Neural-Symbolic Message Passing with Dynamic Pruning

Chongzhi Zhang; Junhao Zheng; Qianli Ma; Zhiping Peng

Neural-Symbolic Message Passing with Dynamic Pruning

T0 review · 1 major / 0 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a frozen neural link predictor plus fuzzy symbolic message passing, with dynamic pruning, answers arbitrary existential first-order queries without complex-query training and runs 2× to over 150× faster than…

desk verdict A genuinely new training-free combination of fuzzy symbolic states and neural one-hop message passing, but the 'arbitrary EFO1' claim breaks on constant-free query components and the empirical reporting needs more discipline. read the letter →

arxiv 2501.14661 v1 pith:CE5JYSPO submitted 2025-01-24 cs.LG cs.AI

Chongzhi Zhang , Junhao Zheng , Zhiping Peng , Qianli Ma This is my paper · ORCID

classification cs.LGcs.AI

keywords ComplexQueryAnsweringKnowledgeGraphsNeural-SymbolicReasoningFuzzyLogicMessagePassingDynamicPruningExistentialFirst-OrderLinkPrediction

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 tries to establish that complex query answering over incomplete knowledge graphs can be done by combining a frozen pretrained neural link predictor with fuzzy symbolic reasoning on the query graph, with no training on complex queries. The central claim is that messages between variable nodes are noisy until a node's state has been updated, and dynamically pruning those un-updated messages improves both accuracy and speed. If true, the proposed Neural-Symbolic Message Passing (NSMP) framework would give an interpretable, training-free route to arbitrary existential first-order queries, with inference times 2× to over 150× faster than the step-by-step neural-symbolic baseline on tested graphs. The paper reports strong performance on FB15k-237 and NELL995, best among message-passing models and competitive with the current state of the art.

What carries the argument

The load-bearing object is the neural-symbolic message encoding function $\varrho$, which for each directed edge combines $f(\rho(\cdot))$, the pretrained link predictor's score converted into a fuzzy vector via softmax similarities, with $\mu(\cdot)$, a symbolic one-hop inference using the relation's 0/1 adjacency matrix and thresholded normalization. Negated edges are handled through a fuzzy-logic negator that replaces the adjacency-matrix product with $\alpha/|V| - pM$. Messages received by a variable are aggregated with the Hadamard product, and the dynamic pruning rule decides which variable nodes may send messages based on whether their state has been updated. The complexity claim, roughly $O(|V|^2)$ for any existential first-order formula, follows because the symbolic sparse operations dominate and the layer count is bounded by the query diameter.

What would settle it

Build a cyclic conjunctive query whose cycle contains no constant node in the early message-passing layers, so an answer entity can only be reached through a message emitted by a variable before that variable's state was updated; run NSMP with and without dynamic pruning on thousands of such queries. If the pruned variant's MRR drops below the unpruned variant's, the assumption that un-updated variables emit only noise fails for those structures.

Watch

Extended reading notes

Core claim

The central discovery is that a message-passing complex query answering model can remain fully neural-symbolic: each variable's state is a fuzzy vector over entities, one-hop messages are computed by a neural-symbolic encoding that adds a neural link predictor's soft score vector to the symbolic result of a sparse adjacency-matrix multiplication, and messages are combined by product fuzzy logic. A dynamic pruning rule then gates messages from variable nodes: only variables whose states have already been updated are allowed to send messages to other variables. This removes what the authors call noise from un-updated variable nodes, avoids wasted computation, and lets the free-variable state converge over a number of layers set by the query diameter. The authors claim this yields arbitrary existential first-order query answering without training on complex queries, improvements over other message-passing baselines especially on negative queries, and a speedup over the step-by-step neural-symbolic baseline ranging from 2× to over 150×, with the largest gains on cyclic queries.

Load-bearing premise

The load-bearing premise is that a variable node whose state has not yet been updated carries no information worth passing, so every message from such a node can be dropped without losing any answer signal; the paper supports this with ablations, not with a theorem.

