupposes that GRAPHYP already models preferences effectively. That presupposition is supported exclusively by references [15–17], all authored by the present authors (Fabre, Egret, and Bellot appear in each). The paper supplies no in-paper data, benchmark, or error analysis for GRAPHYP's preference-modeling capability, so the hybrid's expected advantage is inherited from a self-citation chain rather than established independently in this manuscript.

  2. self citation load bearing [Section 3.3.2 (Fractal Geometric Applications), reflected in Table 3]
    "GRAPHYP uses fractal-based analysis (examining patterns like those found in Mandelbrot sets) to distinguish between different network states [17] (such as rest versus active tasks in neural data)."

    Table 3 advertises 'Fractal Analysis' as 'Strong' for GRAPHYP and 'Enhanced with semantic integration' for the hybrid, and Section 3.3.2 claims that incorporating this fractal analysis into D-LLM 'creates a flexible and interpretable framework.' The capability is attributed to reference [17], another prior paper by the same authors, without any independent validation in the present text. This is a second load-bearing self-citation for one of the hybrid system's advertised reasoning advantages.

full rationale

This is a perspective paper with no formal derivation equations, no fitted parameters, and no statistical predictions, so the classic construction-circularity patterns (self-definitional equivalence, fitted-input-called-prediction, imported uniqueness theorem, ansatz smuggled via citation, renaming a known result) do not apply. The only genuine circularity concern is load-bearing self-citation: the D-LLM architecture is built entirely on GRAPHYP, and GRAPHYP's 'demonstrated effective preference modeling' and fractal-analysis capabilities are justified solely by the authors' own prior publications (refs 15–17). Because the paper's stated contribution includes an 'Empirical Validation' claim, and Section 4.2 asserts 'Empirical validation across three domains demonstrates encouraging results' without reporting any dataset, metric, baseline, or statistical test, the central empirical claim is unsupported rather than demonstrated. That is an evidentiary gap and correctness risk, not a circular reduction, so it does not by itself raise the circularity score. The broader conceptual proposal—coupling structured preference graphs with LLMs for transparent, user-controllable personalization—is grounded in the external literature the paper cites (e.g., Graphologue, GraphTranslator, Apollonion, COMPASS), so the central idea retains independent content. On balance, the circularity score is moderate: the framework's load-bearing component rests on a self-citation chain, but the framework itself is a perspective rather than a derivation whose output equals its input.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The paper introduces no fitted parameters. It relies on the prior GRAPHYP system (self-cited), on ad hoc assumptions about the benefits of explicit reasoning traces and fractal analysis, and on a proposed D-LLM architecture that is not implemented.

assumptions (5)
  • domain assumption GRAPHYP's cognitive communities faithfully capture stable user preferences from search logs (intensity, variety, attention).
    Invoked in Sec. 2.3 and throughout; no validation data provided in this paper, relies on refs 15-17.
  • ad hoc to paper Explicit graph actions and reasoning traces are sufficient to make AI reasoning transparent and reduce hallucinations.
    Claimed in Sec. 3.1.3 and Table 3 without measurement.
  • domain assumption Wittgenstein's language games provide a productive computational model for personalization.
    Used in Sec. 3.1.2 and 3.4.3; conceptual, not operational.
  • ad hoc to paper Fractal geometric analysis of LLM embeddings reveals structures useful for adaptivity.
    Sec. 3.3.2 cites examples but provides no evidence for GRAPHYP's fractal capabilities.
  • standard math Personalized PageRank focuses attention on relevant graph regions in GRAPHYP.
    Sec. 3.4.2; PPR is established, but its benefit in GRAPHYP is assumed.
invented entities (1)
  • D-LLM (Dialogical Large Language Model) framework
    purpose: Proposed new architecture coupling GRAPHYP knowledge graphs with LLMs for personalized, transparent dialogue.
    No implementation or falsifiable prediction is provided; it exists only as a proposal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Matching Game Preferences Through Dialogical Large Language Models: A Perspective." pith.science (2026). https://pith.science/paper/JW6NHET2

