acceptodds
Under review as a conference paper at ICLR 2027

WaveGSSM: Graph Wave State Space Models for Propagating Spatio-Temporal Patterns

Abstract

Spatio-temporal graph models typically encode each snapshot with a GNN and then connect the resulting representations through a temporal module. This space-then-time design is effective, yet it does not explicitly represent how a pattern moves across the graph. We show empirically that, for a propagating process, the same present field can lead to different futures when its recent rate of change differs, motivating an explicit representation of motion in the predictive state. We introduce WaveGSSM, a second-order graph state-space model that maintains two coupled latent states at each node, one for the current pattern and one for its temporal rate of change. A graph-wave transition updates the motion state through graph interactions and uses it to advance the pattern state, coupling spatial propagation and temporal evolution within a single rollout. We evaluate WaveGSSM on four temporal-graph benchmarks and global weather forecasting. It consistently achieves the best mean performance across the temporal-graph benchmarks and reduces the geopotential RMSE by 20.2% on average for 1- to 5-day weather forecasts relative to a backbone-matched snapshot model, while better preserving large-scale atmospheric patterns.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural networks, State space models, graph representation learning, Spatio-temporal graph

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_26hTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces WaveGSSM, a second-order graph state-space model for propagating spatiotemporal signals. The model explicitly maintains a pattern state and its rate of change, using a graph-wave transition to couple spatial propagation with temporal evolution. Synthetic experiments and modal analysis illustrate the role of motion information in prediction. The authors evaluate the model on four temporal graph node classification benchmarks and ERA5 global weather forecasting. WaveGSSM achieves the highest mean performance on all four benchmarks and reduces geopotential-height RMSE by 20.2% on average over forecast days 1 to 5 relative to a backbone-matched GraphCast-style baseline. Ablation experiments further assess the contribution of motion-state carry.

Critical strengths
  1. Jointly modeling the pattern and its rate of change provides a reasonable inductive bias for propagating graph signals. The authors clearly motivate this design through synthetic examples and modal analysis that show how the same current pattern can lead to different futures depending on its rate of change.
  2. The paper is well written, with clear and visually appealing figures. The propagation examples and architecture diagram make the design motivation and state updates easy to follow.
Critical weaknesses
  1. The baseline comparisons are insufficient. The synthetic experiments mainly compare against GCN-LSTM and GCN-SSM. Although the weather experiments include GraphSSM and GNN-LSTM, comparisons with other relevant architectures that couple space and time remain limited. This makes it difficult to assess the proposed structure’s advantages over existing architectures.
  2. The methodological advance remains unclear. The fact that identical current patterns can lead to different futures when their rates of change differ is a basic property of second-order dynamical systems and carries limited weight as a standalone contribution. Assessing the novelty requires a clearer account of the substantive differences from existing graph oscillator, graph wave, and graph state-space methods, and the benefits those differences bring.
Questions for authors
  1. To my knowledge, the spatial stage of a space-then-time architecture can already incorporate temporal dependencies. For example, EvolveGCN uses an RNN to evolve GCN parameters, and GraphSSM mixes information from adjacent snapshots. Although these methods appear in Table 1, the synthetic experiments in Figure 1 mainly compare against GCN-LSTM and GCN-SSM. Include baselines whose spatial encoders incorporate temporal information, or explain their exclusion.
  2. The weather experiments need comparisons with methods that jointly model space and time. For example, CITRUS, cited in the related work, operates on product graphs and has been evaluated on weather forecasting. Compare it using the same data setup, or explain specifically why it is unsuitable for this task.
  3. Most baseline results in Table 1 are taken from the GraphSSM paper, while GraphSSM itself is rerun by the authors. State whether the cited results use the same preprocessing, exact node splits, and validation-based model selection. Where these differ, provide results under a common protocol.
  4. The first contribution and Observation 1 establish that identical patterns can lead to different futures when their rates of change differ. This is a basic property of second-order dynamical systems. Frame it as motivation for the design and avoid presenting the fact itself as a major contribution.
  5. The benefit of an explicit rate-of-change state over standard temporal memory needs further validation. For example, give GCN-LSTM or GCN-SSM explicit differences between consecutive frames while maintaining comparable parameter budgets. This would help separate the benefits of difference cues from those of the second-order state structure.
  6. Specify the differences from GraphCON and Graph Wave Networks in terms of state definitions and update equations. Distinguishing “representation propagation” from “motion-state modeling” describes different uses of the dynamics but does not explain why the latter is more advantageous.
  7. Appendix B.4 states that each weather prediction call reconstructs (U,V)(U,V) from two historical frames without directly carrying the states across forecast days. Explain why reconstruction was chosen over continuously updating the latent states, and clarify that motion-state carry occurs only within each processor call.
  8. The node classification ablation retains the initial V0V_0, whereas the weather ablation also zeros V0V_0 before the first transition. Add a weather ablation that retains V0V_0 and removes only subsequent state carry. Retrain it under the same conditions to isolate the contribution of state carry.
  9. Equation (1) assigns learnable weights to support-graph edges, while Appendix C uses disjoint training, validation, and test graphs. Specify how edge weights are generated or transferred to unseen graphs. If the synthetic experiments use a different parameterization, explain how it differs from the main model.
  10. Report parameter counts for the complete models, including learnable edge weights. Measure inference time for an ERA5 one-day forecast on the same GPU model with a batch size of 1, and state whether it includes grid-to-mesh encoding, mesh processing, and mesh-to-grid decoding.
7 comments
AuthorsOpenReview ↗