Editorial extensions

If this is right

A frozen link predictor suffices for arbitrary existential first-order queries, so complex-query training data and trainable parameters become unnecessary for this task.
Cyclic queries, which step-by-step methods answer by enumerating a variable with cost $O(|V|^n)$, can be handled in roughly $O(|V|^2)$ time, with empirically larger speedups on more cyclic queries.
Interpretability improves because each variable's state is a fuzzy vector whose entries are per-entity membership degrees, visible in the paper's case-study entity rankings.
Negative queries improve substantially relative to prior message-passing models because negation is executed directly through fuzzy logic.
The speedups compound on larger knowledge graphs: the relative gain over the step-by-step baseline is larger on NELL995 than on FB15k-237.

Reading between the lines

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

Editorial extension: the blanket rule that an un-updated variable emits only noise could be replaced by a confidence threshold on its fuzzy state, which would let a variable broadcast as soon as its state becomes informative, potentially rescuing answers in cycles with no constant anchor.
Editorial extension: because NSMP has no trainable parameters, its fuzzy-vector states could serve as a cheap, interpretable initialization or regularizer for trained neural query-answering models, a combination the paper does not explore.
Editorial extension: the complexity argument predicts that on query graphs with more variables in a cycle, the relative speedup over step-by-step enumeration should keep growing roughly like $O(|V|^{n-2})$; measuring that trend across cycle sizes would separate the complexity effect from implementation details.

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.

Desk Editor's Note

read the letter

Short version: the core idea is solid and worth engaging with; the 'arbitrary EFO1' claim is not. NSMP combines a frozen neural link predictor with fuzzy symbolic state vectors on query graphs, and prunes variable-to-variable messages until a node has been updated. That combination is new relative to LMPNN/CLMPT and FIT/QTO, and it pays off: no complex-query training, interpretable fuzzy states, better negative-query MRR than prior message-passing models, and large speedups over FIT. The ablation shows pruning helps, and the speed numbers are consistent.

What I'd want fixed before trusting it. The biggest issue is the scope claim. Definition 3.1 allows EFO1 formulas in which a connected component contains no constant. NSMP starts all variables at zero and only propagates from constants; with or without dynamic pruning, a variable in a constant-free component never receives a non-zero update, so Eq. 19 cannot rank anything. The stress-test example q(x)=∃y R(x,y)∧S(y,x) is valid EFO1 and FIT/QTO can handle it by anchoring a variable; NSMP cannot. So the abstract's 'arbitrary existential first order logic queries' and Proposition 4.1 are false as stated. The benchmark query types all have constants, so the reported numbers stand, but the claim needs to be narrowed or the machinery needs a fallback for constant-free components.

The second concern is empirical hygiene. No code, no error bars, and epsilon is tuned per dataset. Table 10 shows average positive MRR going from 10.5 at epsilon=1e-2 to 27.6 at 1e-14. That is a huge sensitivity. I would not trust the exact MRR deltas over baselines until there is variance reporting and a clear separation of validation and test. The efficiency comparison is more robust because the speedups are large and consistent.

Minor: Proposition 4.3's proof assumes K = |T*(vk)>0| without justification, and the claim that a binary adjacency matrix is always sparser than a thresholded neural one is heuristic. The empirical speedups support the conclusion, so this is a proof gap rather than a fatal flaw.

Bottom line: the paper deserves a serious referee. It is a genuine new combination with useful results, but the arbitrary-EFO1 overclaim and the missing code/variance need to be addressed. I'd send it to review with a request for those revisions.

Referee Report

1 major / 0 minor

Summary. The paper proposes NSMP, a training-free complex query answering (CQA) method that combines a frozen pre-trained neural link predictor (ComplEx-N3) with fuzzy symbolic propagation over query graphs. Neural one-hop inference on atomic formulas is converted into fuzzy membership vectors via softmax over entity similarities, and symbolic one-hop inference is performed with sparse adjacency matrices and a thresholded normalization; the two are summed and normalized. Messages between variable nodes are governed by a dynamic pruning rule in which a variable may send messages only after its state has been updated. States are updated by product fuzzy-logic aggregation, and final answers are ranked by a weighted combination of the free variable's symbolic fuzzy vector and a neural cosine-similarity term. The method is evaluated on FB15k-237 and NELL995 on both the BetaE and FIT query suites, and the paper reports strong MRR, particularly on negative queries and on FIT datasets, plus 2x--150x inference speedups over the FIT baseline. The central claims are that NSMP generalizes to arbitrary existential first-order logic (EFO1) queries without complex-query training, provides interpretability through fuzzy variable states, and gains both accuracy and efficiency from dynamic pruning.

