Pith. sign in

REVIEW 3 major objections 4 minor 116 references

TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read TANGO makes GNN layer updates mimic stable energy descent plus a sideways flow, improving long-range signal propagation.

desk verdict Strong empirical paper on learned energy plus tangential flow for GNNs, with a theoretical story that overreaches: Proposition 4 is a decomposition identity, not a learnability proof. read the letter →

arxiv 2508.05070 v1 pith:DHJH436S submitted 2025-08-07 cs.LG

Moshe Eliasof , Eldad Haber , Carola-Bibiane Schönlieb This is my paper
classification cs.LG
keywords graphneuralnetworksenergy-basedmodelsLyapunovstabilitytangentialflowsoversquashinglong-rangebenchmarksheterophilicnodeclassificationdynamicalsystems
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 introduces TANGO, a graph neural network whose layer updates are read off a learned energy landscape: at each step, node features move partly in the direction that lowers the learned energy and partly along a learned direction orthogonal to that gradient, i.e. tangent to the level sets. The intended payoff is that the dynamics stay stable in the Lyapunov sense, keep evolving even where the energy landscape is flat or ill-conditioned, and can carry information across graph bottlenecks where ordinary message-passing or pure gradient flow stalls. On that basis the authors claim TANGO mitigates oversquashing and delivers leading results across graph property prediction, the Peptides long-range benchmarks, standard GNN benchmarks, and heterophilic node classification. A curious reader should care because the paper proposes a structured way to inject optimization and dynamical-systems reasoning into GNNs while staying within message passing.

What carries the argument

The decomposition of each update into an energy-descent component and an orthogonal (tangential) component. The energy is $V_G(H) = \frac{1}{n}\sum_{v\in V} \big(\mathrm{MLP}_E(\sigma(\mathrm{ENERGY\,GNN}(H)))_v\big)^2$; the tangential field is $T = M - \langle M, \hat{\nabla V_G}\rangle \hat{\nabla V_G}$, where $M$ is the output of a TANGENT GNN and $\hat{\nabla V_G}$ is the normalized energy gradient. This projection is the load-bearing step: it enforces $\langle T, \nabla V_G\rangle = 0$, so the energy can only be lowered by the $-\alpha\nabla V_G$ term while $T$ provides energy-preserving motion.

What would settle it

Compute the Hessian of the learned energy for a small trained TANGO model on a bottleneck graph (e.g., a barbell), form the Newton direction $N=(\nabla^2 V_G)^{-1}\nabla V_G$, and measure the cosine similarity between $N$ and the actual combined update $D=-\alpha\nabla V_G+\beta T$. If the similarity is no higher than for a random orthogonal direction, the claim that TANGO learns a quadratic-convergence direction is falsified for that configuration. A second check: on a barbell graph with one source node, if the non-tangent ablation propagates the signal as far as the full TANGO, the oversquas

Watch

Extended reading notes

Core claim

The central claim is that feature evolution should be governed by the ODE $\frac{dH}{dt} = -\alpha \nabla_H V_G(H) + \beta T_{V_G}(H)$, where $V_G$ is a learned, graph-dependent energy and $T_{V_G}(H)$ is a learned vector field projected onto the orthogonal complement of $\nabla_H V_G(H)$. Because $T_{V_G}$ is orthogonal to the gradient, energy $V_G$ can only decrease through the first term; because the second term survives when $\nabla_H V_G=0$, features are not frozen at flat or saddle regions. The authors further argue that the orthogonal term can, in principle, be learned so that the combined update coincides with the Newton direction $(\nabla^2 V_G)^{-1}\nabla V_G$, which would make con

Load-bearing premise

The load-bearing premise is that the message-passing TANGENT GNN can actually represent the orthogonal correction that turns the combined update into a Newton-like direction for the learned energy; the proof only shows such a decomposition exists for an arbitrary target vector, not that the parametrized network can compute it from graph features, nor that the learned energy's Hessian reflects the graph's bottleneck structure.

Editorial extensions

If this is right

  • Under the stated smoothness assumptions, the learned energy is a Lyapunov function: $V_G$ never increases along TANGO trajectories, so deep stacks inherit a stability guarantee.
  • In flat regions of the learned energy, TANGO still moves features, whereas pure gradient flow (the standard GNN-as-diffusion reading) would stall.
  • The orthogonal term can be tuned so the combined step matches the Newton direction; if the learned network realizes this, convergence no longer degrades with the Hessian's condition number, which the paper ties to oversquashing.
  • TANGO is a drop-in layer on top of MPNN backbones such as GatedGCN or GPS and improves their performance across graph property prediction, LRGB, ZINC/MNIST/CIFAR10/PATTERN/CLUSTER, and heterophilic node classification.
  • Ablations indicate that both ingredients matter: replacing the energy gradient with generic intermediate features, or dropping the orthogonal projection, degrades results compared to the full update.

Reading between the lines

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

  • A testable prediction the paper leaves implicit: TANGO's advantage over gradient-flow GNNs should grow with bottleneck severity (e.g., graph diameter or curvature), because the tangential term compensates exactly for the ill-conditioning that bottlenecks create.
  • The Newton-direction proof is existential: it shows a decomposition exists for any target vector, but not that the TANGENT GNN can compute it from local features. Whether message passing can actually realize the orthogonal Newton correction on given graphs is an open question.
  • The same orthogonal-decomposition trick could be exported to other learned-dynamics settings, such as neural ODEs, to add second-order-like corrections without computing a Hessian and without breaking a Lyapunov guarantee.
  • Because $\alpha_G$ is bounded and $\beta_G$ is unbounded, nothing in the architecture prevents the tangential flow from dominating in practice; a simple extension is to regularize $\beta_G$ or tie it to $\alpha_G$ to keep the energy-dissipation regime intact.
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 / 4 minor

Summary. The paper introduces TANGO, a graph neural architecture that discretizes a continuous-time dynamics in which node features evolve by a weighted sum of (i) the negative gradient of a learned graph energy V_G and (ii) a tangential flow obtained by projecting the output of a TANGENT GNN onto the orthogonal complement of that gradient. The energy is implemented by an ENERGY GNN followed by an MLP, with coefficients α_G (sigmoid-bounded) and β_G (MLP output) balancing the two terms. The authors prove energy dissipation (Proposition 1), evolution in flat regions (Proposition 2), and argue that the tangent term can realize a Newton-like direction (Proposition 4). They support the method with experiments on synthetic graph property prediction, LRGB peptides, Dwivedi et al. benchmarks, and heterophilic node classification, reporting consistent gains over strong baselines and ablations isolating the energy and tangential components.

