Pith. sign in

REVIEW 3 major objections 7 minor 91 references

Flow-Modulated Scoring for Semantic-Aware Knowledge Graph Completion

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that knowledge graph relations can be modeled as dynamic, context-conditioned flows between entity representations, and that this yields near-perfect scores on standard benchmarks with far fewer parameters than embedding…

desk verdict Plausible idea, but the central scoring equation does not type-check and the graph split is unstated, so the near-perfect results are not supported as written. read the letter →

arxiv 2506.23137 v3 pith:IJVAPJQI submitted 2025-06-29 cs.CL cs.AI

classification cs.CLcs.AI
keywords knowledgegraphcompletionrelationpredictionentityconditionalflowmatchingneuralnetworksemanticcontextlearningrepresentationandreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the meaning of a relation between two entities is jointly determined by the static semantic context surrounding the entities and by a dynamic flow that carries one entity's representation toward the other's. It instantiates this belief in a two-stage model, FMS: a Semantic Context Learning module passes messages between relation edges while keeping only the top-K semantically similar neighbors, and a Conditional Flow-Matching module learns a vector field that evolves the head representation into the tail representation and multiplies it into a static plausibility score. The model reports a 99.8% Mean Reciprocal Rank (MRR) and 99.7% Hits@1 on FB15k-237 relation prediction using 0.35M parameters, a 99.9% MRR on WN18RR, and a 25.2% relative MRR gain over the strongest baseline on FB15k-237 entity prediction. If these results hold, they imply that most relation labels in a knowledge graph are nearly determined by their relational context, and that per-entity embeddings, the backbone of most existing systems, are unnecessary. The paper itself closes by noting that its performance depends on the semantic quality of relational features, which could constrain it on sparse or ambiguous relation sets.

What carries the argument

The load-bearing object is the flow-modulated score at the meeting point of two modules. The Semantic Context Learning module treats the knowledge graph as a graph of relations rather than entities: every edge's state is updated by aggregating over neighboring edge states, filtered by an energy-style similarity $Score(e_c, e_n) = \exp(-\lVert g(s_{e_c}) - g(s_{e_n})\rVert^2/\tau)$ that keeps only the Top-K most semantically relevant neighbors, then merged through a multi-head attention gate into the central edge's new state; the final messages $m_h, m_t$ feed a static linear score and also serve as the conditioning context $z$ for the flow module. The Conditional Flow-Matching module learns a time-dependent vector field $v_\theta(t,x)$ whose regression target is the straight-line velocity $u_t(x|z) = t^* - h^*$, with an optimal-transport coupling between the head and tail distributions. The mechanism doing the work is the element-wise modulation $s_{(h,t)} \odot v_\theta(t,x)$: the base score supplies static plausibility, the flow rescales it per entity pair, and the joint loss $L = L_{\mathrm{pred}} + \lambda L_{\mathrm{cfm}}$ binds the two stages into one training objective.

What would settle it

Inspect the released code's graph construction for the Semantic Context Learning module: if the adjacency structure is built once from the full dataset (training plus validation plus test) rather than from the training split alone, then a test query $(h, ?, t)$ sees the ground-truth edge $(h, r, t)$ during message passing and the relation label is visible to the scoring path; re-running the evaluation with a strictly training-only context graph and comparing the MRR would settle whether the 99.8% figure reflects learned relation semantics or label leakage.

Watch

Extended reading notes

Core claim

FMS's central claim is that a relation can be modeled as a conditional vector flow from head to tail entity, and that multiplying a static score by this learned flow gives a complete account of relation semantics. The paper decomposes the relation distribution as $p(r|h,t) \propto p(h,t|r)\cdot p(r)$, then splits $p(h,t|r)$ symmetrically into context terms $p(h|r), p(t|r)$ and dynamic terms $p(t|h,r), p(h|t,r)$; the context terms are instantiated by the Semantic Context Learning module and the dynamic terms by the Conditional Flow-Matching module. The static score takes only the entity messages $m_h, m_t$ as input, deliberately excluding their connecting edge because the ground-truth relation is unobserved, while the flow module regresses a velocity field $v_\theta(t,x)$ against the straight-line target $u_t(x|z) = t^* - h^*$ along the interpolated path $p_t(x|z) = \mathcal{N}(x \mid (1-t)h^* + t\, t^*, \sigma^2 I)$; the resulting field modulates the static score as $s_{(h,t)} = s_{(h,t)} \odot v_\theta(t,x)$ before a softmax over relation types, and the whole model is trained jointly on relation cross-entropy plus the flow-matching loss. On its own terms, this unification of static context and dynamic evolution is what lets FMS report state-of-the-art transductive and inductive results across six relation-prediction datasets and four entity-prediction datasets while learning no entity embeddings.

Load-bearing premise

The near-perfect results rest on the assumption that the graph used by the Semantic Context Learning module when scoring test queries contains only training edges, so a test pair's true relation never reaches the entity messages through message passing; the paper does not state that this split is enforced, and if it is not, the task becomes trivial.