@misc{pith2026250720000,
  author       = {Pith},
  title        = {Pith review of: Matching Game Preferences Through Dialogical Large Language Models: A Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JW6NHET2}},
  note         = {Machine review of arXiv:2507.20000}
}
read the original abstract

This perspective paper explores the future potential of "conversational intelligence" by examining how Large Language Models (LLMs) could be combined with GRAPHYP's network system to better understand human conversations and preferences. Using recent research and case studies, we propose a conceptual framework that could make AI rea-soning transparent and traceable, allowing humans to see and understand how AI reaches its conclusions. We present the conceptual perspective of "Matching Game Preferences through Dialogical Large Language Models (D-LLMs)," a proposed system that would allow multiple users to share their different preferences through structured conversations. This approach envisions personalizing LLMs by embedding individual user preferences directly into how the model makes decisions. The proposed D-LLM framework would require three main components: (1) reasoning processes that could analyze different search experiences and guide performance, (2) classification systems that would identify user preference patterns, and (3) dialogue approaches that could help humans resolve conflicting information. This perspective framework aims to create an interpretable AI system where users could examine, understand, and combine the different human preferences that influence AI responses, detected through GRAPHYP's search experience networks. The goal of this perspective is to envision AI systems that would not only provide answers but also show users how those answers were reached, making artificial intelligence more transparent and trustworthy for human decision-making.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 31 canonical work pages

  1. [1]

    Human Evaluation of Procedural Knowledge Graph Extraction from Text with Large Language Models

    Carriero, V.A.; Azzini, A.; Baroni, I.; Scrocca, M.; Celino, I. Human Evaluation of Procedural Knowledge Graph Extraction from Text with Large Language Models. In Knowledge Engineering and Knowledge Management. EKAW 2024; Alam, M., Rospocher, M., van Erp, M., Hollink, L., Gesese, G.A.; Eds.; Lecture Notes in Computer Science; Springer: Cham, Switzerland, ...

    work page doi:10.1007/978-3-031-77792-9_26 2024

  2. [2]

    Epistemic GraphText: Graph Reasoning in Text Space

    Zhao, J.; Zhuo, L.; Shen, Y.; Qu, M.; Liu, K.; Bronstein, M.; Zhu, Z.; Tang, J. Epistemic GraphText: Graph Reasoning in Text Space. arXiv 2023, arXiv:2310.01089. https://doi.org/10.48550/arXiv.2310.01089

    arXiv doi:10.48550/arxiv.2310.01089 2023

  3. [3]

    Personaliza- tion of Large Language Models: A Survey

    Zhang, Z.; Rossi, R.A.; Kveton, B.; Shao, Y.; Yang, D.; Zamani, H.; Dernoncourt, F.; Barrow, J.; Yu, T.; Kim, S.; et al. Personaliza- tion of Large Language Models: A Survey. arXiv 2025, arXiv:2411.00027. https://doi.org/10.48550/arXiv.2411.00027

    arXiv doi:10.48550/arxiv.2411.00027 2025

  4. [4]

    Artificial intelligence and illusions of understanding in scientific research

    Messeri, L.; Crockett, M.J. Artificial intelligence and illusions of understanding in scientific research. Nature 2024, 627, 49–58. https://doi.org/10.1038/s41586-024-07146-0

    doi:10.1038/s41586-024-07146-0 2024

  5. [5]

    The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity

    Shojaee, P.; Mirzadeh, I.; Alizadeh, K.; Horton, M.; Bengio, S.; Farajtabar, M. The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. arXiv 2025, arXiv:2506.06941. https://doi.org/10.48550/arXiv.2506.06941

    arXiv doi:10.48550/arxiv.2506.06941 2025

  6. [6]

    Aligning Language Models with Demonstrated Feedback; arXiv 2024, arXiv:2406.00888