Significance. If the central theoretical claims were supported, TANGO would be a valuable contribution: it couples a learned Lyapunov-like energy with a principled orthogonal flow, and the empirical evaluation is careful and unusually broad. The paper reports seed-averaged results, respects standard protocols and parameter budgets, and includes two well-designed ablations (TANGO-NON-ENERGY and TANGO-NON-TANGENT) that control for extra capacity and demonstrate that both the energy-gradient structure and the projection contribute to performance. The main weakness is that the advertised theoretical guarantees, especially the quadratic-convergence and oversquashing-mitigation claims, are much weaker than stated: Proposition 1 is an identity by construction, and Proposition 4 establishes only an abstract decomposition, not realizability by the proposed architecture.

major comments (3)
  1. [Section 4, Proposition 1 and Eq. (3)] The energy-dissipation result is true but it is a restatement of the design. In Eq. (3), T_V is defined so that ⟨T_V, ∇V_G⟩ = 0; substituting this into the chain-rule calculation makes Eq. (11) follow immediately. No separate Lyapunov guarantee is proved: the theorem does not establish boundedness of trajectories, existence of a valid equilibrium, or a decreasing neighborhood around one. The paper should either present this as an architectural property rather than a stability theorem, or supply the missing Lyapunov conditions.
  2. [Section 4, Proposition 4 and Appendix B] The proof of Proposition 4 is purely algebraic: for a given Newton direction N = (∇²V_G)^{-1}∇V_G, one sets T_V = N − α_G ∇V_G and chooses α_G = ⟨N, ∇V_G⟩/‖∇V_G‖² to enforce orthogonality. This works for any invertible Hessian and does not involve the TANGENT GNN, the orthogonal projection in Eq. (3), or the MLP producing α_G. The actual architecture computes M = σ(TANGENT_GNN(H;G)) and then projects it (Eq. 9), while α_G is constrained to [0,1] by the sigmoid in Eq. (8). The required α_G can lie outside [0,1] or be negative when the Newton direction is not positively aligned with the gradient. No representability or training-convergence argument is given. Thus the claim that TANGO 'can learn a quadratic convergence direction' is currently unsupported.
  3. [Section 4, Oversquashing discussion and Proposition 3] The bridge from Hessian conditioning to oversquashing mitigation is not established. Proposition 3 is a standard condition-number bound for gradient descent; the subsequent claim that a small second eigenvalue of the graph Laplacian makes the problem ill-conditioned and that TANGO can overcome this via a Newton-like tangential flow requires showing that the message-passing TANGENT GNN, with L_gnn layers, can approximate the inverse of a global Hessian. For bottleneck graphs (e.g., the Barbell example), the Hessian inverse is generically a dense global operator, so local message passing with small depth cannot represent it without further assumptions. Figure 2 is qualitative and does not quantify how much of the improvement comes from the tangential flow versus added nonlinearity. A constructive realizability result or a quantitative sensitivity experiment is needed to support the oversqu
minor comments (4)
  1. [Eq. (12)] There is a typo in the displayed equation: 'withD' appears in the text, and the notation H(l) is used inconsistently with H(ℓ) in the same paragraph.
  2. [Eqs. (8) and (10)] α_G is bounded in [0,1] by sigmoid, while β_G is an unconstrained MLP output. The paper calls both 'non-negative scalars' in Section 3.1, but Eq. (10) does not enforce non-negativity. Please clarify whether β_G can be negative and whether Proposition 4's assumption β_G = 1 is compatible with the implementation.
  3. [Appendix E.3, Table 9] The TANGO-NON-TANGENT ablation removes the projection but keeps the same M as the projection input. This is a good capacity control, but the comparison also changes the scale of the added term. A sentence explaining why this does not confound the effect of orthogonality would be useful.
  4. [References] References [56] and [57] appear to be the same paper (Kreuzer et al., 'Rethinking graph transformers with spectral attention'); likewise [102] and [103] duplicate the same GIN paper by Xu et al. Please de-duplicate.

Circularity Check

3 steps flagged · score 6.0 of 10

The theoretical guarantees reduce to the orthogonal-projection definition; the Newton-direction and oversquashing claims are restatements of an elementary decomposition, while the empirical benchmarks remain independent.

  1. self definitional [Section 4, Proposition 1 (Eq. 11), relying on Eq. (3) in Section 3.1]
    "As discussed in Section 3, we have by design, that ⟨TVG (H), ∇HVG(H)⟩ = 0. Therefore, d/dt VG(H) = −αG(H) ∥∇HVG(H)∥2 . Because αG(H) ≥ 0 by design, the energy is non-increasing"

    T_VG is defined in Eq. (3) as M minus its projection onto the normalized energy gradient, so ⟨T_VG, ∇V_G⟩=0 holds by construction, not as a discovered dynamical property. Substituting this identity into the chain rule yields exactly Eq. (11). The 'Lyapunov guarantee' is therefore a restatement of the orthogonal-projection definition plus the explicitly included −α∇V term, and carries no independent content beyond the design of the update.

  2. self definitional [Section 4, Proposition 2]
    "Suppose ∇HVG(H) = 0, and TVG (H) ̸= 0, then the TANGO flow in Equation (2) reads: dH/dt = βG(H)TVG (H)."

    This is the definition of Eq. (2) with the first summand set to zero. It asserts only that if the gradient vanishes and the tangential network happens to output a nonzero projected vector, the update is the tangential term. The 'can evolve in flat regions' claim is conditional on an unproved existence of T_V≠0; the proposition itself is direct substitution, not a dynamical result.