Editorial extensions

If this is right

  • If the reported numbers hold, relation labels on standard benchmarks are nearly determined by local relational context: the correct relation is ranked first in more than 99% of queries on FB15k-237 and WN18RR, far beyond any previous system.
  • Entity embeddings become unnecessary: FMS learns no entity-specific features, so its parameter count is 0.35M on FB15k-237 against 5.9M for TransE and DistMult and 1.67M for PathCon, making the model scale with the number of relations rather than entities.
  • The two-stage design transfers beyond relation prediction: on FB15k-237 entity prediction it reports a 25.2% relative MRR gain over the strongest baseline, and in inductive settings with unseen entities it outperforms all rule-based, GNN-based, and diffusion-based baselines on nearly every split.
  • The ablation study attributes most of the gain to the energy-based Top-K context selection (removing it drops MRR from 99.8 to 97.8 on FB15k-237 and from 99.9 to 94.3 on WN18RR), while removing the flow-matching module drops it to 98.7 and 98.4; the paper concludes the two components work synergistically.

Reading between the lines

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

  • The paper's own ablation places most of the signal in the context-selection stage, which raises a question it does not test: how much of the near-perfect score a far simpler baseline that memorizes the majority relation per local edge-context pattern would already capture.
  • The flow-modulation step is a wrapper around the base score rather than something tied to the particular linear scorer FMS uses; the same Conditional Flow-Matching modulation could be layered onto stronger base scorers such as RotatE or ComplEx, a combination the paper does not evaluate.
  • The pseudo-time variable $t$ invites a direct extension to temporal knowledge graphs: where real timestamps exist, the learned flow could be re-parameterized by actual time, turning the evolution metaphor into a genuine temporal relation model.
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

3 major / 7 minor

Summary. The paper proposes Flow-Modulated Scoring (FMS), a knowledge-graph-completion framework with two components: a Semantic Context Learning module that performs edge-level message passing with Top-K semantic selection to produce context-aware entity messages, and a Conditional Flow-Matching module that learns a vector field v_θ between head and tail entity embeddings. The final score is obtained by elementwise-modulating the static score with the predicted flow, and the model is trained jointly with a cross-entropy prediction loss and a conditional-flow-matching loss. The paper reports state-of-the-art results on relation prediction (e.g., 99.8% MRR on FB15k-237 and 99.9% MRR on WN18RR), strong entity-prediction results, and competitive inductive results, with a very small parameter count.

Significance. If the proposed mechanism were sound and the results reproducible, FMS would be a significant contribution: it offers a novel conceptual connection between conditional flow matching and relation scoring, a parameter-efficient architecture that avoids explicit entity embeddings, and a promising direction for context-sensitive relation reasoning. The paper also releases code, includes ablations of the key components, and performs hyperparameter sensitivity analysis, all of which are commendable. However, the significance is conditional: the core scoring equation as written is dimensionally inconsistent, the relation-prediction protocol does not specify whether validation/test edges are excluded from the message-passing graph, and the entity-prediction adaptation is described too briefly to be reconstructed. These issues directly affect the headline claims and must be resolved before the results can be evaluated.

major comments (3)
  1. [Section 3.2.3, Eq. (20) and Algorithm 1 (line 9)] The core scoring operation is dimensionally inconsistent. Equation (19) defines s_(h,t) as the output of a Linear layer applied to [m_h, m_t]; since Equation (21) applies SoftMax over relation types, s_(h,t) must be a vector of relation logits. The CFM module, however, is defined as v_θ: [0,1] × R^d → R^d (Section 3.2.2), and the experiments use d=64 (Table 2) while the relation counts are 237 for FB15k-237 and 11 for WN18RR. The elementwise product in Eq. (20) is therefore undefined unless d equals the number of relations, which is neither stated nor reflected in the hyperparameters. This is not a notational issue: the mechanism that distinguishes FMS from a static scorer cannot be evaluated from the paper as written.
  2. [Section 3.2.1 (Eqs. 3–5) with Section 3.1 and Section 4.1] The edge set used for Semantic Context Learning is never restricted to training edges. Section 3.1 defines G=(V,E) without specifying whether E includes validation or test triples, and the initial state of an edge is its relation embedding x_e. If the query edge (h,r,t) is present in E at evaluation time, the true relation r enters the messages m_h and m_t through the incident-edge aggregation in Eqs. (3)–(5). The sentence after Eq. (19) only excludes r from the Linear input, not from the message-passing graph. Since Tables 3 and 4 report near-perfect relation-prediction scores, the paper must state explicitly that E contains only training triples (or describe an equivalent masking procedure); otherwise the results are consistent with trivial label leakage rather than relational generalization.
  3. [Section 5, Eq. (24) and the paragraph following it] The entity-prediction adaptation cannot be reconstructed. The text says the model 'must be trained with the relation r as part of the input condition' to predict the target entity, but no forward-pass equation or architectural change shows where r enters: Eq. (19) depends only on m_h and m_t, and v_θ in Eq. (20) is conditioned only on z=(h*,t*). Equation (24) is therefore just a softmax over relation-agnostic scores s_(h,t), which cannot implement p(t|h,r). Without a description of how r conditions either the static score or the flow, the entity-prediction results in Tables 11, 12, and 14 are unsupported.
minor comments (7)
  1. [Section 3.2.3, after Eq. (19)] The statement that the ground-truth relation r is 'treated as unobserved during the training stage' is confusing, because r is observed in training triples; presumably the intended meaning is that r is treated as unobserved when computing the score for a query.
  2. [Section 3.2.2, Theorem 1] The proof for the t=1 boundary is dismissed with the single sentence 'This is also true for t=1,' and the notation alternates between p_t and pt; please expand the proof and make the notation consistent.
  3. [Section 3.2.2, final paragraph] The acronym 'CSM' appears without definition ('Through CSM, model can learn more direct evolution paths between entities'); this should likely be 'CFM'.
  4. [Section 4.4.1, Table 6] The ablation 'w/o Flow-Matching' still achieves 98.7 MRR on FB15k-237, which is close to the full model's 99.8; a discussion of the marginal benefit of the flow-modulation term relative to its additional complexity would strengthen the motivation.
  5. [Section 4.6 and Section 4.7] The correlation heatmaps and t-SNE visualizations are presented as qualitative evidence for specific learned rules and dynamic representations; consider reporting a quantitative support measure or a small user/rule-accuracy study.
  6. [Section 3.2.2, Eq. (16)] The optimal-transport coupling π(z) is defined abstractly, but the paper does not explain how this coupling is computed or approximated during training; please specify the estimator (e.g., minibatch OT or independent sampling) used in Algorithm 1.
  7. [References and formatting] There are several typos and minor inconsistencies, including 'T able' and 'Y et' in the abstract/body, 'NL995' instead of 'NELL995' in Table 2, duplicated 'Baselines' headings in Section 5.2, and a citation to Schrijver's Combinatorial Optimization for optimal transport, which should be replaced with a standard OT reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the flow-modulation term is trained as an auxiliary regression and the relation classifier is trained by cross-entropy; no load-bearing step reduces to its own inputs by the paper's own equations.