    Shaikh, O.; Lam, M.S.; Hejna, J.; Shao, Y.; Cho, H.; Bernstein, M.S.; Yang, D. Aligning Language Models with Demonstrated Feedback; arXiv 2024, arXiv:2406.00888. https://doi.org/10.48550/arXiv.2406.00888

    arXiv doi:10.48550/arxiv.2406.00888 2024

  7. [7]

    On the Measure of Intelligence

    Chollet, F. On the Measure of Intelligence. arXiv 2019, arXiv:1911.01547. https://doi.org/10.48550/arXiv.1911.01547

    arXiv doi:10.48550/arxiv.1911.01547 2019

  8. [8]

    When large language models meet personalization: Perspectives of challenges and opportunities

    Chen, J.; Liu, Z.; Huang, X.; Wu, C.; Liu, Q.; Jiang, G.; Pu, Y.; Lei, Y.; Chen, X.; Wang, X.; et al. When large language models meet personalization: Perspectives of challenges and opportunities. World Wide Web 2024, 27, 42. https://doi.org/10.1007/s11280- 024-01276-1

    doi:10.1007/s11280- 2024

Show all 63 references
  1. [9]

    Human- in-the-Loop Integration with Domain-Knowledge Graphs for Explainable Federated Deep Learning

    Holzinger, A.; Saranti, A.; Hauschild, A.-C.; Beinecke, J.; Heider, D.; Roettger, R.; Mueller, H.; Baumbach, J.; Pfeifer, B. Human- in-the-Loop Integration with Domain-Knowledge Graphs for Explainable Federated Deep Learning. In Machine Learning and Knowledge Extraction. CD‐MA...

    2023

  2. [10]

    Assessing LLMs Suitability for Knowledge Graph Completion

    Iga, V.I.R.; Silaghi, G.C. Assessing LLMs Suitability for Knowledge Graph Completion. In Neural‐Symbolic Learning and Reason‐ ing. NeSy 2024; Besold, T.R., d’Avila Garcez, A., Jimenez-Ruiz, E., Confalonieri, R., Madhyastha, P., Wagner, B.; Eds.; Lecture Notes in Computer Scien...

    2024 doi

  3. [11]

    Unveiling User Preferences: A Knowledge Graph and LLM-Driven Approach for Conver- sational Recommendation

    Qiu, Z.; Luo, L.; Pan, S.; Liew, A.W. Unveiling User Preferences: A Knowledge Graph and LLM-Driven Approach for Conver- sational Recommendation. arXiv 2024, arXiv:2411.14459. https://doi.org/10.48550/arXiv.2411.14459

    2024 doi

  4. [12]

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

    Bai, Y.; Jones, A.; Ndousse, K.; Askell, A.; Chen, A.; Dassarma, N.; Drain, D.; Fort, S.; Ganguli, D.; Henighan, T.J et al. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv 2022, arXiv:2204.05862. https://doi.org/10.48550/arXiv.2...

    2022 doi arXiv

  5. [13]

    A short survey: Exploring knowledge graph-based neural-symbolic system from application perspective

    Zhu, S.; Sun, S. A short survey: Exploring knowledge graph-based neural-symbolic system from application perspective. arXiv 2024, arXiv:2405.03524. https://doi.org/10.48550/arXiv.2405.03524

    2024 doi arXiv

  6. [14]

    A Very Brief Introduction to Machine Learning with Applications to Communication Systems

    Simeone, O. A Very Brief Introduction to Machine Learning with Applications to Communication Systems. IEEE Trans. Cogn. Commun. Netw. 2018, 4, 648–664. https://doi.org/10.1109/TCCN.2018.2881442

    2018

  7. [15]

    Retrieving Adversarial Cliques in Cognitive Communities: A New Conceptual Framework for Scientific Knowledge Graphs

    Fabre, R.; Azeroual, O.; Bellot, P.; Schöpfel, J.; Egret, D. Retrieving Adversarial Cliques in Cognitive Communities: A New Conceptual Framework for Scientific Knowledge Graphs. Future Internet 2022, 14, 262. https://doi.org/10.3390/fi14090262