1 more flagged steps
  1. renaming known result [Section 4, Proposition 4 and Appendix B proof]
    "Then, we can express a Newton direction by the decomposition: N = αG(H) ∇HVG(H) + TVG (H). Solving for the orthogonal component yields: TVG (H) = N − αG(H) ∇HVG(H). To enforce orthogonality, we require: ⟨N − αG(H) ∇HVG(H), ∇HVG(H)⟩ = 0."

    The proof is the elementary fact that, for any nonzero vector g, any vector N can be uniquely written as αg + v with v⊥g, where α = ⟨N,g⟩/||g||². Plugging g=∇V_G and N=(∇²V_G)^{-1}∇V_G gives the displayed identities. This involves no property of the TANGENT GNN, the projection in Eq. (3), or the parameterization of α_G; it is a change of variables valid for any invertible Hessian. Thus the claim 'TANGO can learn a quadratic convergence direction' renames the orthogonal-decomposition identity as an architectural capability. It does not show the message-passing network can output an M whose projection equals the required v, nor that the sigmoid-bounded α_G in Eq. (8) can take the required value (which may be negative or >1).

full rationale

The empirical evaluation (Tables 1-7) is a standard train/test comparison against external baselines and is not circular: no fitted parameter is relabeled as a prediction, and no load-bearing self-citation is used. The violations are concentrated in the theoretical section. Propositions 1 and 2 are immediate consequences of the construction of T_VG in Eq. (3) and of the dynamics in Eq. (2). Proposition 4, the sole bridge to the oversquashing-mitigation claim, proves only that the Newton direction admits an orthogonal decomposition relative to the energy gradient; since every vector admits such a decomposition, the claimed ability of TANGO to 'learn' a Newton direction is the decomposition identity restated, not a property of the architecture. The sigmoid-bounded α_G and the message-passing TANGENT GNN are not used in the proof, so the step from existence of an abstract vector to learnability by TANGO is an additional, unsupported assumption. Because the central oversquashing narrative rests on this proposition, the theory is partially circular, while the benchmark results remain independent evidence. Score 6 reflects partial circularity confined to the theoretical claims.

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

The method introduces no new physical entities, but the central claims rely on standard control-theoretic results, a connectedness/regularity assumption, an unchecked assumption that the learned energy is task-aligned, and several tuned hyperparameters. The projection-based tangential flow is the core invented mechanism.

free parameters (5)
  • Euler step size epsilon = grid-searched over {0.001, 0.1, 1.0}
    Forward Euler discretization step in Equation (4), tuned per benchmark (Table 6).
  • Unrolled steps L = grid {1,5,10,20} or {2,4,8,16,32} depending on benchmark
    Number of dynamical layers, treated as a hyperparameter.
  • Internal GNN depth L_gnn = grid {1,2,4,8,16}
    Depth of ENERGY GNN and TANGENT GNN, tuned per benchmark.
  • Feature dimension d = grid {10,20,30} or {64,128,256}
    Hidden dimensionality, tuned per benchmark.
  • Weight decay = grid {0, 1e-6, 1e-5}
    Regularization strength, tuned per benchmark.
assumptions (5)
  • domain assumption The input graph G is connected and the energy VG is twice differentiable and bounded from below.
    Assumed in Section 4 'Assumptions and Notations' and used throughout the theoretical analysis.
  • standard math Lyapunov stability theorem from Khalil and Grizzle
    Invoked in Section 2 and Theorem 1 to interpret energy non-increase as stability.
  • standard math Classical linear convergence rate of gradient descent for a scalar objective
    Proposition 3, quoted from Nocedal and Wright, used to motivate ill-conditioning and the value of second-order directions.
  • domain assumption Forward Euler discretization preserves the qualitative behavior of the continuous-time dynamics
    The discrete update in Equation (4) is treated as inheriting the energy dissipation and tangential properties, though no discrete-time convergence proof is given.
  • ad hoc to paper The learned energy VG is a valid task surrogate whose minima align with correct predictions
    Implicit throughout Sections 1 and 3.1; no analysis ties energy minimization to task accuracy. If false, the Lyapunov guarantees are irrelevant to downstream performance.
invented entities (2)
  • Learned graph energy VG(H)
    purpose: Defines the gradient descent direction and the orthogonal projection in Equations (2)-(7).
    A learned output of an MLP over GNN features; not a physical quantity and has no falsifiable prediction outside the model.
  • Tangential flow T_VG(H)
    purpose: Learnable energy-preserving update that moves features along level sets of VG.
    Constructed by orthogonal projection of a learned message-passing output in Equation (3); no independent evidence beyond the construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows." pith.science (2026). https://pith.science/paper/DHJH436S

@misc{pith2026250805070,
  author       = {Pith},
  title        = {Pith review of: TANGO: Graph Neural Dynamics via Learned Energy and Tangential Flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHJH436S}},
  note         = {Machine review of arXiv:2508.05070}
}
read the original abstract

We introduce TANGO -- a dynamical systems inspired framework for graph representation learning that governs node feature evolution through a learned energy landscape and its associated descent dynamics. At the core of our approach is a learnable Lyapunov function over node embeddings, whose gradient defines an energy-reducing direction that guarantees convergence and stability. To enhance flexibility while preserving the benefits of energy-based dynamics, we incorporate a novel tangential component, learned via message passing, that evolves features while maintaining the energy value. This decomposition into orthogonal flows of energy gradient descent and tangential evolution yields a flexible form of graph dynamics, and enables effective signal propagation even in flat or ill-conditioned energy regions, that often appear in graph learning. Our method mitigates oversquashing and is compatible with different graph neural network backbones. Empirically, TANGO achieves strong performance across a diverse set of node and graph classification and regression benchmarks, demonstrating the effectiveness of jointly learned energy functions and tangential flows for graph neural networks.

Figures

Figures reproduced from arXiv: 2508.05070 by the authors.