full rationale

The FMS derivation chain is self-contained. Section 3.1's Bayes factorization (Eqs. 1-2) is used only as motivation; the semantic context module and the conditional flow-matching module are separately parameterized and jointly trained. The static score (Eq. 19) is a learned linear function of the context messages mh and mt, and the flow-modulated score (Eq. 20) multiplies that score by a vector field vθ. The flow-matching loss (Eq. 15) regresses vθ toward t* - h* with z = (mh, mt), so this term is an internal consistency regularizer rather than a fitted relation predictor. The final relation probability (Eq. 21) is trained with cross-entropy (Eq. 22), not by reusing the flow target as a label. No parameter is fitted to a subset of relation labels and then reported as a prediction of those labels; the relation classifier is trained on the training triplets only. The paper contains no load-bearing self-citation chain and no imported uniqueness theorem; the cited external tools (flow matching, PathCon, RED-GNN) provide standard independent machinery. The only notable concern is an experimental-specification ambiguity: Section 3.1 defines G=(V,E) as the knowledge graph without stating whether E is restricted to training triples for the relation-prediction task, and the note after Eq. 19 excludes the connecting edge r only from the Linear input, not from the message-passing graph. If test edges were included in E, the queries would leak their labels through the context messages. However, the paper nowhere states that test edges are included, so this is a potential data-leakage/correctness risk rather than a circular derivation established by the paper's own equations. Per the hard rule against conditional or speculative circularity, this ambiguity does not raise the circularity score.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. Its free parameters are standard hyperparameters plus an unreported noise scale σ. The model relies on standard flow matching theory (continuity equation, Gaussian conditional paths) and on an ad hoc score modulation whose dimensional consistency is not established.

free parameters (7)
  • Temperature tau = 0.95
    Used in Eq (3) for top-K edge selection; tuned per dataset.
  • Top-K values = 10/10/4/4/3/3 for FB15K, FB15K-237, WN18, WN18RR, NELL995, DDB14
    Number of neighboring edges selected; dataset-specific hyperparameter.
  • CFM loss weight lambda = 1.2
    Balances flow-matching loss and prediction loss in Eq (23).
  • Flow noise sigma = not reported
    Controls Gaussian noise in conditional path Eq (17); value never specified.
  • Dimension = 64
    Embedding dimension for relation states.
  • Context hops = 2 or 3
    Number of message passing layers.
  • Neighbor samples = 8, 16
    Number of sampled neighboring edges.
assumptions (3)
  • standard math Continuity equation and conditional flow matching objective from Tong et al. (2023) hold as stated.
    Used to justify Eq (7) and CFM loss; standard result.
  • domain assumption Bayes decomposition p(r|h,t) ∝ p(h,t|r)p(r) and symmetric factorization Eq (2) are valid for the relation prediction task.
    Motivates the two-module architecture; no derivation that these factors correspond to static context and dynamic evolution.
  • ad hoc to paper The elementwise modulation s ⊙ vθ in Eq (20) is a meaningful operation.
    The static score is a vector over relations while vθ is d-dimensional; the paper provides no rule for how the multiplication is defined.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flow-Modulated Scoring for Semantic-Aware Knowledge Graph Completion." pith.science (2026). https://pith.science/paper/IJVAPJQI

@misc{pith2026250623137,
  author       = {Pith},
  title        = {Pith review of: Flow-Modulated Scoring for Semantic-Aware Knowledge Graph Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJVAPJQI}},
  note         = {Machine review of arXiv:2506.23137}
}
read the original abstract

Knowledge graph completion demands effective modeling of multifaceted semantic relationships between entities. Yet, prevailing methods, which rely on static scoring functions over learned embeddings, struggling to simultaneously capture rich semantic context and the dynamic nature of relations. To overcome this limitation, we propose the Flow-Modulated Scoring (FMS) framework, conceptualizing a relation as a dynamic evolutionary process governed by its static semantic environment. FMS operates in two stages: it first learns context-aware entity embeddings via a Semantic Context Learning module, and then models a dynamic flow between them using a Conditional Flow-Matching module. This learned flow dynamically modulates a base static score for the entity pair. By unifying context-rich static representations with a conditioned dynamic flow, FMS achieves a more comprehensive understanding of relational semantics. Extensive experiments demonstrate that FMS establishes a new state of the art across both canonical knowledge graph completion tasks: relation prediction and entity prediction. On the standard relation prediction benchmark FB15k-237, FMS achieves a near-perfect MRR of 99.8\% and Hits@1 of 99.7\% using a mere 0.35M parameters, while also attaining a 99.9\% MRR on WN18RR. Its dominance extends to entity prediction, where it secures a 25.2\% relative MRR gain in the transductive setting and substantially outperforms all baselines in challenging inductive settings. By unifying a dynamic flow mechanism with rich static contexts, FMS offers a highly effective and parameter-efficient new paradigm for knowledge graph completion. Code published at: https://github.com/yuanwuyuan9/FMS.