    2022 doi

  8. [16]

    A Multiverse Graph to Help Scientific Reasoning from Web Usage: Interpretable Patterns of Assessor Shifts in GRAPHYP

    Fabre, R.; Azeroual, O.; Schöpfel, J.; Bellot, P.; Egret, D. A Multiverse Graph to Help Scientific Reasoning from Web Usage: Interpretable Patterns of Assessor Shifts in GRAPHYP. Future Internet 2023, 15, 147. https://doi.org/10.3390/fi15040147

    2023 doi

  9. [17]

    Challenging Scientific Categorizations Through Dispute Learning

    Fabre, R.; Bellot, P.; Egret, D. Challenging Scientific Categorizations Through Dispute Learning. Appl. Sci. 2025, 15, 2241. https://doi.org/10.3390/app15042241

    2025 doi

  10. [18]

    Think Only When You Need with Large Hybrid-Reasoning Models

    Jiang, L.; Wu, X.; Huang, S.; Dong, Q.; Chi, Z.; Dong, L.; Zhang, X.; Lv, T.; Cui, L.; Wei, F. Think Only When You Need with Large Hybrid-Reasoning Models. arXiv 2025, arXiv:2505.14631. https://doi.org/10.48550/arXiv.2505.14631

    2025 doi arXiv

  11. [19]

    Law-Following AI: Designing AI Agents to Obey Human Laws

    O’Keefe, C.; Ramakrishnan, K.; Tay, J.; Winter, C. Law-Following AI: Designing AI Agents to Obey Human Laws. Fordham Law Rev. 2025, 94. Available online: https://law-ai.org/law-following-ai/ (accessed on 20 June 2025)

    2025

  12. [20]

    The benefits and dangers of anthropomorphic conversational agents

    Peter, S.; Riemer, K.; West, J.D. The benefits and dangers of anthropomorphic conversational agents. Proc. Natl. Acad. Sci. USA 2025, 122, e2415898122. https://doi.org/10.1073/pnas.2415898122

    2025 doi

  13. [21]

    User Modeling in the Era of Large Language Models: Current Research and Future Directions

    Tan, Z.; Jiang, M. User Modeling in the Era of Large Language Models: Current Research and Future Directions. arXiv 2023, arXiv:2312.11518. https://doi.org/10.48550/arXiv.2312.11518

    2023 doi arXiv

  14. [22]

    The Method of Language-Games as a Method of Logic

    Kuusela, O. The Method of Language-Games as a Method of Logic. In Wittgenstein on Logic as the Method of Philosophy: Re‐ examining the Roots and Development of Analytic Philosophy, online ed.; Oxford Academic: Oxford, UK, 2019. https://doi.org/10.1093/oso/9780198829751.003.0005

    2019

  15. [23]

    Epistemic Alignment: A Mediating Framework for User-LLM Knowledge Delivery

    Clark, N.; Shen, H.; Howe, W.; Mitra, T. Epistemic Alignment: A Mediating Framework for User-LLM Knowledge Delivery. arXiv 2025, arXiv:2504.01205. https://doi.org/10.48550/arXiv.2504.01205

    2025 doi arXiv

  16. [24]

    An Introduction to Variational Methods for Graphical Models

    Jordan, M.I.; Ghahramani, Z.; Jaakkola, T.S.; Saul, L.K. An Introduction to Variational Methods for Graphical Models. In Learning in Graphical Models; Jordan, M.I.; Ed.; NATO ASI Series; Springer: Dordrecht, The Netherlands, 1998; Volume 89, pp. 105–161. https://doi.org/10.100...

    1998 doi

  17. [25]

    The pivot penalty in research

    Hill, R.; Yin, Y.; Stein, C.; Wang, X.; Wang, D.; Jones, B.F. The pivot penalty in research. Nature 2025, 642, 999–1006. https://doi.org/10.1038/s41586-025-09048-1

    2025 doi

  18. [26]

    Graphologue: Exploring Large Language Model Responses with Interactive Diagrams

    Jiang, P.; Rayan, J.; Dow, S.P.; Xia, H. Graphologue: Exploring Large Language Model Responses with Interactive Diagrams. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ‘23), San Francisco, CA, USA, 29 October–1 November 2023; A...

