Pith. sign in

REVIEW 4 major objections 5 minor 30 references

From Eigenmodes to Proofs: Integrating Graph Spectral Operators with Symbolic Interpretable Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Spectral NSR claims logical rules can be encoded as frequency-selective graph filters and that reasoning in the Laplacian eigenbasis beats transformer, message-passing, and neuro-symbolic baselines on ProofWriter and CLUTRR.

desk verdict A genuinely novel spectral-template proposal with clean GSP math, but the headline empirical claims are one unreproducible table away from support. read the letter →

arxiv 2509.07017 v1 pith:KIMT32AQ submitted 2025-09-07 cs.AI cs.CLcs.LG

Andrew Kiruluta , Priscilla Burity This is my paper · ORCID
classification cs.AIcs.CLcs.LG
keywords graphsignalprocessingneuro-symbolicreasoningspectralfiltersLaplacianlogicalruletemplatesChebyshevpolynomialapproximationinterpretableadversarialrobustness
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 argues that logical reasoning can be carried out entirely in the frequency domain of a knowledge graph. The authors propose Spectral NSR, a neuro-symbolic architecture that encodes each logical rule as a diagonal filter in the Laplacian eigenbasis, so that inference becomes a combination of frequency-selective graph filters followed by thresholding into predicates. On ProofWriter and CLUTRR they report accuracy of 88.1% and 77.4%, with lower latency than transformer, message-passing, and differentiable-logic baselines, and an 87% agreement between spectral activations and ground-truth proof steps. If the claims hold, the trade-off between neural flexibility and symbolic transparency is resolved by choosing the right representational domain rather than by coupling two separate systems.

What carries the argument

The central object is the graph Laplacian $L = D - A$ with eigendecomposition $L = U\Lambda U^\top$, whose eigenvectors form the graph Fourier basis. The load-bearing identity is the spectral filtering formula $y = h_\theta(L)x^{(0)} = U h_\theta(\Lambda)U^\top x^{(0)}$, which turns convolution on the graph into pointwise multiplication in frequency, exactly as in classical Fourier analysis. Reasoning is carried by spectral templates $\phi_r(\lambda)$, one per logical rule, aggregated into a rule-mixture response $\phi^*(\lambda) = \sum_r w_r \phi_r(\lambda)$; Chebyshev polynomial recurrences $b_{k+1} = 2\tilde{L}b_k - b_{k-1}$ let the filter be applied without computing the eigendecomposition, at $O(K|E|)$ cost. These pieces together carry the claim that logical content is encoded in which frequency bands are amplified or suppressed.

What would settle it

Construct a ProofWriter-style graph where the evidence for a universally true conclusion sits on a checkerboard-like set of nodes, so the supporting signal is concentrated in high-frequency modes, and run the model with only the low-pass template $\phi_r(\lambda)=1/(1+\tau\lambda)$; the claimed frequency semantics predicts failure, while a purely correlational spectral model may still succeed. Alternatively, with the learned eigenvectors fixed, randomly reassign which eigenvalue multiplies each eigenvector in the filter response and measure accuracy without retraining: the paper's frequency-to-logic claim predicts a sharp drop.

Watch

Extended reading notes

Core claim

Spectral NSR's central claim is that symbolic rules and graph structure share a single representational language: the eigenmodes of the graph Laplacian. A rule is not bolted onto a neural network; it is a spectral template $\phi_r(\lambda)$ applied to the belief vector in the Fourier basis, and the composite update $b' = U(\sum_r w_r \phi_r(\Lambda))U^\top x^{(0)}$ is the whole inference step. Low graph frequencies are claimed to carry general rules, mid frequencies to refine relational structure, and high frequencies to mark contradictions and exceptions. The paper reports that this scheme reaches 88.1% accuracy on ProofWriter and 77.4% on CLUTRR, loses only 6.4% under adversarial perturbation, and aligns 87% of its spectral activations with symbolic proof steps, which it presents as evidence that the frequency decomposition carries genuine logical content rather than being an opaque surrogate.

Load-bearing premise

The load-bearing premise is that logical rules can be faithfully captured as diagonal filters in the Laplacian eigenbasis, with low frequencies carrying general rules and high frequencies carrying exceptions; the paper gives no proof and no direct empirical test that eigenvalue ordering corresponds to logical content.