Significance. If the arbitrary-EFO1 claim were fully supported, NSMP would be a notable contribution: a training-free CQA method with interpretable symbolic states, strong performance on negative queries, and drastically faster inference than step-by-step neural-symbolic methods, especially on cyclic queries. The core message-passing equations are coherent, the use of a frozen pre-trained link predictor is principled, and the empirical trend is consistent across two KGs and many query types; the speedups reported in Fig. 3 and Table 8 are concrete and align with the complexity analysis in App. B. The paper also provides explicit proofs, a hyperparameter sensitivity study, and a case study on interpretability, which are valuable. However, the 'arbitrary EFO1' claim is not supported for query graphs that contain a connected component with no constant entity, and the reported performance is sensitive to the tuned epsilon threshold without error bars, so the empirical significance is somewhat fragile.

major comments (1)

[App. E, Table 16] The efficiency ablation shows a 23.8% reduction in average inference time from dynamic pruning on FB15k-237. This is a useful result, but the paper does not quantify the overhead of the pruning logic itself; adding a small breakdown would make the efficiency claim more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: NSMP's inference equations compose a frozen external link predictor with symbolic one-hop inference, and the claims are empirically benchmarked against independent baselines.

full rationale

NSMP's derivation chain is self-contained rather than circular. Equations (12)-(15) define neural-symbolic messages as a normalized combination of a neural fuzzy vector (derived from a frozen, externally released ComplEx-N3 checkpoint) and a symbolic one-hop inference over KG adjacency matrices. Equations (16)-(18) aggregate messages with product fuzzy logic and update states, and Equation (19) reads out the answer distribution at the free variable. None of these equations is defined in terms of the reported MRR targets, nor is any target metric used as an input to the equations. The claims of generalization to arbitrary EFO1 queries are algorithmically stated, with the known limitation that queries without any constant anchor do not receive non-zero initial messages; that is a correctness or coverage issue, not circularity. The complexity and speedup comparisons are against independent external methods (FIT, QTO, LMPNN) and use cited complexity results plus measured wall-clock times; same-group citations such as CLMPT are used for design choices and as a baseline, not as an unverified load-bearing uniqueness theorem. Hyperparameters such as epsilon, alpha, lambda, and depth L are selected by grid search on validation/test benchmarks, which is a data-fit component that could raise generalization concerns, but it does not make the central derivation equivalent to its inputs by construction. No equation reduces to a fitted parameter renamed as a prediction, and no uniqueness theorem from the authors is used to forbid alternative designs. Therefore no significant circularity is found.

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

Most of NSMP's behavior follows from standard message passing, fuzzy logic, and pretrained link prediction. The load-bearing additions are the fuzzy-state representation, the dynamic pruning rule, and the sparsity assumption in the complexity proof. I count four explicit or implicit assumptions and four tuned hyperparameters; the empirical claims are independent of the prior literature's equations, so circularity is low.

free parameters (4)

epsilon (threshold in normalization N) = 1e-14 (FB15k-237 and NELL995)
Tuned by grid search; controls sparsity of fuzzy vectors and has large effect on MRR (Table 10), e.g., AVG(P) rises from 10.5 to 27.6 as epsilon decreases.
alpha (fuzzy negation term) = 100 for FB15k-237, 1000 for NELL995
Tuned by grid search; used in negated-edge symbolic inference (Eq. 9-10) and affects negative-query scores (Table 11).
lambda (neural-symbolic balance) = 0.3 for FB15k-237, 0.1 for NELL995
Tuned by grid search; balances final fuzzy state and neural embedding similarity (Eq. 19); lambda=0 gives much worse results (Table 12).
L (number of NSMP layers / depth) = D+1 by default, with manual choice per query type reaching 23.3 avg MRR
Layer count is a user-set hyperparameter; Table 3 shows different query types prefer different depths, and manual selection improves average MRR.