Response to Reviewer 26hT (Part 1/3)

We sincerely thank the reviewer for the thorough and detailed review. The questions have helped us clarify several important aspects of the paper. Since the two weaknesses overlap with the questions, we address them together: W1 with Q1–Q2, and W2 with Q4 and Q6.


W1 & Q1. To my knowledge, the spatial stage of a space-then-time architecture can already incorporate temporal dependencies. For example, EvolveGCN uses an RNN to evolve GCN parameters, and GraphSSM mixes information from adjacent snapshots. Although these methods appear in Table 1, the synthetic experiments in Figure 1 mainly compare against GCN-LSTM and GCN-SSM. Include baselines whose spatial encoders incorporate temporal information, or explain their exclusion.

A1. We appreciate the suggestion to include these baselines. EvolveGCN [R7] and GraphSSM [R8] both pair a GNN with a separate temporal module (an RNN that evolves GCN weights, or an SSM over snapshots), and neither keeps a node-level motion state updated through the graph. Figure 1 uses GCN-LSTM and GCN-SSM as simple space-then-time baselines to illustrate the problem. Following your suggestion, we further added the official EvolveGCN-H and GraphSSM, adapted to the Figure 1 task:

ModelParamsPattern correlation ↑Localization error ↓
GCN-LSTM10,7850.9353 ± 0.01160.01852 ± 0.00355
GCN-SSM10,7310.6839 ± 0.00950.09470 ± 0.00338
EvolveGCN-H10,4680.3667 ± 0.39230.24221 ± 0.20927
GraphSSM10,9460.9783 ± 0.00610.00683 ± 0.00133
Our WaveGSSM10,6410.9873 ± 0.00110.00481 ± 0.00099

WaveGSSM is best on both metrics, with about 30% lower localization error than GraphSSM. We have added these results to Appendix C (page 20) of the revised manuscript.


W1 & Q2. The weather experiments need comparisons with methods that jointly model space and time. For example, CITRUS, cited in the related work, operates on product graphs and has been evaluated on weather forecasting. Compare it using the same data setup, or explain specifically why it is unsuitable for this task.

A2. Thank you for pointing us to CITRUS [R6]. We adapted CITRUS's product-graph processor to our ERA5 setting, using the same data setup. The results are shown as follows:

ModelDay 1Day 3Day 5Mean (Days 1–5)
CITRUS23.5347.7167.0846.57
Our WaveGSSM12.7130.0450.0130.72
RMSE reduction45.99%37.04%25.44%34.03%

Geopotential-height RMSE (m, lower is better), averaged over six pressure levels. WaveGSSM has lower RMSE at all three leads and reduces the mean RMSE over days 1–5 by 34.03%. We have added this comparison to Section 4.3 (page 8) and Appendix F.1 (page 26) of the revised manuscript.


Q3. Most baseline results in Table 1 are taken from the GraphSSM paper, while GraphSSM itself is rerun by the authors. State whether the cited results use the same preprocessing, exact node splits, and validation-based model selection. Where these differ, provide results under a common protocol.