Editorial extensions

If this is right

  • Spectral filtering can be evaluated with Chebyshev recurrences in $O(K|E|)$ time and $O(N)$ memory, so reasoning cost scales linearly with graph edges rather than with the quadratic cost of dense attention over long contexts.
  • Because rule aggregation is a bandwise sum of spectral operators, multiple rules can be composed and mixed in the same frequency domain, and each rule's contribution can be read off from the band energies of the output.
  • The reported robustness drop of only 6.4%, against drops of roughly 17--29% for the baselines, implies that spectral perturbation training and bounded filter responses can blunt structural adversarial attacks that degrade message-passing networks.
  • The 87% proof-band agreement on ProofWriter suggests that spectral activations can be used to produce or verify human-readable proof traces, not just predictions.
  • The CLUTRR-to-GraphQA transfer result of 71.3% with minimal fine-tuning indicates that reusable reasoning skill is carried by frequency profiles rather than by surface graph structure.

Reading between the lines

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

  • If the frequency-to-logic correspondence is real, the eigenvalue distribution of a knowledge graph becomes a predictor of which logical rules are learnable; that is a testable hypothesis the paper does not run.
  • The Chebyshev order $K$ functions as a spectral analogue of reasoning depth, so varying $K$ per instance should trace an accuracy-versus-latency frontier comparable to layer-depth trade-offs in transformers, an experiment the paper describes only qualitatively.
  • The same machinery could be turned into a consistency checker: logical contradictions in a knowledge base should appear as anomalous high-frequency energy, a signature that could be measured directly on inconsistent subsets of ProofWriter.
  • Because rule templates are linear filters, rules with overlapping spectral support may interfere; if that interference proves harmful, gated or non-linear spectral experts become necessary, which would refine rather than overturn the paper's central claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript introduces Spectral NSR, a neuro-symbolic reasoning architecture in which logical rules are encoded as diagonal spectral templates in the graph Laplacian eigenbasis and inference is claimed to proceed via frequency-selective filtering. Section 2 develops standard graph-signal-processing machinery—Laplacian eigendecomposition, functional calculus, Chebyshev polynomial filters, and Parseval energy accounting—and appends a symbolic rule-aggregation step. Section 3 lists a large set of architectural extensions (basis learning, rational filters, MoSE, proof-guided training, uncertainty quantification, LLM coupling, co-spectral transfer, adversarial robustness, GPU kernels, hypergraph Laplacians, causal interventions). Section 4 reports accuracy, latency, robustness, interpretability, and transfer numbers for ProofWriter and CLUTRR in a single table, with no experimental protocol. The abstract and conclusion make broad state-of-the-art claims based on these results.

Significance. Section 2's mathematical derivation is internally consistent and correctly reproduces textbook graph-signal-processing results: the functional calculus h(L)=U h(Λ)U⊤, the Chebyshev recurrence b_{k+1}=2L̃b_k−b_{k−1}, and the O(K|E|) complexity claim are all standard and sound. The paper also gives explicit gradients with respect to Chebyshev coefficients. These pieces are useful as a formulation exercise. However, the significance of the claimed contribution depends on two unsupported premises: that the eigenmode ordering of a knowledge-graph Laplacian carries logical semantics (low frequencies = general rules, high frequencies = contradictions), and that the empirical results in Table 1 are trustworthy. Neither is established. The proof-band agreement metric is a training objective, not an independent evaluation. The paper makes falsifiable predictions, but none are backed by reproducible experiments, code, or data. As a result, the claimed advances in accuracy, latency, robustness, and interpretability cannot be assessed.