    2023

  19. [27]

    Enhancing Dialogue Generation via Dynamic Graph Knowledge Aggregation

    Tang, C.; Zhang, H.; Loakman, T.; Lin, C.; Guerin, F. Enhancing Dialogue Generation via Dynamic Graph Knowledge Aggregation. arXiv 2023, arXiv:2306.16195. https://doi.org/10.48550/arXiv.2306.16195

    2023 doi arXiv

  20. [28]

    Building knowledge-grounded dialogue systems with graph-based semantic modelling

    Yang, Y.; Huang, H.; Gao, Y.; Li, J. Building knowledge-grounded dialogue systems with graph-based semantic modelling. Knowl.‐Based Syst. 2024, 298, 11943. https://doi.org/10.1016/j.knosys.2024.111943

    2024

  21. [29]

    LLM Inference Enhanced by External Knowledge: A Survey

    Lin, Y.; Chen, Q.; Cheng, Y.; Zhang, J.; Liu, Y.; Hsia, L.; Chen, Y. LLM Inference Enhanced by External Knowledge: A Survey. arXiv 2025, arXiv:2505.24377. https://doi.org/10.48550/arXiv.2505.24377. Appl. Sci. 2025, 15, 8307 28 of 29

    2025 doi arXiv

  22. [30]

    Research Trends for the Interplay between Large Language Models and Knowledge Graphs

    Khorashadizadeh, H.; Zahra Amara, F.; Ezzabady, M.; Ieng, F.; Tiwari, S.; Mihindukulasooriya, N.; Groppe, J.; Sahri, S.; Be- namara, F.; Groppe, S. Research Trends for the Interplay between Large Language Models and Knowledge Graphs. arXiv 2024, arXiv:2406.08223. https://doi.o...

    2024 doi arXiv

  23. [31]

    LLM-Powered Knowledge Graphs for Enterprise Intelligence and Analytics

    Kumar, R.; Ishan, K.; Kumar, H.; Singla, A. LLM-Powered Knowledge Graphs for Enterprise Intelligence and Analytics. arXiv 2025, arXiv:2503.07993. https://doi.org/10.48550/arXiv.2503.07993

    2025 doi arXiv

  24. [32]

    Challenging Machines to the Language Game: Wittgensteinian Philosophy and Future Dimensions of Artificial Intel- ligence

    Chen, N. Challenging Machines to the Language Game: Wittgensteinian Philosophy and Future Dimensions of Artificial Intel- ligence. Fudan J. Hum. Soc. Sci. 2015, 8, 487–500. https://doi.org/10.1007/s40647-015-0085-8

    2015 doi

  25. [33]

    Philosophical Investigations into AI Alignment: A Wittgensteinian Framework

    Pérez-Escobar, J.A.; Sarikaya, D. Philosophical Investigations into AI Alignment: A Wittgensteinian Framework. Philos. Technol. 2024, 37, 1–25. https://doi.org/10.1007/s13347-024-00761-9

    2024 doi

  26. [34]

    Essai d’une Théorie des Langues Spéciales

    Van Gennep, A. Essai d’une Théorie des Langues Spéciales. 1908. Available online: https://elianedaphy.org/IMG/pdf/VanGennep_1908_LanguesSpeciales.pdf (accessed on 20 June 2025)

    1908

  27. [35]

    Tree-of-Traversals: A Zero-Shot Reasoning Algorithm for Black-box Language Models with Knowledge Graphs

    Markowitz, E.; Ramakrishna, A.; Dhamala, J.; Mehrabi, N.; Peris, C.; Gupta, R.; Chang, K.; Galstyan, A. Tree-of-Traversals: A Zero-Shot Reasoning Algorithm for Black-box Language Models with Knowledge Graphs. In Proceedings of the 62nd Annual Meeting of the Association for Com...