A3. Thank you for pointing this out. The cited numbers follow the GraphSSM benchmark. To address this concern, we reran the baselines under a common protocol as the reviewer suggested (the same preprocessing, the same saved node splits as WaveGSSM, and validation-only model selection). Some examples are shown below (Micro-F1 %):

BaselineDBLP-3BrainRedditDBLP-10
DeepWalk47.09 ± 2.0552.80 ± 2.2426.60 ± 0.4664.33 ± 0.63
Node2Vec47.98 ± 1.2951.32 ± 1.8525.13 ± 1.5067.98 ± 0.87
STAR85.15 ± 1.0989.48 ± 2.4942.10 ± 1.9371.10 ± 0.58

The reruns are consistent with Table 1, and WaveGSSM remains better than these baselines on all four datasets.

AuthorsOpenReview ↗

Response to Reviewer 26hT (Part 2/3)

W2 & Q4. The first contribution and Observation 1 establish that identical patterns can lead to different futures when their rates of change differ. This is a basic property of second-order dynamical systems. Frame it as motivation for the design and avoid presenting the fact itself as a major contribution.

A4. We agree with this suggestion. Our intention was to use Observation 1 to motivate the coupled pattern–motion state in WaveGSSM. We have revised the first contribution in Section 1 (page 3) to present it as a basic second-order property that motivates our design.


Q5. The benefit of an explicit rate-of-change state over standard temporal memory needs further validation. For example, give GCN-LSTM or GCN-SSM explicit differences between consecutive frames while maintaining comparable parameter budgets. This would help separate the benefits of difference cues from those of the second-order state structure.

A5. This is a useful control. We added explicit consecutive-frame differences to GCN-LSTM and GCN-SSM on Reddit, with comparable parameter budgets.

ModelParamsMicro-F1 (%) ↑
GCN-LSTM2,490,45443.29 ± 0.85
GCN-LSTM + differences2,490,45446.75 ± 1.30
GCN-SSM2,484,62843.94 ± 1.19
GCN-SSM + differences2,484,62846.83 ± 1.14
WaveGSSM w/o carry2,486,65045.40 ± 0.59
Our WaveGSSM2,486,65047.47 ± 1.68

Difference cues clearly help both baselines, and WaveGSSM still has the highest mean. The last two rows separate the two effects. Both start from the same difference-based motion state V0\mathbf{V}_0, and only the full model carries the motion state across steps. Removing this second-order structure lowers Micro-F1 from 47.47 to 45.40, so the gain within WaveGSSM comes from the carried motion state rather than from the difference cue itself. We have added these results to Appendix E.3 (page 25) of the revised manuscript.


W2 & Q6. Specify the differences from GraphCON and Graph Wave Networks in terms of state definitions and update equations. Distinguishing "representation propagation" from "motion-state modeling" describes different uses of the dynamics but does not explain why the latter is more advantageous.

A6. The key difference is what the second-order state represents. In GraphCON [R1] and GWN [R2], it is a layer-wise device for message passing on a static graph. In WaveGSSM, it is the motion of the observed field over time, which the graph directly updates:

GraphCON (nn = layer index, static graph): Yn=Yn−1+Δt[σ(Fθ(Xn−1))−γXn−1−αYn−1],Xn=Xn−1+ΔtYn\mathbf{Y}^{n}=\mathbf{Y}^{n-1}+\Delta t\big[\sigma(\mathcal{F}_\theta (\mathbf{X}^{n-1}))-\gamma \mathbf{X}^{n-1}-\alpha \mathbf{Y}^{n-1}\big],\qquad \mathbf{X}^{n}=\mathbf{X}^{n-1}+\Delta t\mathbf{Y}^{n}

GWN-sym (nn = layer index, static graph, Aˉ\bar{\mathbf{A}} = symmetrically normalized adjacency): Xn+1=(2I+Δt2Aˉ)Xn−Xn−1\mathbf{X}^{n+1}=\big(2\mathbf{I}+\Delta t^2\bar{\mathbf{A}}\big)\mathbf{X}^{n}-\mathbf{X}^{n-1}

WaveGSSM (tt = observed time step): Vt=[Vt−1+Δt(−LtUt−1K+Ft)](Id+ΔtΓ)−1,Ut=Ut−1+ΔtVt\mathbf{V}_t=\big[\mathbf{V}_{t-1}+\Delta t(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\big](\mathbf{I}_d+\Delta t\mathbf{\Gamma})^{-1},\qquad \mathbf{U}_t=\mathbf{U}_{t-1}+\Delta t\mathbf{V}_t

In GraphCON and GWN, the rate state is the change across layers (Yn\mathbf{Y}^n, or implicitly Xn−Xn−1\mathbf{X}^n-\mathbf{X}^{n-1}), and the input only sets the initial values. In WaveGSSM, Vt\mathbf{V}_t is the rate of change between snapshots, and every snapshot enters through Ft\mathbf{F}_t.