Figure 1
Figure 1. Illustration of TANGO dynamics in a 2D feature space. We plot the level sets of a learned energy function and visualize the energy descent direction (green), the learned tangential direction (blue), and their combined vector (orange). The tangential component enables movement along level sets, while the descent component reduces energy. Together, they allow for more effective navigation of the learned energy landsca… view at source ↗
Figure 2
Figure 2. Comparison of propagation behaviors between gradient flow and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

116 extracted references · 57 canonical work pages

  1. [1]

    Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing

    Sami Abu-El-Haija, Bryan Perozzi, Amol Kapoor, Nazanin Alipourfard, Kristina Lerman, Hrayr Harutyunyan, Greg Ver Steeg, and Aram Galstyan. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In International Conference on Machine Learning, pages 21–29. PMLR, 2019

    2019

  2. [2]

    On the bottleneck of graph neural networks and its practical implications

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. In International Conference on Learning Representations, 2021. URL https: //openreview.net/forum?id=i80OPhOCVH2

    2021

  3. [3]

    On vanishing gradients, over- smoothing, and over-squashing in gnns: Bridging recurrent and graph learning

    Álvaro Arroyo, Alessio Gravina, Benjamin Gutteridge, Federico Barbero, Claudio Gallicchio, Xiaowen Dong, Michael Bronstein, and Pierre Vandergheynst. On vanishing gradients, over- smoothing, and over-squashing in gnns: Bridging recurrent and graph learning. arXiv preprint arXiv:2502.10818, 2025. URL https://arxiv.org/abs/2502.10818

    arXiv 2025

  4. [4]

    Accurate prediction of protein structures and interactions using a three-track neural network

    Minkyung Baek, Frank DiMaio, Ivan Anishchenko, Justas Dauparas, Sergey Ovchinnikov, Gyu Rie Lee, Jue Wang, Qian Cong, Lisa N Kinch, R Dustin Schaeffer, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science, 373(6557):871–876, 2021

    2021

  5. [5]

    Directional graph networks

    Dominique Beaini, Saro Passaro, Vincent Létourneau, Will Hamilton, Gabriele Corso, and Pietro Liò. Directional graph networks. In International Conference on Machine Learning, pages 748–758. PMLR, 2021

    2021

  6. [6]

    Graph Mamba: Towards Learning on Graphs with State Space Models, 2024

    Ali Behrouz and Farnoosh Hashemi. Graph Mamba: Towards Learning on Graphs with State Space Models, 2024. URL https://arxiv.org/abs/2402.08678

    arXiv 2024

  7. [7]

    Ravinder Bhattoo, Sayan Ranu, and N. M. Anoop Krishnan. Learning articulated rigid body dynamics with lagrangian graph neural networks. In Advances in Neural Information Processing Systems, volume 35, pages 29789–29800, 2022. URL https://arxiv.org/ abs/2209.11588

    work page Pith review arXiv 2022

  8. [8]

    Understanding oversquashing in gnns through the lens of effective resistance

    Mitchell Black, Zhengchao Wan, Amir Nayyeri, and Yusu Wang. Understanding oversquashing in gnns through the lens of effective resistance. In Proceedings of the 40th International Conference on Machine Learning, pages 2528–2547. PMLR, 2023

    2023

Show all 116 references
  1. [9]

    Beyond low-frequency information in graph convolutional networks

    Deyu Bo, Xiao Wang, Chuan Shi, and Huawei Shen. Beyond low-frequency information in graph convolutional networks. Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):3950–3957, May 2021. doi: 10.1609/aaai.v35i5.16514. URL https://ojs.aaai. org/index.php/AAAI/a...

    2021 doi

  2. [10]

    Convex optimization

    Stephen P Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004

    2004

  3. [11]

    Worrall, and Max Welling

    Johannes Brandstetter, Daniel E. Worrall, and Max Welling. Message passing neural PDE solvers. In International Conference on Learning Representations , 2022. URL https: //openreview.net/forum?id=vSix3HPYKSU

    2022

  4. [12]

    Residual Gated Graph ConvNets

    Xavier Bresson and Thomas Laurent. Residual Gated Graph ConvNets. arXiv preprint arXiv:1711.07553, 2018

    2018 arXiv

  5. [13]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021

    2021 arXiv

  6. [14]

    A note on over-smoothing for graph neural networks.arXiv preprint arXiv:2006.13318, 2020

    Chen Cai and Yusu Wang. A note on over-smoothing for graph neural networks.arXiv preprint arXiv:2006.13318, 2020

    2006 arXiv

  7. [15]

    Beltrami flow and neural diffusion on graphs

    Benjamin Chamberlain, James Rowbottom, Davide Eynard, Francesco Di Giovanni, Xiaowen Dong, and Michael Bronstein. Beltrami flow and neural diffusion on graphs. In Advances in Neural Information Processing Systems, volume 34, 2021

    2021

  8. [16]

    GRAND: Graph neural diffusion

    Benjamin Paul Chamberlain, James Rowbottom, Maria Gorinova, Stefan Webb, Emanuele Rossi, and Michael M Bronstein. GRAND: Graph neural diffusion. InInternational Conference on Machine Learning (ICML), pages 1407–1418. PMLR, 2021. 10

    2021

  9. [17]

    Simple and Deep Graph Convolutional Networks

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and Deep Graph Convolutional Networks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Resea...

    2020

  10. [18]

    Neural ordinary differential equations

    Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. In Advances in Neural Information Processing Systems, pages 6571– 6583, 2018

    2018

  11. [19]

    Adaptive universal generalized pagerank graph neural network

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. Adaptive universal generalized pagerank graph neural network. In International Conference on Learning Representations,

  12. [20]

    Gread: Graph neural reaction-diffusion networks

    Jeongwhan Choi, Seoyoung Hong, Noseong Park, and Sung-Bae Cho. Gread: Graph neural reaction-diffusion networks. In ICML, 2023

    2023

  13. [21]

    Beletsky, Konrad M

    Krzysztof Choromanski, Marcin Kuczynski, Jacek Cieszkowski, Paul L. Beletsky, Konrad M. Smith, Wojciech Gajewski, Gabriel De Masson, Tomasz Z. Broniatowski, Antonina B. Gorny, Leszek M. Kaczmarek, and Stanislaw K. Andrzejewski. Performers: A new approach to scaling transformer...