major comments (4)
  1. [Section 4, Table 1] No experimental setup is reported: the table lists no dataset splits, hyperparameters, number of seeds, hardware, baseline implementations, or tuning procedures, and no code or data are provided. The absolute numbers (ProofWriter 88.1%, CLUTRR 77.4%, robustness drop −6.4) therefore cannot be reproduced or compared, and the central claims of superior accuracy, latency, and robustness are unsupported.
  2. [Section 3, 'Proof-Guided Training and Spectral Curriculum'; Section 4, 'Interpretability Results'] The 87% proof-band agreement is not evidence of emergent logical faithfulness: the training loss explicitly penalizes spectral energy distributions that do not correspond to valid proof bands, so the reported agreement is an optimized training objective rather than an independent measurement. The statement that this 'demonstrates the ability of spectral methods to faithfully ground reasoning in logical structure' is therefore circular.
  3. [Section 2, 'Projection to symbolic predicates and inference'] The pipeline is not fully spectral: after thresholding, the predicates feed a symbolic forward-chaining or resolution engine. This contradicts the abstract's claim that inference is performed 'directly in the graph spectral domain'. The framework as described is a spectral front-end followed by a classical symbolic solver, not a fully spectral reasoner.
  4. [Section 1 and Section 2, 'Symbolic rules as spectral templates'] The load-bearing assumption that low-frequency Laplacian modes encode general rules and high-frequency modes encode contradictions/exceptions is asserted without proof or empirical evidence. No experiment ties eigenvalue position to logical content; interpretability and correctness claims depend directly on this assumption. A concrete test would be needed, for example measuring whether proof-step alignments change systematically when eigenvalues are permuted.
minor comments (5)
  1. [Abstract and Section 1] 'We introduceSpectral NSR' is missing a space; also the abstract's 'fully spectral' claim is contradicted by the symbolic inference engine described in Section 2, as noted in major comment 3.
  2. [Table 1] Several rows contain run-together numbers (e.g., '36.1-7.8', '58.3 -22.5') and inconsistent column spacing; the table needs reformatting.
  3. [References] References are duplicated and inconsistently numbered: [1]–[5] overlap with later entries, and the reference list mixes arXiv and venue formats without a consistent style.
  4. [Section 4, Transfer Results] The CLUTRR-to-GraphQA transfer result is reported without any description of the fine-tuning protocol; the reader cannot tell what 'minimal fine-tuning' means or how the baselines were adapted.
  5. [Section 4, Evaluation Metrics] Proof-band agreement is listed as a metric but is never defined mathematically; the paper should specify how overlap between spectral activations and ground-truth proof steps is computed.

Circularity Check

1 steps flagged · score 6.0 of 10

The 87% proof-band agreement is a training target, not independent interpretability evidence; the remaining headline claims rest on an unreported table.

  1. fitted input called prediction [Section 3 (Proof-Guided Training and Spectral Curriculum) and Section 4 (Interpretability Results)]
    "Logical soundness is enforced by aligning spectral activations with symbolic proof steps. Spectral energy distributions are penalized if they do not correspond to valid proof bands, discouraging hallucinated reasoning. ... Spectral NSR achieves higher proof-band agreement scores, with 87% of spectral activations aligning with ground-truth symbolic proof steps on ProofWriter ... This demonstrates the ability of spectral methods to faithfully ground reasoning in logical structure."

    The proof-band agreement metric is defined in Section 4 as the overlap between spectral activations and ground-truth symbolic proof steps. The training objective in Section 3 explicitly penalizes spectral energy distributions whenever they do not correspond to valid proof bands. Therefore the reported 87% agreement is a direct consequence of the loss term, not an independent measurement of logical faithfulness. Reporting this optimized alignment as evidence that model decisions align closely with symbolic proof structures makes the headline interpretability claim reduce by construction to the training penalty. This is the pattern of a fitted input being called a prediction.

full rationale

The spectral derivation in Section 2 is self-contained and non-circular: the graph Fourier transform, functional calculus, Chebyshev filtering, and rule-template aggregation are standard identities that do not assume the paper's conclusions. However, the principal interpretability result is circular. Section 3's proof-guided training explicitly penalizes spectral energy distributions that do not correspond to valid proof bands, and Section 4 then reports the resulting proof-band agreement (87%) as evidence that spectral methods faithfully ground reasoning in logical structure. Since the reported metric is the very quantity optimized by the loss, it cannot independently validate the interpretability claim. No self-citation chain or uniqueness import is present; the rule-as-spectral-template mapping is an unproven assumption but not a circular reduction. The unreported experimental details for Table 1 raise reproducibility concerns, but those are correctness risks, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The framework rests on standard spectral graph theory plus several unverified domain assumptions about the semantic meaning of graph frequencies. The parameter count is moderate because the filter coefficients, rule weights, thresholds, and gate parameters are all learned or hand-chosen, while no values are reported. No invented entity has independent evidence.