assumptions (4)

domain assumption Product fuzzy logic with Hadamard product correctly aggregates conjunctions of messages (Eq. 16).
The paper assumes fuzzy set intersection is product t-norm; this is standard fuzzy logic but is a modeling choice, not a derived property of neural link predictor scores.
domain assumption Pretrained neural link predictor scores, after softmax/entity similarity conversion (Eq. 11), are valid fuzzy membership degrees that can be summed with symbolic propagation values (Eq. 12-15).
No calibration or correctness proof is provided that the converted scores are suitable for fuzzy aggregation; the method's success rests on this assumption.
ad hoc to paper Messages from variable nodes whose states have not been updated are pure noise and can be pruned without losing necessary signal (Sec 4.2.1).
This is the core dynamic pruning rule; the paper provides intuition and ablation support but no theorem guaranteeing no information loss for all query graphs.
ad hoc to paper For the sparsity comparison, the symbolic adjacency matrix of NSMP is strictly sparser than the neural adjacency matrix of FIT/QTO, L_N < L_Q (App B.3).
The proof assumes this inequality without a rigorous derivation; it anchors Proposition 4.3 and the claimed theoretical superiority on acyclic queries.

invented entities (1)

Fuzzy symbolic variable state (s_v) independent evidence
purpose: Encodes each query variable as a probability-like distribution over entities, enabling symbolic one-hop inference and interpretable top-k explanations.
The state is directly observable in case-study rankings (App H) and its utility is tested by ablations; it is a representational construct, not a hidden explanatory entity, so independent evidence exists.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural-Symbolic Message Passing with Dynamic Pruning." pith.science (2026). https://pith.science/paper/CE5JYSPO

@misc{pith2026250114661,
  author       = {Pith},
  title        = {Pith review of: Neural-Symbolic Message Passing with Dynamic Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CE5JYSPO}},
  note         = {Machine review of arXiv:2501.14661}
}

abstract

Complex Query Answering (CQA) over incomplete Knowledge Graphs (KGs) is a challenging task. Recently, a line of message-passing-based research has been proposed to solve CQA. However, they perform unsatisfactorily on negative queries and fail to address the noisy messages between variable nodes in the query graph. Moreover, they offer little interpretability and require complex query data and resource-intensive training. In this paper, we propose a Neural-Symbolic Message Passing (NSMP) framework based on pre-trained neural link predictors. By introducing symbolic reasoning and fuzzy logic, NSMP can generalize to arbitrary existential first order logic queries without requiring training while providing interpretable answers. Furthermore, we introduce a dynamic pruning strategy to filter out noisy messages between variable nodes. Experimental results show that NSMP achieves a strong performance. Additionally, through complexity analysis and empirical verification, we demonstrate the superiority of NSMP in inference time over the current state-of-the-art neural-symbolic method. Compared to this approach, NSMP demonstrates faster inference times across all query types on benchmark datasets, with speedup ranging from 2$\times$ to over 150$\times$.

Figures

Figures reproduced from arXiv: 2501.14661 by the authors.

**Figure 2.** A toy example to show the process of dynamic pruning. The blue arrow represents the passing of the message [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Relative speedup of NSMP over FIT in terms of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Graphical representation of the query types of the BetaE dataset considered in our experiment, where [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

**Figure 5.** Figure 5: Graphical representation of the query types of the FIT dataset considered in our experiment, where [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 39 canonical work pages

[1]

S., and Zhang, C

Amayuelas, A., Zhang, S., Rao, X. S., and Zhang, C. Neural methods for logical reasoning over knowledge graphs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=tgcAoUVHRIB

work page 2022
[2]

Complex query answering with neural link predictors

Arakelyan, E., Daza, D., Minervini, P., and Cochez, M. Complex query answering with neural link predictors. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=Mos9F9kDwkz

work page 2021
[3]

Adapting neural link predictors for data-efficient complex query answering

Arakelyan, E., Minervini, P., Daza, D., Cochez, M., and Augenstein, I. Adapting neural link predictors for data-efficient complex query answering. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

work page 2023
[4]

Query2particles: Knowledge graph reasoning with particle embeddings

Bai, J., Wang, Z., Zhang, H., and Song, Y. Query2particles: Knowledge graph reasoning with particle embeddings. In Findings of the Association for Computational Linguistics: NAACL 2022, pp.\ 2703--2714, 2022

work page 2022
[5]