Figures

Figures reproduced from arXiv: 2506.23137 by the authors.

Figure 1
Figure 1. The context-dependent nature of relational semantics, exempli [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the dynamic evolution of relationships. The exam [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. The architecture of the Flow-Modulated Scoring (FMS) framework. In Semantic Context Learning Module, nodes(square shapes) do not [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hyper-parameters analysis of FMS on FB15k-237. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The learned correlation between the contextual relations of head/tail entities and the predicted relations. Case studies are shown for the [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: T-SNE visualization of Static Score versus Flow-Modulated Score [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

91 extracted references · 63 canonical work pages

  1. [1]

    Comprehending knowledge graphs with large language models for recommender systems,

    Z. Cui, Y. Weng, X. Tang, F. Lyu, D. Liu, X. He, and C. Ma, “Comprehending knowledge graphs with large language models for recommender systems,” in Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, ser. SIGIR ’25. New York, NY, USA: Association for Computing Machinery, 2025, p. 1229–1239. [Onl...

  2. [2]

    Data scarcity in recom- mendation systems: A survey,

    Z. Chen, W. Gan, J. Wu, K. Hu, and H. Lin, “Data scarcity in recom- mendation systems: A survey,” ACM Transactions on Recommender Systems, vol. 3, no. 3, pp. 1–31, 2025

  3. [3]

    A comprehensive large-scale biomedical knowledge graph for ai-powered data-driven biomed- ical research,

    Y. Zhang, X. Sui, F. Pan, K. Yu, K. Li, S. Tian, A. Erdengasileng, Q. Han, W. Wang, J. Wang et al. , “A comprehensive large-scale biomedical knowledge graph for ai-powered data-driven biomed- ical research,” Nature Machine Intelligence, pp. 1–13, 2025

  4. [4]

    A universal question-answering platform for knowledge graphs,

    R. Omar, I. Dhall, P . Kalnis, and E. Mansour, “A universal question-answering platform for knowledge graphs,” Proceedings of the ACM on Management of Data , vol. 1, no. 1, pp. 1–25, 2023

  5. [5]

    An efficient fuzzy system for complex query answering on knowledge graphs,

    Y. Lu, H. Chen, Y. Rao, J. Yu, W. Hua, and Q. Li, “An efficient fuzzy system for complex query answering on knowledge graphs,”IEEE Transactions on Knowledge and Data Engineering , vol. 37, no. 9, pp. 4962–4976, 2025

  6. [6]

    Constructing knowledge graphs and their biomedical applications,

    D. N. Nicholson and C. S. Greene, “Constructing knowledge graphs and their biomedical applications,” Computational and Structural Biotechnology Journal, vol. 18, pp. 1414–1428, 2020

  7. [7]

    Knowledge graphs,

    A. Hogan, E. Blomqvist, M. Cochez, C. d’Amato, G. de Melo, C. Gutierrez, S. Kirrane, J. E. Labra Gayo, R. Navigli, S. Neumaier et al. , “Knowledge graphs,” ACM Computing Surveys (CSUR) , vol. 54, no. 4, pp. 1–37, 2021

  8. [8]

    Hierarchy-aware multi-hop question answering over knowledge graphs,

    J. Dong, Q. Zhang, X. Huang, K. Duan, Q. Tan, and Z. Jiang, “Hierarchy-aware multi-hop question answering over knowledge graphs,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 2519–2527

Show all 91 references
  1. [9]

    A survey of knowledge graph reasoning on graph types: Static, dynamic, and multi-modal,

    K. Liang, L. Meng, M. Liu, Y. Liu, W. Tu, S. Wang, S. Zhou, X. Liu, F. Sun, and K. He, “A survey of knowledge graph reasoning on graph types: Static, dynamic, and multi-modal,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 12, pp. 9456–9478, 2024

  2. [10]

    Unifying large language models and knowledge graphs: A roadmap,

    S. Pan, L. Luo, Y. Wang, C. Chen, J. Wang, and X. Wu, “Unifying large language models and knowledge graphs: A roadmap,” IEEE Transactions on Knowledge and Data Engineering , vol. 36, no. 7, pp. 3580–3599, 2024

  3. [11]

    A survey on knowledge graphs: Representation, acquisition, and applica- tions,

    S. Ji, S. Pan, E. Cambria, P . Marttinen, and S. Y. Philip, “A survey on knowledge graphs: Representation, acquisition, and applica- tions,” IEEE transactions on neural networks and learning systems , vol. 33, no. 2, pp. 494–514, 2021

  4. [12]

    Bilinear scoring function search for knowledge graph learning,

    Y. Zhang, Q. Yao, and J. T. Kwok, “Bilinear scoring function search for knowledge graph learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 2, pp. 1458–1473, 2023