    2020 arXiv

  14. [22]

    From block-toeplitz matrices to differential equations on graphs: towards a general theory for scalable masked transformers

    Krzysztof Choromanski, Han Lin, Haoxian Chen, Tianyi Zhang, Arijit Sehanobish, Valerii Likhosherstov, Jack Parker-Holder, Tamas Sarlos, Adrian Weller, and Thomas Weingarten. From block-toeplitz matrices to differential equations on graphs: towards a general theory for scalable...

    2022

  15. [23]

    Principal Neighbourhood Aggregation for Graph Nets

    Gabriele Corso, Luca Cavalleri, Dominique Beaini, Pietro Liò, and Petar Veliˇckovi´c. Principal Neighbourhood Aggregation for Graph Nets. In Advances in Neural Information Processing Systems, volume 33, pages 13260–13271. Curran Associates, Inc., 2020

    2020

  16. [24]

    On over-squashing in message passing neural networks: the impact of width, depth, and topology

    Francesco Di Giovanni, Lorenzo Giusti, Federico Barbero, Giulia Luise, Pietro Liò, and Michael Bronstein. On over-squashing in message passing neural networks: the impact of width, depth, and topology. In Proceedings of the 40th International Conference on Machine Learning, IC...

    2023

  17. [25]

    Chamberlain, Thomas Markovich, and Michael M

    Francesco Di Giovanni, James Rowbottom, Benjamin P. Chamberlain, Thomas Markovich, and Michael M. Bronstein. Graph neural networks as gradient flows. In International Conference on Learning Representations (ICLR), 2023. URL https://arxiv.org/abs/2206.10991

    2023 arXiv

  18. [26]

    Gbk-gnn: Gated bi-kernel graph neural networks for modeling both homophily and heterophily

    Lun Du, Xiaozhou Shi, Qiang Fu, Xiaojun Ma, Hengyu Liu, Shi Han, and Dongmei Zhang. Gbk-gnn: Gated bi-kernel graph neural networks for modeling both homophily and heterophily. In Proceedings of the ACM Web Conference 2022, WWW ’22, page 1550–1558, New York, NY , USA, 2022. Ass...

    2022

  19. [27]

    Implicit generation and modeling with energy based models

    Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. In Advances in Neural Information Processing Systems, volume 32, 2019

    2019

  20. [28]

    A Generalization of Transformer Networks to Graphs

    Vijay Prakash Dwivedi and Xavier Bresson. A Generalization of Transformer Networks to Graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021

    2021

  21. [29]

    Graph neural networks with learnable structural and positional representations

    Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Graph neural networks with learnable structural and positional representations. InInternational Conference on Learning Representations, 2022. URL https://openreview.net/forum? id=wTTjnvGphYj

    2022

  22. [30]

    Long Range Graph Benchmark

    Vijay Prakash Dwivedi, Ladislav Rampášek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long Range Graph Benchmark. In Advances in Neural In- formation Processing Systems , volume 35, pages 22326–22340. Curran Associates, Inc., 2022. 11

    2022

  23. [31]

    Benchmarking graph neural networks

    Vijay Prakash Dwivedi, Chaitanya K Joshi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. Journal of Machine Learning Research, 24(43):1–48, 2023

    2023

  24. [32]

    PDE-GCN: Novel architectures for graph neural networks motivated by partial differential equations

    Moshe Eliasof, Eldad Haber, and Eran Treister. PDE-GCN: Novel architectures for graph neural networks motivated by partial differential equations. Advances in Neural Information Processing Systems, 34:3836–3849, 2021

    2021

  25. [33]

    Feature transportation improves graph neural networks

    Moshe Eliasof, Eldad Haber, and Eran Treister. Feature transportation improves graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11874–11882, 2024

    2024

  26. [34]

    On the temporal domain of differential equation inspired graph neural networks

    Moshe Eliasof, Eldad Haber, Eran Treister, and Carola-Bibiane B Schönlieb. On the temporal domain of differential equation inspired graph neural networks. In International Conference on Artificial Intelligence and Statistics, pages 1792–1800. PMLR, 2024

    2024

  27. [35]

    Lyapunov-based graph neural networks for adaptive control of multi-agent systems

    Brandon C Fallin, Cristian F Nino, Omkar Sudhir Patil, Zachary I Bell, and Warren E Dixon. Lyapunov-based graph neural networks for adaptive control of multi-agent systems. arXiv preprint arXiv:2503.15360, 2025

    2025 arXiv

  28. [36]

    Bronstein, and Ismail Ilkan Ceylan

    Ben Finkelshtein, Xingyue Huang, Michael M. Bronstein, and Ismail Ilkan Ceylan. Cooperative Graph Neural Networks. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=ZQcqXCuoxD

    2024

  29. [37]

    A large-scale database for graph representation learning

    Scott Freitas, Yuxiao Dong, Joshua Neil, and Duen Horng Chau. A large-scale database for graph representation learning. In J. Vanschoren and S. Yeung, editors, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, volume 1, 2021

    2021

  30. [38]

    Physics-informed graph neural galerkin networks: A unified framework for solving pde-governed forward and inverse problems

    Han Gao, Matthew J Zahr, and Jian-Xun Wang. Physics-informed graph neural galerkin networks: A unified framework for solving pde-governed forward and inverse problems. Computer Methods in Applied Mechanics and Engineering, 390:114502, 2022

    2022

  31. [39]

    Diffusion Improves Graph Learning

    Johannes Gasteiger, Stefan Weiß enberger, and Stephan Günnemann. Diffusion Improves Graph Learning. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019

    2019

  32. [40]

    On the trade-off between over-smoothing and over-squashing in deep graph neural networks

    Jhony H Giraldo, Konstantinos Skianis, Thierry Bouwmans, and Fragkiskos D Malliaros. On the trade-off between over-smoothing and over-squashing in deep graph neural networks. In Proceedings of the 32nd ACM international conference on information and knowledge management, pages...