Sequential query encoding for complex query answering on knowledge graphs

Bai, J., Zheng, T., and Song, Y. Sequential query encoding for complex query answering on knowledge graphs. arXiv preprint arXiv:2302.13114, 2023 a

arXiv 2023
[6]

Answering complex logical queries on knowledge graphs via query computation tree optimization

Bai, Y., Lv, X., Li, J., and Hou, L. Answering complex logical queries on knowledge graphs via query computation tree optimization. In International Conference on Machine Learning, pp.\ 1472--1491. PMLR, 2023 b

work page 2023
[7]

Translating embeddings for modeling multi-relational data

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

2013
[8]

Fuzzy logic based logical query answering on knowledge graphs

Chen, X., Hu, Z., and Sun, Y. Fuzzy logic based logical query answering on knowledge graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 3939--3948, 2022

work page 2022

Show all 52 references

[9]

Probabilistic entity representation model for reasoning over knowledge graphs

Choudhary, N., Rao, N., Katariya, S., Subbian, K., and Reddy, C. Probabilistic entity representation model for reasoning over knowledge graphs. Advances in Neural Information Processing Systems, 34: 0 23440--23451, 2021

2021
[10]

Cohen, W., Yang, F., and Mazaitis, K. R. Tensorlog: A probabilistic database implemented using deep-learning infrastructure. Journal of Artificial Intelligence Research, 67: 0 285--325, 2020

2020
[11]

and Cochez, M

Daza, D. and Cochez, M. Message passing query embedding. arXiv preprint arXiv:2002.02406, 2020

2002 arXiv
[12]

Convolutional 2d knowledge graph embeddings

Dettmers, T., Minervini, P., Stenetorp, P., and Riedel, S. Convolutional 2d knowledge graph embeddings. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

2018
[13]

S., Riley, P

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In International conference on machine learning, pp.\ 1263--1272. PMLR, 2017

2017
[14]

Metamathematics of fuzzy logic, volume 4

H \'a jek, P. Metamathematics of fuzzy logic, volume 4. Springer Science & Business Media, 2013

2013
[15]

Embedding logical queries on knowledge graphs

Hamilton, W., Bajaj, P., Zitnik, M., Jurafsky, D., and Leskovec, J. Embedding logical queries on knowledge graphs. Advances in neural information processing systems, 31, 2018

2018
[16]

Ji, S., Pan, S., Cambria, E., Marttinen, P., and Philip, S. Y. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE transactions on neural networks and learning systems, 33 0 (2): 0 494--514, 2021

2021
[17]

Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[18]

P., Mesiar, R., and Pap, E

Klement, E. P., Mesiar, R., and Pap, E. Triangular norms, volume 8. Springer Science & Business Media, 2013

2013
[19]

Canonical tensor decomposition for knowledge base completion

Lacroix, T., Usunier, N., and Obozinski, G. Canonical tensor decomposition for knowledge base completion. In International Conference on Machine Learning, pp.\ 2863--2872. PMLR, 2018

2018
[20]

House: Knowledge graph embedding with householder parameterization

Li, R., Zhao, J., Li, C., He, D., Wang, Y., Liu, Y., Sun, H., Wang, S., Deng, W., Shen, Y., et al. House: Knowledge graph embedding with householder parameterization. In International conference on machine learning, pp.\ 13209--13224. PMLR, 2022

2022
[21]

and Sirangelo, C

Libkin, L. and Sirangelo, C. Open and closed world assumptions in data exchange. Description Logics, 477, 2009

2009
[22]

Mask and reason: Pre-training knowledge graph transformers for complex logical queries

Liu, X., Zhao, S., Su, K., Cen, Y., Qiu, J., Zhang, M., Wu, W., Dong, Y., and Tang, J. Mask and reason: Pre-training knowledge graph transformers for complex logical queries. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 1120--1130, 2022

2022
[23]

Logic embeddings for complex query answering

Luus, F., Sen, P., Kapanipathi, P., Riegel, R., Makondo, N., Lebese, T., and Gray, A. Logic embeddings for complex query answering. arXiv preprint arXiv:2103.00418, 2021

2021 arXiv
[24]

Nguyen, C. D. M., French, T., Liu, W., and Stewart, M. Cyle: Cylinder embeddings for multi-hop reasoning over knowledge graphs. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pp.\ 1736--1751, 2023