    2024

  28. [36]

    GraphOTTER: Evolving LLM-based Graph Reasoning for Complex Table Question Answering

    Li, Q.; Huang, C.; Li, S.; Xiang, Y.; Xiong, D.; Lei, W. GraphOTTER: Evolving LLM-based Graph Reasoning for Complex Table Question Answering. In Proceedings of the 31st International Conference on Computational Linguistics, Abu Dhabi, United Arab Emirates, 19–24 January 2025; ...

    2025

  29. [37]

    GraphReason: Enhancing Reasoning Capabilities of Large Language Models through A Graph-Based Verification Ap- proach

    Cao, L. GraphReason: Enhancing Reasoning Capabilities of Large Language Models through A Graph-Based Verification Ap- proach. In Proceedings of the 2nd Workshop on Natural Language Reasoning and Structured Explanations, Bangkok, Thailand, 11–16 August 2024; Association for Com...

    2024

  30. [38]

    Graph-based Approaches and Functionalities in Retrieval-Augmented Generation: A Comprehensive Survey

    Zhu, Z.; Huang, T.; Wang, K.; Ye, J.; Chen, X.; Luo, S. Graph-based Approaches and Functionalities in Retrieval-Augmented Generation: A Comprehensive Survey. arXiv 2025, arXiv:2504.10499. https://doi.org/10.48550/arXiv.2504.10499

    2025 doi

  31. [39]

    Unlocking the Potential of Generative AI through Neuro-Symbolic Architectures: Benefits and Limitations

    Bougzime, O.; Jabbar, S.; Cruz, C.; Demoly, F. Unlocking the Potential of Generative AI through Neuro-Symbolic Architectures: Benefits and Limitations. arXiv 2025, arXiv:2502.11269. https://doi.org/10.48550/arXiv.2502.11269

    2025 doi arXiv

  32. [40]

    Besta, M.; Memedi, F.; Zhang, Z.; Gerstenberger, R.; Piao, G.; Blach, N.; Nyczyk, P.; Copik, M.; Kwaśniewski, G.; Muller, J. et al. Demystifying Chains, Trees, and Graphs of Thoughts. arXiv 2024, arXiv:2401.14295. https://doi.org/10.48550/arXiv.2401.14295

    2024 doi

  33. [41]

    LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities

    Zhu, Y.; Wang, X.; Chen, J.; Qiao, S.; Ou, Y.; Yao, Y.; Deng, S.; Chen, H.; Zhang, N. LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities. arXiv 2023, arXiv:2305.13168. https://doi.org/10.48550/arXiv.2305.13168

    2023 doi arXiv

  34. [42]

    On the fractal patterns of language structures

    Ribeiro, L.C.; Bernardes, A.T.; Mello, H. On the fractal patterns of language structures. PLoS ONE 2023, 18, e0285630. https://doi.org/10.1371/journal.pone.0285630

    2023 doi

  35. [43]

    Fractal Patterns May Illuminate the Success of Next-Token Prediction

    Alabdulmohsin, I.; Tran, V.Q.; Dehghani, M. Fractal Patterns May Illuminate the Success of Next-Token Prediction. arXiv 2024, arXiv:2402.01825. https://doi.org/10.48550/arXiv.2402.01825

    2024 doi arXiv

  36. [44]

    Reason-Align-Respond: Aligning LLM Reasoning with Knowledge Graphs for KGQA

    Shen, X.; Wang, F.; Xia, R. Reason-Align-Respond: Aligning LLM Reasoning with Knowledge Graphs for KGQA. arXiv 2025, arXiv:2505.20971. https://doi.org/10.48550/arXiv.2505.20971

    2025 doi arXiv

  37. [45]