    2023

  33. [41]

    Anti-Symmetric DGN: a stable architecture for Deep Graph Networks

    Alessio Gravina, Davide Bacciu, and Claudio Gallicchio. Anti-Symmetric DGN: a stable architecture for Deep Graph Networks. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=J3Y7cgZOOS

    2023

  34. [42]

    On oversquashing in graph neural networks through the lens of dynamical systems

    Alessio Gravina, Moshe Eliasof, Claudio Gallicchio, Davide Bacciu, and Carola-Bibiane Schönlieb. On oversquashing in graph neural networks through the lens of dynamical systems. In The 39th Annual AAAI Conference on Artificial Intelligence, 2025

    2025

  35. [43]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=uYLFoz1vlAC

    2022

  36. [44]

    Egc: Image generation and classification via a diffusion energy-based model

    Qiushan Guo, Yifan Zhang, Yifan Wang, Yizhou Wang, and Hongsheng Li. Egc: Image generation and classification via a diffusion energy-based model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12345–12354, 2023

    2023

  37. [45]

    Drew: Dynamically rewired message passing with delay

    Benjamin Gutteridge, Xiaowen Dong, Michael M Bronstein, and Francesco Di Giovanni. Drew: Dynamically rewired message passing with delay. In International Conference on Machine Learning, pages 12252–12267. PMLR, 2023

    2023

  38. [46]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 1025–1035. Curran Associates Inc., 2017. ISBN 9781510860964. 12

    2017

  39. [47]

    From continuous dynamics to graph neural networks: Neural diffusion and beyond

    Andi Han, Dai Shi, Lequan Lin, and Junbin Gao. From continuous dynamics to graph neural networks: Neural diffusion and beyond. arXiv preprint arXiv:2310.10121, 2023

    2023 arXiv

  40. [48]

    Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks

    Simon Heilig, Alessio Gravina, Alessandro Trenta, Claudio Gallicchio, and Davide Bacciu. Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks. In The Thirteenth International Conference on Learning Representations, 2025. URL https: //openreview...

    2025

  41. [49]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9 (8):1735–1780, 1997. doi: 10.1162/neco.1997.9.8.1735

    1997 doi

  42. [50]

    Strategies for Pre-training Graph Neural Networks

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. Strategies for Pre-training Graph Neural Networks. In International Con- ference on Learning Representations, 2020. URL https://openreview.net/forum?id= HJlWWJSFDH

    2020

  43. [51]

    Global self- attention as a replacement for graph convolution

    Md Shamim Hussain, Mohammed J Zaki, and Dharmashankar Subramanian. Global self- attention as a replacement for graph convolution. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 655–665, 2022

    2022

  44. [52]

    Spectral graph pruning against over-squashing and over-smoothing

    Adarsh Jamadandi, Celia Rubio-Madrigal, and Rebekka Burkholz. Spectral graph pruning against over-squashing and over-smoothing. In Advances in Neural Information Processing Systems, 2024

    2024

  45. [53]

    Nonlinear systems, volume 3

    Hassan K Khalil and Jessy W Grizzle. Nonlinear systems, volume 3. Prentice hall Upper Saddle River, NJ, 2002

    2002

  46. [54]

    Kipf and M

    T. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. Proceedings of the International Conference on Learning Representations, 2016

    2016

  47. [55]

    Bayan Bruss, and Tom Goldstein

    Kezhi Kong, Jiuhai Chen, John Kirchenbauer, Renkun Ni, C. Bayan Bruss, and Tom Goldstein. GOAT: A global transformer on large-scale graphs. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceed- ings of the ...

    2023

  48. [57]

    Rethinking graph transformers with spectral attention

    Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 34:21618–21629, 2021

    2021

  49. [58]

    Sven Kreuzer, Michael Reiner, and Stefan D. D. De Villiers. Sant: Structural attention networks for graphs. Proceedings of the 38th International Conference on Machine Learning (ICML), 2021

    2021

  50. [59]

    A tutorial on energy-based learning

    Yann LeCun, Sumit Chopra, Raia Hadsell, Marc’Aurelio Ranzato, and Fu-Jie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0):1–59, 2006

    2006

  51. [60]

    Finding global homophily in graph neural networks when meeting heterophily

    Xiang Li, Renyu Zhu, Yao Cheng, Caihua Shan, Siqiang Luo, Dongsheng Li, and Weining Qian. Finding global homophily in graph neural networks when meeting heterophily. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Procee...

    2022

  52. [61]

    Toloker Graph: Interaction of Crowd Annotators, February 2023

    Daniil Likhobaba, Nikita Pavlichenko, and Dmitry Ustalov. Toloker Graph: Interaction of Crowd Annotators, February 2023. URL https://doi.org/10.5281/zenodo.7620796. 13

    2023 doi

  53. [62]

    GraphEBM: Molecular graph generation with energy-based models

    Meng Liu, Keqiang Yan, Bora Oztekin, and Shuiwang Ji. GraphEBM: Molecular graph generation with energy-based models. In Energy Based Models Workshop - ICLR 2021, 2021. URL https://openreview.net/forum?id=Gc51PtL_zYw

    2021

  54. [63]

    Li, Jian Tang, Guy Wolf, and Stefanie Jegelka

    Sitao Luan, Chenqing Hua, Qincheng Lu, Liheng Ma, Lirong Wu, Xinyu Wang, Minkai Xu, Xiao-Wen Chang, Doina Precup, Rex Ying, Stan Z. Li, Jian Tang, Guy Wolf, and Stefanie Jegelka. The heterophilic graph learning handbook: Benchmarks, models, theoretical analysis, applications a...

    2024 arXiv

  55. [64]

    Graph inductive biases in transformers without message passing

    Liheng Ma, Chen Lin, Derek Lim, Adriana Romero-Soriano, Puneet K Dokania, Mark Coates, Philip Torr, and Ser-Nam Lim. Graph inductive biases in transformers without message passing. In International Conference on Machine Learning, pages 23321–23337. PMLR, 2023

    2023

  56. [65]

    Qdc: Quantum diffusion convolution kernels on graphs, 2023

    Thomas Markovich. Qdc: Quantum diffusion convolution kernels on graphs, 2023

    2023

  57. [66]

    A fractional graph laplacian approach to oversmoothing

    Sohir Maskey, Raffaele Paolino, Aras Bacho, and Gitta Kutyniok. A fractional graph laplacian approach to oversmoothing. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=kS7ED7eE74

    2023

  58. [67]

    Simplifying approach to node classification in graph neural networks

    Sunil Kumar Maurya, Xin Liu, and Tsuyoshi Murata. Simplifying approach to node classification in graph neural networks. Journal of Computational Science , 62:101695,