2023
[25]

and Leskovec, J

Ren, H. and Leskovec, J. Beta embeddings for multi-hop logical reasoning in knowledge graphs. Advances in Neural Information Processing Systems, 33: 0 19716--19726, 2020

2020
[26]

Query2box: Reasoning over knowledge graphs in vector space using box embeddings

Ren, H., Hu, W., and Leskovec, J. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. arXiv preprint arXiv:2002.05969, 2020

2002 arXiv
[27]

Smore: Knowledge graph completion and multi-hop reasoning in massive knowledge graphs

Ren, H., Dai, H., Dai, B., Chen, X., Zhou, D., Leskovec, J., and Schuurmans, D. Smore: Knowledge graph completion and multi-hop reasoning in massive knowledge graphs. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 1472--1482, 2022

2022
[28]

Knowledge graph embedding for link prediction: A comparative analysis

Rossi, A., Barbosa, D., Firmani, D., Matinata, A., and Merialdo, P. Knowledge graph embedding for link prediction: A comparative analysis. ACM Transactions on Knowledge Discovery from Data (TKDD), 15 0 (2): 0 1--49, 2021

2021
[29]

O., Bedrax - Weiss, T., Pereira, F., and Cohen, W

Sun, H., Arnold, A. O., Bedrax - Weiss, T., Pereira, F., and Cohen, W. W. Faithful embeddings for knowledge base queries. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neu...

2020
[30]

Rotate: Knowledge graph embedding by relational rotation in complex space

Sun, Z., Deng, Z.-H., Nie, J.-Y., and Tang, J. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=HkgEQnRqYQ

2019
[31]

and Chen, D

Toutanova, K. and Chen, D. Observed versus latent features for knowledge base and text inference. In Proceedings of the 3rd workshop on continuous vector space models and their compositionality, pp.\ 57--66, 2015

2015
[32]

Complex embeddings for simple link prediction

Trouillon, T., Welbl, J., Riedel, S., Gaussier, \'E ., and Bouchard, G. Complex embeddings for simple link prediction. In International conference on machine learning, pp.\ 2071--2080. PMLR, 2016

2016
[33]

N., Kaiser, ., and Polosukhin, I

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems, 30, 2017

2017
[34]

Graph attention networks

Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

2017 arXiv
[35]

Benchmarking the combinatorial generalizability of complex query answering on knowledge graphs

Wang, Z., Yin, H., and Song, Y. Benchmarking the combinatorial generalizability of complex query answering on knowledge graphs. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1 (NeurIPS Datasets and Benchmarks 2021), 2022 a

2021
[36]

Logical queries on knowledge graphs: Emerging interface of incomplete relational data

Wang, Z., Yin, H., and Song, Y. Logical queries on knowledge graphs: Emerging interface of incomplete relational data. IEEE Data Eng. Bull., 45 0 (4): 0 3--18, 2022 b

2022
[37]

Y., and See, S

Wang, Z., Fei, W., Yin, H., Song, Y., Wong, G. Y., and See, S. Wasserstein-fisher-rao embedding: Logical query embeddings with local comparison and global transport. arXiv preprint arXiv:2305.04034, 2023 a

2023 arXiv
[38]

Logical message passing networks with one-hop inference on atomic formulas

Wang, Z., Song, Y., Wong, G., and See, S. Logical message passing networks with one-hop inference on atomic formulas. In The Eleventh International Conference on Learning Representations, 2023 b . URL https://openreview.net/forum?id=SoyOsp7i_l

2023
[39]

Xiong, W., Hoang, T., and Wang, W. Y. Deeppath: A reinforcement learning method for knowledge graph reasoning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp.\ 564--573, 2017

2017
[40]

How powerful are graph neural networks? In International Conference on Learning Representations, 2019

Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=ryGs6iA5Km

2019
[41]

Neural-symbolic entangled framework for complex query answering

Xu, Z., Zhang, W., Ye, P., Chen, H., and Chen, H. Neural-symbolic entangled framework for complex query answering. Advances in Neural Information Processing Systems, 35: 0 1806--1819, 2022

2022
[42]

W.-t., He, X., Gao, J., and Deng, L

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

2015
[43]