    Hydra: Structured Cross-Source Enhanced Large Language Model Reasoning

    Tan, X.; Wang, X.; Liu, Q.; Xu, X.; Yuan, X.; Zhu, L.; Zhang, W. Hydra: Structured Cross-Source Enhanced Large Language Model Reasoning. arXiv 2025, arXiv:2505.17464. https://doi.org/10.48550/arXiv.2505.17464

    2025 doi

  38. [46]

    A Hybrid RAG System with Comprehensive Enhancement on Complex Reasoning

    Yuan, Y.; Liu, C.; Yuan, J.; Sun, G.; Li, S.; Zhang, M. A Hybrid RAG System with Comprehensive Enhancement on Complex Reasoning. arXiv 2024, arXiv:2408.05141. https://doi.org/10.48550/arXiv.2408.05141

    2024 doi arXiv

  39. [47]

    A Graph-based Approach for Adaptive Serious Games

    Hare, R.; Patel, N.; Tang, Y.; Patel, P. A Graph-based Approach for Adaptive Serious Games. In Proceedings of the IEEE Intl Conference on Dependable, Autonomic and Secure Computing, International Conference on Pervasive Intelligence and Com- puting, International Conference on...

    2022

  40. [48]

    Efficient Algorithms for Personalized PageRank Computation: A Survey

    Yang, M.; Wang, H.; Wei, Z.; Wang, S.; Wen, J.R. Efficient Algorithms for Personalized PageRank Computation: A Survey. In IEEE Trans. Knowl. Data Eng. 2024, 36, 4582-4602. https://doi.org/10.1109/TKDE.2024.3376000. Appl. Sci. 2025, 15, 8307 29 of 29

    2024

  41. [49]

    The BS-meter: A ChatGPT-Trained Instrument to Detect Sloppy Language-Games

    Trevisan, A. The BS-meter: A ChatGPT-Trained Instrument to Detect Sloppy Language-Games. arXiv 2024, arXiv:2411.15129. https://doi.org/10.48550/arXiv.2411.15129

    2024 doi arXiv

  42. [50]

    Injecting Knowledge Graphs into Large Language Models

    Coppolillo, E. Injecting Knowledge Graphs into Large Language Models. arXiv 2025, arXiv:2505.07554. https://doi.org/10.48550/arXiv.2505.07554

    2025 doi arXiv

  43. [51]

    GENEVA: GENErating and Visualizing branching narratives using LLMs

    Leandro, J.; Rao, S.; Xu, M.; Xu, W.; Jojic, N.; Brockett, C.; Dolan, B. GENEVA: GENErating and Visualizing branching narratives using LLMs. In Proceedings of the 2024 IEEE Conference on Games (CoG), Milan, Italy, 5–8 August 2024; pp. 1–5. https://doi.org/10.1109/CoG60054.2024...

    2024

  44. [52]

    Personalized Graph-Based Retrieval for Large Language Models

    Au, S.; Dimacali, C.J.; Pedirappagari, O.; Park, N.; Dernoncourt, F.; Wang, Y.; Kanakaris, N.; Deilamsalehy, H.; Rossi, R.A.; Ahmed, N.K. Personalized Graph-Based Retrieval for Large Language Models. arXiv 2025, arXiv:2501.02157. https://doi.org/10.48550/arXiv.2501.02157

    2025 doi arXiv

  45. [53]

    A Graph Neural Network Reasoner for Game Description Language

    Gunawan, A.; Ruan, J.; Huang, X. A Graph Neural Network Reasoner for Game Description Language. In Proceedings of the 19th International Conference on Principles of Knowledge Representation and Reasoning—Special Session on KR and Machine Learning, Haifa, Israel, 31 July–5 Augu...

    2022

  46. [54]

    Integrating Graphs with Large Language Models: Methods and Prospects

    Pan, S.; Zheng, Y.; Liu, Y. Integrating Graphs with Large Language Models: Methods and Prospects. IEEE Intell. Syst. 2024, 39, 64–68. https://doi.org/10.1109/MIS.2023.3332242

    2024

  47. [55]

    GraphTranslator: Aligning Graph Model to Large Language Model for Open-ended Tasks

    Zhang, M.; Sun, M.; Wang, P.; Fan, S.; Mo, Y.; Xu, X.; Liu, H.; Yang, C.; Shi, C. GraphTranslator: Aligning Graph Model to Large Language Model for Open-ended Tasks. In Proceedings of the ACM Web Conference 2024 (WWW ‘24), Singapore, 13–17 May 2024; Association for Computing M...