  59. [68]

    Attending to graph transformers

    Luis Müller, Mikhail Galkin, Christopher Morris, and Ladislav Rampášek. Attending to graph transformers. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=HhbqHBBrfZ

    2024

  60. [69]

    Nocedal and S

    J. Nocedal and S. Wright. Numerical Optimization. Springer, New York, 1999

    1999

  61. [70]

    Revisiting graph neural networks: All we have is low-pass filters

    Hoang Nt and Takanori Maehara. Revisiting graph neural networks: All we have is low-pass filters. arXiv preprint arXiv:1905.09550, 2019

    1905 arXiv

  62. [71]

    Graph neural networks exponentially lose expressive power for node classification

    Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=S1ldO2EFPr

    2020

  63. [72]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

  64. [73]

    Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks

    Jie Peng, Runlin Lei, and Zhewei Wei. Beyond over-smoothing: Uncovering the trainability challenges in deep graph neural networks. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1878–1887, 2024

    2024

  65. [74]

    A critical look at the evaluation of GNNs under heterophily: Are we re- ally making progress? In The Eleventh International Conference on Learning Representations,

    Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. A critical look at the evaluation of GNNs under heterophily: Are we re- ally making progress? In The Eleventh International Conference on Learning Representations,

  66. [77]

    Recipe for a general, powerful, scalable graph transformer (graphgps)

    Petr Rampášek, Mikhail Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer (graphgps). In Advances in Neural Information Processing Systems, volume 35, pages 28877–28890, 2022. URL https://arxi...

    2022 arXiv

  67. [78]

    Recipe for a General, Powerful, Scalable Graph Transformer

    Ladislav Rampášek, Mikhail Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a General, Powerful, Scalable Graph Transformer. Advances in Neural Information Processing Systems, 35, 2022

    2022

  68. [79]

    Graph neural networks for materials science and chemistry

    Patrick Reiser, Marlen Neubert, André Eberhard, Luca Torresi, Chen Zhou, Chen Shao, Houssam Metni, Clint van Hoesel, Henrik Schopmans, Timo Sommer, et al. Graph neural networks for materials science and chemistry. Communications Materials, 3(1):93, 2022

    2022

  69. [80]

    Lyanet: A lyapunov framework for training neural odes

    Ivan Dario Jimenez Rodriguez, Aaron Ames, and Yisong Yue. Lyanet: A lyapunov framework for training neural odes. In International conference on machine learning, pages 18687–18703. PMLR, 2022

    2022

  70. [81]

    Graph-coupled oscillator networks

    T Konstantin Rusch, Ben Chamberlain, James Rowbottom, Siddhartha Mishra, and Michael Bronstein. Graph-coupled oscillator networks. In International Conference on Machine Learning, pages 18888–18909. PMLR, 2022

    2022

  71. [82]

    Konstantin Rusch, Michael M

    T. Konstantin Rusch, Michael M. Bronstein, and Siddhartha Mishra. A Survey on Oversmooth- ing in Graph Neural Networks. arXiv preprint arXiv:2303.10993, 2023

    2023 arXiv

  72. [83]

    Deep neural networks motivated by partial differential equations

    Lars Ruthotto and Eldad Haber. Deep neural networks motivated by partial differential equations. Journal of Mathematical Imaging and Vision, 62:352–364, 2020

    2020

  73. [84]

    Qm/mm methods for biomolecular systems

    Hans M Senn and Walter Thiel. Qm/mm methods for biomolecular systems. Angewandte Chemie International Edition, 48(7):1198–1229, 2009

    2009

  74. [85]

    Exposition on over-squashing prob- lem on gnns: Current methods, benchmarks and challenges

    Dai Shi, Andi Han, Lequan Lin, Yi Guo, and Junbin Gao. Exposition on over-squashing prob- lem on gnns: Current methods, benchmarks and challenges. arXiv preprint arXiv:2311.07073, 2023

    2023 arXiv

  75. [86]

    Masked label prediction: Unified message passing model for semi-supervised classifica- tion

    Yunsheng Shi, Zhengjie Huang, Shikun Feng, Hui Zhong, Wenjing Wang, and Yu Sun. Masked label prediction: Unified message passing model for semi-supervised classifica- tion. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Confer- ence on Artificial Int...

    2021 doi

  76. [87]

    Rahmani, and Marzieh Aghaei

    Behzad Shirzad, Amir M. Rahmani, and Marzieh Aghaei. Exphormer: Sparse attention for graphs. Proceedings of the 40th International Conference on Machine Learning (ICML), 2023

    2023

  77. [88]

    Where did the gap go? reassessing the long-range graph benchmark

    Jan Tönshoff, Martin Ritzert, Eran Rosenbluth, and Martin Grohe. Where did the gap go? reassessing the long-range graph benchmark. In The Second Learning on Graphs Conference,

  78. [89]

    Walking out of the weis- feiler leman hierarchy: Graph learning beyond message passing

    Jan Tönshoff, Martin Ritzert, Hinrikus Wolf, and Martin Grohe. Walking out of the weis- feiler leman hierarchy: Graph learning beyond message passing. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id= vgXnEyeWVY

    2023

  79. [90]

    Bronstein

    Jake Topping, Francesco Di Giovanni, Benjamin Paul Chamberlain, Xiaowen Dong, and Michael M. Bronstein. Understanding over-squashing and bottlenecks on graphs via curvature. In International Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=7UmjRGzp-A

    2022

  80. [91]

    Capturing graphs with hypo-elliptic diffusions

    Csaba Toth, Darrick Lee, Celia Hacker, and Harald Oberhauser. Capturing graphs with hypo-elliptic diffusions. In Advances in Neural Information Processing Systems, 2022

    2022

  81. [92]

    Vaswani et al

    A. Vaswani et al. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017

    2017

  82. [93]

    URL https://openreview.net/forum?id=rIUjwxc5lj

  83. [94]

    Graph-mamba: Towards long-range graph sequence modeling with selective state spaces

    Chloe Wang, Oleksii Tsepa, Jun Ma, and Bo Wang. Graph-mamba: Towards long-range graph sequence modeling with selective state spaces. arXiv preprint arXiv:2402.00789, 2024

    2024 arXiv

  84. [95]

    The heterophilic snowflake hypothesis: Training and empowering gnns for heterophilic graphs

    Kun Wang, Guibin Zhang, Xinnan Zhang, Junfeng Fang, Xun Wu, Guohao Li, Shirui Pan, Wei Huang, and Yuxuan Liang. The heterophilic snowflake hypothesis: Training and empowering gnns for heterophilic graphs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery ...