free parameters (5)
  • Chebyshev filter coefficients theta_k
    Learned coefficients parameterize the spectral response h_theta(lambda) in Section 2; values are not reported.
  • Rule weights w_r
    Weights aggregate rule operators in the belief update b' = sum_r w_r Phi_r x^(0); values are not reported.
  • Diffusion and contradiction filter parameters tau and beta
    Appear in g_tau(lambda)=1/(1+tau*lambda) and phi_hp(lambda)=lambda/(lambda+beta) in Section 2; choices are not specified.
  • Predicate thresholds tau and temperature alpha
    Used in hard-thresholding and sigmoid projection to predicates in Section 2; values are not reported.
  • MoSE gating parameters alpha_b(x)
    Learned gating function in the mixture-of-spectral-experts extension; no training details provided.
assumptions (5)
  • standard math The graph Laplacian admits an orthonormal eigendecomposition L = U Lambda U^T and functional calculus h(L) = U h(Lambda) U^T.
    Standard spectral theorem for real symmetric matrices, used throughout Section 2.
  • ad hoc to paper Logical rules can be faithfully encoded as diagonal spectral templates phi_r(Lambda) in the graph Fourier basis.
    This is the core representational assumption of the framework, introduced in 'Symbolic rules as spectral templates' without proof or empirical validation.
  • ad hoc to paper Low graph frequencies correspond to general rules and high frequencies to contradictions or exceptions.
    Assumed in Sections 1 and 2 to justify interpretability claims; no evidence connects eigenvalue order to logical content.
  • domain assumption Knowledge graph reasoning can be modeled as belief propagation via graph signal filtering.
    The whole pipeline assumes a fixed or learned graph with a belief vector x^(0) whose evolution encodes logical inference; this is stated but not derived in Section 2.
  • ad hoc to paper Penalizing spectral energy distributions that do not match ground-truth proof steps enforces logical soundness.
    The 'proof-guided training' extension in Section 3 imposes this penalty and later reports proof-band agreement as evidence of interpretability.
invented entities (2)
  • Spectral templates phi_r(lambda)
    purpose: Represent logical rules as frequency responses applied through diagonal operators Phi_r = U phi_r(Lambda) U^T.
    No falsifiable prediction separates a rule template from an arbitrary filter; the mapping from logic to spectra is asserted, not measured.
  • Proof bands
    purpose: Spectral energy bands claimed to correspond to ground-truth symbolic proof steps for interpretability scoring.
    The bands are defined relative to proof annotations and are trained into the model via a penalty, so they provide no independent evidence of logical faithfulness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Eigenmodes to Proofs: Integrating Graph Spectral Operators with Symbolic Interpretable Reasoning." pith.science (2026). https://pith.science/paper/KIMT32AQ

@misc{pith2026250907017,
  author       = {Pith},
  title        = {Pith review of: From Eigenmodes to Proofs: Integrating Graph Spectral Operators with Symbolic Interpretable Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KIMT32AQ}},
  note         = {Machine review of arXiv:2509.07017}
}
read the original abstract

We introduce Spectral NSR, a fully spectral neuro-symbolic reasoning framework that embeds logical rules as spectral templates and performs inference directly in the graph spectral domain. By leveraging graph signal processing (GSP) and frequency-selective filters grounded in the Laplacian eigenstructure of knowledge graphs, the architecture unifies the interpretability of symbolic reasoning with the scalability and adaptability of spectral learning. Beyond the core formulation, we incorporate a comprehensive set of extensions, including dynamic graph and basis learning, rational and diffusion filters for sharper spectral selectivity, mixture-of-spectral-experts for modular specialization, proof-guided training with spectral curricula, and uncertainty quantification for calibrated confidence. Additional enhancements such as large language model coupling, co-spectral transfer alignment, adversarial robustness, efficient GPU kernels, generalized Laplacians, and causal interventions further expand the versatility of the framework. Empirical evaluation on state-of-the-art reasoning benchmarks such as ProofWriter and CLUTRR demonstrates that Spectral NSR achieves superior accuracy, faster inference, improved robustness to adversarial perturbations, and higher interpretability compared to leading baselines including transformers, message-passing neural networks, and neuro-symbolic logic programming systems. Spectral attribution and proof-band agreement analyses confirm that model decisions align closely with symbolic proof structures, while transfer experiments validate effective domain adaptation through co-spectral alignment. These results establish Spectral NSR as a scalable and principled foundation for the next generation of reasoning systems, offering transparency, robustness, and generalization beyond conventional approaches.