  5. [13]

    Multi-modal knowledge graph construction and application: A survey,

    X. Zhu, Z. Li, X. Wang, X. Jiang, P . Sun, X. Wang, Y. Xiao, and N. J. Yuan, “Multi-modal knowledge graph construction and application: A survey,” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 2, pp. 715–735, 2024

  6. [14]

    Rotate: Knowledge graph embedding by relational rotation in complex space,

    Z. Sun, Z.-H. Deng, J.-Y. Nie, and J. Tang, “Rotate: Knowledge graph embedding by relational rotation in complex space,” arXiv preprint arXiv:1902.10197, 2019

  7. [15]

    Context-aware graph inference with knowledge distillation for visual dialog,

    D. Guo, H. Wang, and M. Wang, “Context-aware graph inference with knowledge distillation for visual dialog,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 10, pp. 6056–6073, 2022

  8. [16]

    Transfer-and-fusion: Integrated link prediction across knowledge graphs,

    Y. Cui, Z. Sun, and W. Hu, “Transfer-and-fusion: Integrated link prediction across knowledge graphs,” IEEE Transactions on Knowl- edge and Data Engineering, vol. 37, no. 5, pp. 3062–3074, 2025

  9. [17]

    Inductive relation prediction by subgraph reasoning,

    K. Teru, E. Denis, and W. Hamilton, “Inductive relation prediction by subgraph reasoning,” in International conference on machine learning. PMLR, 2020, pp. 9448–9457

  10. [18]

    Duality- induced regularizer for semantic matching knowledge graph embeddings,

    J. Wang, Z. Zhang, Z. Shi, J. Cai, S. Ji, and F. Wu, “Duality- induced regularizer for semantic matching knowledge graph embeddings,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 2, pp. 1652–1667, 2023

  11. [19]

    Relational message passing for knowledge graph completion,

    H. Wang, H. Ren, and J. Leskovec, “Relational message passing for knowledge graph completion,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , 2021, pp. 1697–1707

  12. [20]

    Translating embeddings for modeling multi- relational data,

    A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings for modeling multi- relational data,” Advances in neural information processing systems , vol. 26, 2013

  13. [21]

    Knowledge graph embedding by translating on hyperplanes,

    Z. Wang, J. Zhang, J. Feng, and Z. Chen, “Knowledge graph embedding by translating on hyperplanes,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 28, no. 1, 2014

  14. [22]

    Embedding entities and relations for learning and inference in knowledge bases,

    B. Yang, W.-t. Yih, X. He, J. Gao, and L. Deng, “Embedding entities and relations for learning and inference in knowledge bases,” in International Conference on Learning Representations (ICLR), 2015

  15. [23]

    Path- based reasoning approach for knowledge graph completion using cnn-bilstm with attention mechanism,

    B. Jagvaral, W.-K. Lee, J.-S. Roh, M.-S. Kim, and Y.-T. Park, “Path- based reasoning approach for knowledge graph completion using cnn-bilstm with attention mechanism,” Expert Systems with Appli- cations, vol. 142, p. 112960, 2020

  16. [24]

    Convd: Attention enhanced dynamic convolutional embeddings for knowledge graph completion,

    W. Guo, Z. Li, X. Wang, Z. Chen, J. Zhao, J. Li, and Y. Yuan, “Convd: Attention enhanced dynamic convolutional embeddings for knowledge graph completion,” IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 9, pp. 5049–5062, 2025

  17. [25]

    Replacing paths with connection-biased attention for knowledge graph completion,

    S. Dutta, A. Gittens, M. J. Zaki, and C. C. Aggarwal, “Replacing paths with connection-biased attention for knowledge graph completion,” ser. AAAI’25/IAAI’25/EAAI’25. AAAI Press, 2025. [Online]. Available: https://doi.org/10.1609/aaai.v39i14.33632

  18. [26]

    Knowledge graph reasoning with relational digraph,

    Y. Zhang and Q. Yao, “Knowledge graph reasoning with relational digraph,” in Proceedings of the ACM web conference 2022 , 2022, pp. 912–924

  19. [27]

    Making large language models perform better in knowledge graph completion,

    Y. Zhang, Z. Chen, L. Guo, Y. Xu, W. Zhang, and H. Chen, “Making large language models perform better in knowledge graph completion,” in Proceedings of the 32nd ACM international conference on multimedia, 2024, pp. 233–242

  20. [28]

    Llms for knowledge graph construction and reasoning: recent capabilities and future opportunities,

    Y. Zhu, X. Wang, J. Chen, S. Qiao, Y. Ou, Y. Yao, S. Deng, H. Chen, and N. Zhang, “Llms for knowledge graph construction and reasoning: recent capabilities and future opportunities,” World Wide Web , vol. 27, no. 5, Aug. 2024. [Online]. Available: https://doi.org/10.1007/s1128...

  21. [29]

    Kg-bert: Bert for knowledge graph completion,

    L. Yao, C. Mao, and Y. Luo, “Kg-bert: Bert for knowledge graph completion,” arXiv preprint arXiv:1909.03193, 2019. 15

  22. [30]

    Llm-based multi-level knowledge generation for few-shot knowledge graph completion,

    Q. Li, Z. Chen, C. Ji, S. Jiang, and J. Li, “Llm-based multi-level knowledge generation for few-shot knowledge graph completion,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence , ser. IJCAI ’24, 2024. [Online]. Available: https://d...