  85. [96]

    How powerful are spectral graph neural networks

    Xiyuan Wang and Muhan Zhang. How powerful are spectral graph neural networks. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning , volume 162 of Proceedin...

    2022

  86. [97]

    Dissecting the Diffusion Process in Linear Graph Convolutional Networks

    Yifei Wang, Yisen Wang, Jiansheng Yang, and Zhouchen Lin. Dissecting the Diffusion Process in Linear Graph Convolutional Networks. In Advances in Neural Information Processing Systems, volume 34, pages 5758–5769. Curran Associates, Inc., 2021

    2021

  87. [98]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Represen- tations, 2018. URL https://openreview.net/forum?id=rJXMpikCZ. 15

    2018

  88. [99]

    Recent successes of the energy landscape theory of protein folding and function

    Peter G Wolynes. Recent successes of the energy landscape theory of protein folding and function. Quarterly reviews of biophysics, 38(4):405–410, 2005

    2005

  89. [100]

    Continuous graph neural networks

    Louis-Pascal Xhonneux, Meng Qu, and Jian Tang. Continuous graph neural networks. In Proceedings of the 37th International Conference on Machine Learning, pages 10432–10441,

  90. [101]

    A theory of generative convnet

    Jianwen Xie, Yuting Lu, Ruiqi Gao, Honglak Zhuang, and Ying Nian Wu. A theory of generative convnet. International Conference on Machine Learning, pages 2635–2644, 2016

    2016

  91. [103]

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

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations , 2019. URL https://openreview.net/forum?id=ryGs6iA5Km

    2019

  92. [104]

    ACMP: Allen-cahn message passing with attractive and repulsive forces for graph neural networks

    Yuelin Wang, Kai Yi, Xinliang Liu, Yu Guang Wang, and Shi Jin. ACMP: Allen-cahn message passing with attractive and repulsive forces for graph neural networks. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview. net/forum?id=4fZc_79Lrqs

    2023

  93. [105]

    H., Lihong Wang, S

    Manzil Zaheer, Guru prasad G. H., Lihong Wang, S. V . K. N. L. Wang, Yujia Li, Jakub Koneˇcný, Shalmali Joshi, Danqi Chen, Jennifer R. R., Zhenyu Zhang, Shalini Devaraj, and Srinivas Narayanan. Bigbird: Transformers for longer sequences. Proceedings of the 37th International C...

    2020 arXiv

  94. [106]

    Rethinking the expressive power of GNNs via graph biconnectivity

    Bohang Zhang, Shengjie Luo, Liwei Wang, and Di He. Rethinking the expressive power of GNNs via graph biconnectivity. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=r9hNv76KoT3

    2023

  95. [107]

    From stars to subgraphs: Uplifting any GNN with local structure awareness

    Lingxiao Zhao, Wei Jin, Leman Akoglu, and Neil Shah. From stars to subgraphs: Uplifting any GNN with local structure awareness. InInternational Conference on Learning Representations,

  96. [108]

    Beyond homophily in graph neural networks: Current limitations and effective designs

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information P...

    2020

  97. [109]

    Rossi, Anup Rao, Tung Mai, Nedim Lipka, Nesreen K

    Jiong Zhu, Ryan A. Rossi, Anup Rao, Tung Mai, Nedim Lipka, Nesreen K. Ahmed, and Danai Koutra. Graph neural networks with heterophily. Proceedings of the AAAI Conference on Artificial Intelligence, 35(12):11168–11176, May 2021. doi: 10.1609/aaai.v35i12.17332. URL https://ojs.a...

    2021 doi

  98. [110]

    Ordinary differential equations on graph networks

    Juntang Zhuang, Nicha Dvornek, Xiaoxiao Li, and James S Duncan. Ordinary differential equations on graph networks. 2020. 17 A Additional Related Work Oversquashing in Graph Learning. Graph neural networks (GNNs) typically operate through message-passing mechanisms, aggregating...

    2020

  99. [111]

    Graphormer: A transformer for graphs

    Zhitao Ying and Jure Leskovec. Graphormer: A transformer for graphs. Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, 2021

    2021

  100. [115]

    URL https://openreview.net/forum?id=Mspk_WYKoEH. 16

  101. [119]

    However, the quadratic complexity of full attention in these models raises scalability concerns, motivating interest in sparse attention mechanisms [105, 21, 87]

    to preserve structural identity. However, the quadratic complexity of full attention in these models raises scalability concerns, motivating interest in sparse attention mechanisms [105, 21, 87]. An alternative line of work explores non-local dynamics to enhance expressivity w...

  102. [120]

    Experimental Setup

    and report mean and standard deviation across three different random seeds for each experiment. Experimental Setup. We follow the evaluation protocol established in Dwivedi et al.[30], including dataset usage, training strategy, and model capacity constraints. Hyperparameter t...

  103. [2019]

    URL https://arxiv.org/abs/1911.07532

    1911 arXiv

  104. [2020]

    URL https://proceedings.mlr.press/v119/xhonneux20a.html

  105. [2021]

    URL https://openreview.net/forum?id=n6jl7fLxrP

  106. [2022]

    doi: https://doi.org/10.1016/j.jocs.2022.101695

    ISSN 1877-7503. doi: https://doi.org/10.1016/j.jocs.2022.101695. URL https: //www.sciencedirect.com/science/article/pii/S1877750322000990

    2022

  107. [2023]

    URL https://openreview.net/forum?id=tJbbQfw-5wv

  108. [2024]

    ISBN 9798400704901

    Association for Computing Machinery. ISBN 9798400704901. doi: 10.1145/3637528. 3671791

    doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.