Figures

Figures reproduced from arXiv: 2509.07017 by the authors.

Figure 1
Figure 1. Spectral NSR architecture (robust, library-free layout). Top row: Inputs → spectral basis (L = D − A, L = UΛU ⊤) → Graph Fourier (ˆx = U ⊤x (0)) → spectral filtering (hθ(λ) = P k θkTk(λ˜), y = U hθ(Λ)U ⊤x (0)) → MoSE (h ⋆ (Λ) = P b αb(x) h (b) (Λ), dynamic K, B) → rule templates (Φr = Uϕr(Λ)U ⊤, b ′ = P r wrΦrx (0)) → predicate projection and symbolic inference. Bottom: extensions tap into the pipeline (dashed): lea… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [1]

    D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The emerging field of signal processing on graphs.IEEE Signal Processing Magazine, 30(3):83–98, 2013

    work page 2013

  2. [2]

    Manhaeve, S

    R. Manhaeve, S. Dumanˇ ci´ c, A. Kimmig, T. Demeester, and L. De Raedt. DeepProbLog: Neural probabilistic logic programming.Advances in Neural Information Processing Systems, 31, 2018

    work page 2018

  3. [3]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl. Neural message passing for quantum chemistry.Proceedings of the 34th International Conference on Machine Learning, 70:1263–1272, 2017

    work page 2017

  4. [4]

    Artificial intelligence, logic and formalizing common sense

    John McCarthy. Artificial intelligence, logic and formalizing common sense. InPhilosophical Logic and Artificial Intelligence, pages 161–190. Springer, 1981. 11 Graph Signal Processing 12

    work page 1981

  5. [5]

    Besold, Artur d’Avila Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, Kai-Uwe K¨ uhnberger, Luis C

    Tarek R. Besold, Artur d’Avila Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, Kai-Uwe K¨ uhnberger, Luis C. Lamb, Daniel Lowd, Prakash Panangaden, Gadi Pinkas, Henri Prade, Garry Schlobach, Evan T. Sher, Stuart M. Shapiro, and Michael Z. Wilk- erson. Neural-symbolic learning and reasoning: A survey and interpretation.Neurocomputin...

    work page 2017

  6. [6]

    Deep learning.Nature, 521(7553):436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, 2015

    2015

  7. [7]

    Towards a rigorous science of interpretable machine learn- ing.arXiv preprint arXiv:1702.08608, 2017

    Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learn- ing.arXiv preprint arXiv:1702.08608, 2017

    arXiv 2017

  8. [8]

    Zachary C. Lipton. The mythos of model interpretability.Communications of the ACM, 61(10):36–43, 2018

    work page 2018

Show all 30 references
  1. [9]

    Artur d’Avila Garcez and Luis C. Lamb. Neurosymbolic AI: The 3rd wave.arXiv preprint arXiv:2012.05876, 2019

    2012 arXiv

  2. [10]

    Neuro-symbolic AI.Dagstuhl Reports, 10(9):36–58, 2021

    Luciano Serafini and Artur d’Avila Garcez. Neuro-symbolic AI.Dagstuhl Reports, 10(9):36–58, 2021

    2021

  3. [11]

    DeepProbLog: Neural probabilistic logic programming

    Robin Manhaeve, Sebastijan Dumanˇ ci´ c, Angelika Kimmig, Thomas Demeester, and Luc De Raedt. DeepProbLog: Neural probabilistic logic programming. InAdvances in Neural Information Processing Systems, pages 3749–3759, 2018

    2018

  4. [12]

    Fan Yang, Zhilin Yang, and William W. Cohen. Differentiable learning of logical rules for knowledge base reasoning. InAdvances in Neural Information Processing Systems, pages 2316–2325, 2017