  23. [31]

    Integrating knowledge graph data with large language models for explainable inference,

    C. E. Quintero-Narvaez and R. Monroy, “Integrating knowledge graph data with large language models for explainable inference,” in Proceedings of the 17th ACM International Conference on Web Search and Data Mining, 2024, pp. 1198–1199

  24. [32]

    Sma-gnn: A symbol-aware graph neural network for signed link prediction in recommender systems,

    Y. Zhao, H. Lin, S. Wen, J. Shen, and B. Hua, “Sma-gnn: A symbol-aware graph neural network for signed link prediction in recommender systems,” in Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .2, ser. KDD ’25. New York, NY, USA: Associ...

  25. [33]

    A survey on augmenting knowledge graphs (kgs) with large language models (llms): models, evaluation metrics, benchmarks, and challenges,

    N. Ibrahim, S. Aboulela, A. Ibrahim, and R. Kashef, “A survey on augmenting knowledge graphs (kgs) with large language models (llms): models, evaluation metrics, benchmarks, and challenges,” Discover Artificial Intelligence, vol. 4, no. 1, p. 76, 2024

  26. [34]

    Drgi: Deep relational graph infomax for knowledge graph completion,

    S. Liang, J. Shao, D. Zhang, J. Zhang, and B. Cui, “Drgi: Deep relational graph infomax for knowledge graph completion,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 3, pp. 2486–2499, 2023

  27. [35]

    An adaptive framework embedded with llm for knowledge graph construc- tion,

    Q. Wang, C. Li, Y. Liu, Q. Zhu, J. Song, and T. Shen, “An adaptive framework embedded with llm for knowledge graph construc- tion,” IEEE Transactions on Multimedia, vol. 27, pp. 2912–2923, 2025

  28. [36]

    Kgdm: A diffusion model to capture multiple relation semantics for knowl- edge graph embedding,

    X. Long, L. Zhuang, A. Li, J. Wei, H. Li, and S. Wang, “Kgdm: A diffusion model to capture multiple relation semantics for knowl- edge graph embedding,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8850–8858

  29. [37]

    Fact embedding through diffusion model for knowledge graph completion,

    X. Long, L. Zhuang, A. Li, H. Li, and S. Wang, “Fact embedding through diffusion model for knowledge graph completion,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2020–2029

  30. [38]

    Knowledge graph embedding: An overview,

    X. Ge, Y. C. Wang, B. Wang, C.-C. J. Kuo et al. , “Knowledge graph embedding: An overview,” APSIP A Transactions on Signal and Information Processing, vol. 13, no. 1, 2024

  31. [39]

    Knowledge graph em- bedding: A survey from the perspective of representation spaces,

    J. Cao, J. Fang, Z. Meng, and S. Liang, “Knowledge graph em- bedding: A survey from the perspective of representation spaces,” ACM Computing Surveys, vol. 56, no. 6, pp. 1–42, 2024

  32. [40]

    Complex embeddings for simple link prediction,

    T. Trouillon, J. Welbl, S. Riedel, ´E. Gaussier, and G. Cevaert, “Complex embeddings for simple link prediction,” in Proceedings of the 33rd International Conference on Machine Learning (ICML) , 2016, pp. 2071–2080

  33. [41]

    Rot-pro: Modeling transitivity by projection in knowledge graph embedding,

    T. Song, J. Luo, and L. Huang, “Rot-pro: Modeling transitivity by projection in knowledge graph embedding,” in Proceedings of the Thirty-Fifth Annual Conference on Advances in Neural Information Processing Systems (NeurIPS), 2021

  34. [42]

    Learning hierarchy-aware knowledge graph embeddings for link prediction,

    Z. Zhang, J. Cai, Y. Zhang, and J. Wang, “Learning hierarchy-aware knowledge graph embeddings for link prediction,” in Thirty- Fourth AAAI Conference on Artificial Intelligence. AAAI Press, 2020, pp. 3065–3072

  35. [43]

    Boxe: a box embedding model for knowledge base completion,

    R. Abboud, u. u. Ceylan, T. Lukasiewicz, and T. Salvatori, “Boxe: a box embedding model for knowledge base completion,” in Proceedings of the 34th International Conference on Neural Information Processing Systems, ser. NIPS ’20. Red Hook, NY, USA: Curran Associates Inc., 2020

  36. [44]

    Analogical inference enhanced knowledge graph embedding,

    Z. Yao, W. Zhang, M. Chen, Y. Huang, Y. Yang, and H. Chen, “Analogical inference enhanced knowledge graph embedding,” in Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligen...

  37. [45]

    Paths- over-graph: Knowledge graph empowered large language model reasoning,

    X. Tan, X. Wang, Q. Liu, X. Xu, X. Yuan, and W. Zhang, “Paths- over-graph: Knowledge graph empowered large language model reasoning,” in Proceedings of the ACM on Web Conference 2025, 2025, pp. 3505–3522

  38. [46]

    SimKGC: Simple contrastive knowledge graph completion with pre-trained language models,

    L. Wang, W. Zhao, Z. Wei, and J. Liu, “SimKGC: Simple contrastive knowledge graph completion with pre-trained language models,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , S. Muresan, P . Nakov, and A. Vi...