    2024

  48. [56]

    Apollonion: Profile-centric Dialog Agent

    Chen, S.; Zhao, Z.; Zhao, Y.; Li, X. Apollonion: Profile-centric Dialog Agent. arXiv 2024, arXiv:2404.08692. https://doi.org/10.48550/arXiv.2404.08692

    2024 doi arXiv

  49. [57]

    Aligning LLMs with Individual Preferences via Interaction

    Wu, S.; Fung, M.; Qian, C.; Kim, J.; Hakkani-Tur, D.; Ji, H. Aligning LLMs with Individual Preferences via Interaction. arXiv 2024, arXiv:2410.03642. https://doi.org/10.48550/arXiv.2410.03642

    2024 doi arXiv

  50. [58]

    Combining Knowledge Graphs and Large Language Models

    Kau, A.; He, X.; Nambissan, A.; Astudillo, A.; Yin, H.; Aryani, A. Combining Knowledge Graphs and Large Language Models. arXiv 2024, arXiv:2407.06564. https://doi.org/10.48550/arXiv.2407.06564

    2024 doi arXiv

  51. [59]

    Bias in Large Language Models: Origin, Evaluation, and Mitigation

    Guo, Y.; Guo, M.; Su, J.; Yang, Z.; Zhu, M.; Li, H.; Qiu, M.; Liu, S.S. Bias in Large Language Models: Origin, Evaluation, and Mitigation. arXiv 2024, arXiv:2411.10915. https://doi.org/10.48550/arXiv.2411.10915

    2024 doi arXiv

  52. [60]

    Trust, Accountability, and Autonomy in Knowledge Graph-based AI for Self-determination

    Ibáñez, L.-D.; Domingue, J.; Kirrane, S.; Seneviratne, O.; Third, A.; Vidal, M.-E. Trust, Accountability, and Autonomy in Knowledge Graph-based AI for Self-determination. arXiv 2023, arXiv:2310.19503. https://doi.org/10.48550/arXiv.2310.19503

    2023 doi arXiv

  53. [61]

    Zhao, X.; Blum, M.; Yang, R.; Yang, B.; Márquez Carpintero, L.; Pina-Navarro, M.; Wang, T.; Li, X.; Li, H.; Fu, Y. et al. AGENTi- Graph: An Interactive Knowledge Graph Platform for LLM-based Chatbots Utilizing Private Data. arXiv 2024, arXiv:2410.11531. https://doi.org/10.4855...

    2024 doi arXiv

  54. [62]

    Building Knowledge from Interactions: An LLM-Based Architecture for Adaptive Tutoring and Social Reasoning

    Garello, L.; Belgiovine, G.; Russo, G.; Rea, F.; Sciutti, A. Building Knowledge from Interactions: An LLM-Based Architecture for Adaptive Tutoring and Social Reasoning. arXiv 2025, arXiv:2504.01588. https://doi.org/10.48550/arXiv.2504.01588

    2025 doi arXiv

  55. [63]

    Dynamic Knowledge Exchange and Dual-diversity Review: Concisely Unleashing the Potential of a Multi-Agent Research Team

    Yu, W.; Tang, S.; Huang, Y.; Dong, N.; Fan, L.; Qi, H.; Liu, W.; Diao, X.; Chen, X.; Ouyang, W. Dynamic Knowledge Exchange and Dual-diversity Review: Concisely Unleashing the Potential of a Multi-Agent Research Team. arXiv 2025, arXiv:2506.18348. https://doi.org/10.48550/arXiv...

    2025 doi arXiv

Pith tools

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