Gammae: Gamma embeddings for logical queries on knowledge graphs

Yang, D., Qing, P., Li, Y., Lu, H., and Lin, X. Gammae: Gamma embeddings for logical queries on knowledge graphs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp.\ 745--760, 2022

2022
[44]

Rethinking complex queries on knowledge graphs with neural link predictors

Yin, H., Wang, Z., and Song, Y. Rethinking complex queries on knowledge graphs with neural link predictors. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=1BmveEMNbG

2024
[45]

Do transformers really perform badly for graph representation? Advances in Neural Information Processing Systems, 34: 0 28877--28888, 2021

Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., and Liu, T.-Y. Do transformers really perform badly for graph representation? Advances in Neural Information Processing Systems, 34: 0 28877--28888, 2021

2021
[46]

R., and Smola, A

Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdinov, R. R., and Smola, A. J. Deep sets. Advances in neural information processing systems, 30, 2017

2017
[47]

Conditional logical message passing transformer for complex query answering

Zhang, C., Peng, Z., Zheng, J., and Ma, Q. Conditional logical message passing transformer for complex query answering. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 4119--4130, 2024 a

2024
[48]

Pathformer: Recursive path query encoding for complex logical query answering

Zhang, C., Peng, Z., Zheng, J., Wang, L., Shi, R., and Ma, Q. Pathformer: Recursive path query encoding for complex logical query answering. arXiv preprint arXiv:2406.14880, 2024 b

2024 arXiv
[49]

Cone: Cone embeddings for multi-hop reasoning over knowledge graphs

Zhang, Z., Wang, J., Chen, J., Ji, S., and Wu, F. Cone: Cone embeddings for multi-hop reasoning over knowledge graphs. Advances in Neural Information Processing Systems, 34: 0 19172--19183, 2021

2021
[50]

Neural-symbolic models for logical queries on knowledge graphs

Zhu, Z., Galkin, M., Zhang, Z., and Tang, J. Neural-symbolic models for logical queries on knowledge graphs. In International Conference on Machine Learning, pp.\ 27454--27478. PMLR, 2022

2022
[51]

T., and Zhao, D

Zou, L., Mo, J., Chen, L., \"O zsu, M. T., and Zhao, D. gstore: answering sparql queries via subgraph matching. Proceedings of the VLDB Endowment, 4 0 (8): 0 482--493, 2011

2011
[52]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

S., and Zhang, C

Amayuelas, A., Zhang, S., Rao, X. S., and Zhang, C. Neural methods for logical reasoning over knowledge graphs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=tgcAoUVHRIB

work page 2022

[2] [2]

Complex query answering with neural link predictors

Arakelyan, E., Daza, D., Minervini, P., and Cochez, M. Complex query answering with neural link predictors. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=Mos9F9kDwkz

work page 2021

[3] [3]

Adapting neural link predictors for data-efficient complex query answering

Arakelyan, E., Minervini, P., Daza, D., Cochez, M., and Augenstein, I. Adapting neural link predictors for data-efficient complex query answering. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

work page 2023

[4] [4]

Query2particles: Knowledge graph reasoning with particle embeddings

Bai, J., Wang, Z., Zhang, H., and Song, Y. Query2particles: Knowledge graph reasoning with particle embeddings. In Findings of the Association for Computational Linguistics: NAACL 2022, pp.\ 2703--2714, 2022

work page 2022

[5] [5]

Sequential query encoding for complex query answering on knowledge graphs

Bai, J., Zheng, T., and Song, Y. Sequential query encoding for complex query answering on knowledge graphs. arXiv preprint arXiv:2302.13114, 2023 a

arXiv 2023

[6] [6]

Answering complex logical queries on knowledge graphs via query computation tree optimization

Bai, Y., Lv, X., Li, J., and Hou, L. Answering complex logical queries on knowledge graphs via query computation tree optimization. In International Conference on Machine Learning, pp.\ 1472--1491. PMLR, 2023 b

work page 2023

[7] [7]

Translating embeddings for modeling multi-relational data

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

2013

[8] [8]

Fuzzy logic based logical query answering on knowledge graphs

Chen, X., Hu, Z., and Sun, Y. Fuzzy logic based logical query answering on knowledge graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 3939--3948, 2022

work page 2022