  39. [47]

    Kgdm: A diffusion model to capture multiple relation semantics for knowledge graph embedding,

    X. Long, L. Zhuang, A. Li, J. Wei, H. Li, and S. Wang, “Kgdm: A diffusion model to capture multiple relation semantics for knowledge graph embedding,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 8, pp. 8850–8858, Mar

  40. [48]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

  41. [49]

    Path-based explanation for knowledge graph completion,

    H. Chang, J. Ye, A. Lopez-Avila, J. Du, and J. Li, “Path-based explanation for knowledge graph completion,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 231–242

  42. [50]

    Semsi-gat: Semantic similarity-based interaction graph attention network for knowl- edge graph completion,

    X. Wang, K. Zhang, M. Niu, and X. Wang, “Semsi-gat: Semantic similarity-based interaction graph attention network for knowl- edge graph completion,” IEEE Transactions on Knowledge and Data Engineering, vol. 37, no. 5, pp. 2958–2970, 2025

  43. [51]

    Replac- ing paths with connection-biased attention for knowledge graph completion,

    S. Dutta, A. Gittens, M. J. Zaki, and C. C. Aggarwal, “Replac- ing paths with connection-biased attention for knowledge graph completion,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 14, 2025, pp. 14 885–14 892

  44. [52]

    Score-based generative modeling through stochastic differential equations,

    Y. Song, J. Sohl-Dickstein, D. P . Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020

  45. [53]

    Srdiff: Single image super-resolution with diffusion probabilistic models,

    H. Li, Y. Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y. Chen, “Srdiff: Single image super-resolution with diffusion probabilistic models,” Neurocomputing, vol. 479, pp. 47–59, 2022

  46. [54]

    Diffusion-lm improves controllable text generation,

    X. Li, J. Thickstun, I. Gulrajani, P . S. Liang, and T. B. Hashimoto, “Diffusion-lm improves controllable text generation,” Advances in Neural Information Processing Systems, vol. 35, pp. 4328–4343, 2022

  47. [55]

    Csdi: Conditional score-based diffusion models for probabilistic time series impu- tation,

    Y. Tashiro, J. Song, Y. Song, and S. Ermon, “Csdi: Conditional score-based diffusion models for probabilistic time series impu- tation,” Advances in Neural Information Processing Systems , vol. 34, pp. 24 804–24 816, 2021

  48. [56]

    Neural ordinary differential equations,

    R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural ordinary differential equations,” Advances in neural infor- mation processing systems, vol. 31, 2018

  49. [57]

    Flow matching for generative modeling,

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” in International Conference on Learning Representations (ICLR), 2023

  50. [58]

    Building normalizing flows with stochastic interpolants,

    M. S. Albergo and E. Vanden-Eijnden, “Building normalizing flows with stochastic interpolants,” in International Conference on Learning Representations (ICLR), 2023

  51. [59]

    Rectified flow: A marginal preserving approach to opti- mal transport,

    Q. Liu, “Rectified flow: A marginal preserving approach to opti- mal transport,” arXiv preprint arXiv:2209.14577, 2022

  52. [60]

    Mean flows for one-step generative modeling,

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,” arXiv preprint arXiv:2505.13447, 2025

  53. [61]

    Flow matching on general geome- tries,

    R. T. Chen and Y. Lipman, “Flow matching on general geome- tries,” in International Conference on Learning Representations (ICLR), 2024

  54. [62]

    Multisample flow matching: Straight- ening flows with minibatch couplings,

    A.-A. Pooladian, H. Ben-Hamu, C. Domingo-Enrich, B. Amos, Y. Lipman, and R. T. Chen, “Multisample flow matching: Straight- ening flows with minibatch couplings,” in International Conference on Learning Representations (ICLR), 2023

  55. [63]

    Improving and generalizing flow-based generative models with minibatch optimal transport,

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y. Zhang, J. Rector- Brooks, G. Wolf, and Y. Bengio, “Improving and generalizing flow-based generative models with minibatch optimal transport,”

  56. [64]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,

    N. Ma, M. Goldstein, M. S. Albergo, N. M. Boffi, E. Vanden- Eijnden, and S. Xie, “Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers,” in European Conference on Computer Vision. Springer, 2024, pp. 23–40

  57. [65]

    Discrete flow matching,

    I. Gat, T. Remez, N. Shaul, F. Kreuk, R. T. Chen, G. Synnaeve, Y. Adi, and Y. Lipman, “Discrete flow matching,” Advances in Neural Information Processing Systems , vol. 37, pp. 133 345–133 385, 2024

  58. [66]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

  59. [67]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Infor- mation Processing Systems, 2017

  60. [68]

    Diffusione: Reasoning on knowledge graphs via diffusion-based graph neural networks,

    Z. Cao, J. Li, Z. Wang, and J. Li, “Diffusione: Reasoning on knowledge graphs via diffusion-based graph neural networks,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 222–230. 16

  61. [69]

    Schrijver, Combinatorial Optimization

    A. Schrijver, Combinatorial Optimization . Berlin; New York: Springer, 2003, see p. 362

  62. [70]

    Observed versus latent features for knowledge base and text inference,

    K. Toutanova and D. Chen, “Observed versus latent features for knowledge base and text inference,” in Proceedings of the 3rd Work- shop on Continuous Vector Space Models and their Compositionality . Beijing, China: Association for Computational Linguistics, Jul. 2015, pp. 57–66

  63. [71]

    Convo- lutional 2d knowledge graph embeddings,

    T. Dettmers, P . Minervini, P . Stenetorp, and S. Riedel, “Convo- lutional 2d knowledge graph embeddings,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 32, no. 1, 2018

  64. [72]