    2017

  5. [13]

    End-to-end differentiable proving

    Tim Rockt¨ aschel and Sebastian Riedel. End-to-end differentiable proving. InAdvances in Neural Information Processing Systems, pages 3788–3800, 2017

    2017

  6. [14]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. InProceedings of the 34th International Con- ference on Machine Learning, pages 1263–1272, 2017

    2017

  7. [15]

    Shuman, Sunil K

    David I. Shuman, Sunil K. Narang, Pascal Frossard, Antonio Ortega, and Pierre Van- dergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains.IEEE Signal Processing Magazine, 30(3):83–98, 2013

    2013

  8. [16]

    Moura, and Pierre Van- dergheynst

    Antonio Ortega, Pascal Frossard, Jelena Kovaˇ cevi´ c, Jos´ e M.F. Moura, and Pierre Van- dergheynst. Graph signal processing: Overview, challenges, and applications.Proceedings of the IEEE, 106(5):808–828, 2018

    2018

  9. [17]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Micha¨ el Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. InAdvances in Neural Information Processing Systems, pages 3844–3852, 2016

    2016

  10. [18]

    Ng, Michael I

    Andrew Y. Ng, Michael I. Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. InAdvances in Neural Information Processing Systems, pages 849–856, 2002. 12 Graph Signal Processing 13

    2002

  11. [19]

    Spectral methods for community detection and graph partitioning.Physical Review E, 88(4):042822, 2013

    Mark Newman. Spectral methods for community detection and graph partitioning.Physical Review E, 88(4):042822, 2013

    2013

  12. [20]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Infor- mation Processing Systems, pages 5998–6008, 2017

    2017

  13. [21]

    Graph attention networks.International Conference on Learning Representa- tions, 2018

    Petar Veliˇ ckovi´ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li` o, and Yoshua Bengio. Graph attention networks.International Conference on Learning Representa- tions, 2018

    2018

  14. [22]

    Adversarial attacks on neural networks for graph data.Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018

    Daniel Z¨ ugner, Amir Akbarnejad, and Stephan G¨ unnemann. Adversarial attacks on neural networks for graph data.Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018

    2018

  15. [23]

    ProofWriter: Generating implications, proofs, and abductive statements over natural language.arXiv preprint arXiv:2103.08696, 2021

    Oyvind Tafjord, Bhavana Dalvi, and Peter Clark. ProofWriter: Generating implications, proofs, and abductive statements over natural language.arXiv preprint arXiv:2103.08696, 2021

    2021 arXiv

  16. [24]

    Hamilton

    Koustuv Sinha, Shagun Sodhani, Joelle Pineau, and William L. Hamilton. CLUTRR: A diagnostic benchmark for inductive reasoning.arXiv preprint arXiv:1908.06177, 2019

    1908 arXiv

  17. [25]

    Program induction by rationale generation: Learning to solve and explain algebraic word problems.Proceedings of ACL, 2017

    Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems.Proceedings of ACL, 2017

    2017

  18. [26]

    Make up your mind! Adversarial generation of inconsistent natural language explanations.ACL, 2020

    Oana-Maria Camburu, Brendan Shillingford, Pasquale Minervini, Thomas Lukasiewicz, and Phil Blunsom. Make up your mind! Adversarial generation of inconsistent natural language explanations.ACL, 2020

    2020

  19. [27]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, pages 5998–6008, 2017

    2017

  20. [28]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry.International Conference on Machine Learn- ing, pages 1263–1272, 2017

    2017

  21. [29]

    DeepProbLog: Neural probabilistic logic programming.Advances in Neural Informa- tion Processing Systems, 2018

    Robin Manhaeve, Sebastijan Dumanˇ ci´ c, Angelika Kimmig, Thomas Demeester, and Luc De Raedt. DeepProbLog: Neural probabilistic logic programming.Advances in Neural Informa- tion Processing Systems, 2018

    2018

  22. [30]

    End-to-end differentiable proving.Advances in Neural Information Processing Systems, 2017

    Tim Rockt¨ aschel and Sebastian Riedel. End-to-end differentiable proving.Advances in Neural Information Processing Systems, 2017. 13

    2017

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.