    Deeppath: A reinforcement learning method for knowledge graph reasoning,

    W. Xiong, T. Hoang, and W. Y. Wang, “Deeppath: A reinforcement learning method for knowledge graph reasoning,” arXiv preprint arXiv:1707.06690, 2017

  65. [73]

    Simple embedding for link prediction in knowledge graphs,

    S. M. Kazemi and D. Poole, “Simple embedding for link prediction in knowledge graphs,” Advances in neural information processing systems, vol. 31, 2018

  66. [74]

    Quaternion knowledge graph embeddings,

    S. Zhang, Y. Tay, L. Yao, and Q. Liu, “Quaternion knowledge graph embeddings,” Advances in neural information processing systems , vol. 32, 2019

  67. [75]

    Drum: End-to-end differentiable rule mining on knowledge graphs,

    A. Sadeghian, M. Armandpour, P . Ding, and D. Z. Wang, “Drum: End-to-end differentiable rule mining on knowledge graphs,” Advances in neural information processing systems , vol. 32, 2019

  68. [76]

    Path-based reasoning with k-nearest neighbor and position embedding for knowledge graph comple- tion,

    Z. Peng, H. Yu, and X. Jia, “Path-based reasoning with k-nearest neighbor and position embedding for knowledge graph comple- tion,” Journal of Intelligent Information Systems, pp. 1–21, 2022

  69. [77]

    Modeling relational data with graph convolutional networks,

    M. Schlichtkrull, T. N. Kipf, P . Bloem, R. van den Berg, I. Titov, and M. Welling, “Modeling relational data with graph convolutional networks,” The Semantic Web, vol. 10843, pp. 593–607, 2018

  70. [78]

    Joint language semantic and structure embedding for knowledge graph completion,

    J. Shen, C. Wang, L. Gong, and D. Song, “Joint language semantic and structure embedding for knowledge graph completion,”arXiv preprint arXiv:2209.08721, 2022

  71. [79]

    Scientific language models for biomedical knowledge base completion: An empirical study,

    R. Nadkarni, D. Wadden, I. Beltagy, N. A. Smith, H. Ha- jishirzi, and T. Hope, “Scientific language models for biomedical knowledge base completion: An empirical study,” arXiv preprint arXiv:2106.09700, 2021

  72. [80]

    Knowledge graph refinement based on triplet bert-networks,

    A. K. Nassiri, N. Pernelle, F. Sa ¨ıs, and G. Quercini, “Knowledge graph refinement based on triplet bert-networks,” arXiv preprint arXiv:2211.10460, 2022

  73. [81]

    Fine-grained evaluation of rule-and embedding-based systems for knowledge graph completion,

    C. Meilicke, M. Fink, Y. Wang, D. Ruffinelli, R. Gemulla, and H. Stuckenschmidt, “Fine-grained evaluation of rule-and embedding-based systems for knowledge graph completion,” in The semantic web–ISWC 2018: 17th international semantic web confer- ence, Monterey, CA, USA, Octobe...

  74. [82]

    Differentiable learning of logical rules for knowledge base reasoning,

    F. Yang, Z. Yang, and W. W. Cohen, “Differentiable learning of logical rules for knowledge base reasoning,” Advances in neural information processing systems, vol. 30, 2017

  75. [83]

    Visualizing data using t-sne,

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of machine learning research , vol. 9, no. Nov, pp. 2579–2605, 2008

  76. [84]

    Kinship,

    G. Hinton, “Kinship,” UCI Machine Learning Repository, 1986

  77. [85]

    The unified medical language system (umls): in- tegrating biomedical terminology,

    O. Bodenreider, “The unified medical language system (umls): in- tegrating biomedical terminology,” Nucleic Acids Research, vol. 32, no. suppl 1, pp. D267–D270, 2004

  78. [86]

    Tdn: Triplet distributor network for knowledge graph completion,

    J. Wang, B. Wang, J. Gao, X. Li, Y. Hu, and B. Yin, “Tdn: Triplet distributor network for knowledge graph completion,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 12, pp. 13 002–13 014, 2023

  79. [87]

    Rnnlogic: Learning logic rules for reasoning on knowledge graphs,

    M. Qu, J. Chen, L.-P . Xhonneux, Y. Bengio, and J. Tang, “Rnnlogic: Learning logic rules for reasoning on knowledge graphs,” arXiv preprint arXiv:2010.04029, 2020

  80. [88]

    Neural bellman-ford networks: A general graph neural network framework for link prediction,

    Z. Zhu, Z. Zhang, L.-P . Xhonneux, and J. Tang, “Neural bellman-ford networks: A general graph neural network framework for link prediction,” in Advances in Neural Information Processing Systems , M. Ranzato, A. Beygelzimer, Y. Dauphin, P . Liang, and J. W. Vaughan, Eds., vol....

  81. [89]

    Adaprop: Learning adaptive propagation for graph neural network based knowledge graph reasoning,

    Y. Zhang, Z. Zhou, Q. Yao, X. Chu, and B. Han, “Adaprop: Learning adaptive propagation for graph neural network based knowledge graph reasoning,” in Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining , 2023, pp. 3446–3457. Siyuan Li is current...

  82. [2023]

    Available: https://arxiv.org/abs/2302.00482

    [Online]. Available: https://arxiv.org/abs/2302.00482

  83. [2024]

    Available: https://ojs.aaai.org/index.php/AAAI/ article/view/28732

    [Online]. Available: https://ojs.aaai.org/index.php/AAAI/ article/view/28732

Pith tools

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