This matters for prediction because a forecast depends on how the field is moving at the last snapshot (Eq. 8). Applied to temporal graphs, GraphCON and GWN restart their second-order state at each snapshot and leave motion to a separate temporal module. WaveGSSM instead carries the motion across snapshots and updates it through the graph, so that spatial propagation and temporal evolution advance in one transition and patterns travel along edges over time. In line with this, WaveGSSM achieves higher Micro-F1 than GraphCON+LSTM and GWN-sym+LSTM on all four benchmarks (Table 9 in the revised manuscript, and A1 in our response to Reviewer B7C8). We have added the update equations, state definitions, and this comparison to Appendix E.2 (pages 24–25) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 26hT (Part 3/3)

Q7. Appendix B.4 states that each weather prediction call reconstructs (U,V)(\mathbf{U},\mathbf{V}) from two historical frames without directly carrying the states across forecast days. Explain why reconstruction was chosen over continuously updating the latent states, and clarify that motion-state carry occurs only within each processor call.

A7. Thank you for the careful reading of Appendix B.4. Yes, in weather forecasting, motion-state carry occurs only within each processor call, and we have stated this explicitly in Section 4.1 (page 7) of the revised manuscript.

We rebuild the wave state from the two most recent fields for two reasons. First, no observations are available after initialization, so a latent state carried over ten days could drift from the forecast fields without correction. Rebuilding keeps it anchored to them. Second, every call then has the same form as in training and the same two-frame interface as GraphCast* [R5], so the backbone-matched comparison isolates the processor. Motion information still passes across days, since V0\mathbf{V}_0 is rebuilt from the latest two fields.


Q8. The node classification ablation retains the initial V0\mathbf{V}_0, whereas the weather ablation also zeros V0\mathbf{V}_0 before the first transition. Add a weather ablation that retains V0\mathbf{V}_0 and removes only subsequent state carry. Retrain it under the same conditions to isolate the contribution of state carry.

A8. We are grateful for this suggestion. We have implemented the requested weather ablation, which retains V0\mathbf{V}_0 and removes only the subsequent carry. Since full ERA5 training takes a long time, we are retraining it on 16 NVIDIA H100 GPUs, so that we can report the results within the discussion period. We will post them here once training and evaluation are complete. In the meantime, the node-classification ablation follows exactly this design, and removing carry there lowers Micro-F1 and Macro-F1 by 1.81%–3.36% (Table 2).


Q9. Equation (1) assigns learnable weights to support-graph edges, while Appendix C uses disjoint training, validation, and test graphs. Specify how edge weights are generated or transferred to unseen graphs. If the synthetic experiments use a different parameterization, explain how it differs from the main model.

A9. The edge strengths wˉ(ℓ)\bar{\mathbf w}^{(\ell)} in Eq. (1) are used only in the real-data experiments (node classification and ERA5), where the support graph G⋆G_\star is fixed, so no transfer is needed. The synthetic diagnostic in Appendix C uses a simplified WaveGSSM that replaces them with each graph's normalized adjacency A^G\widehat{\mathbf A}_G and shares all parameters across graphs, so it applies directly to unseen graphs. All compared models use the same A^G\widehat{\mathbf A}_G.


Q10. Report parameter counts for the complete models, including learnable edge weights. Measure inference time for an ERA5 one-day forecast on the same GPU model with a batch size of 1, and state whether it includes grid-to-mesh encoding, mesh processing, and mesh-to-grid decoding.

A10. We are happy to report these numbers. All models were timed on the same NVIDIA H100 GPU with batch size 1, covering grid-to-mesh encoding, mesh processing, and mesh-to-grid decoding. Parameter counts include all learnable edge weights.

ModelParams (M)One-day inference (ms)Day-1 geopotential RMSE (m) ↓
GraphCast*0.7207.9320.58
GNN-LSTM0.85215.7220.18
GraphSSM1.30214.1427.18
WaveGSSM1.82028.7312.71

WaveGSSM reduces day-1 RMSE by 38% relative to GraphCast* at under 30 ms per forecast day. Model size does not explain this gain: GraphSSM has more parameters but a higher error, the no-carry ablation with identical parameters is worse (Table 14), and WaveGSSM also performs best in our parameter-matched comparisons on synthetic and Reddit data (A1 and A5 above, and our response to Reviewer 2rMi, Q2). We have added these results to Appendix D.2 (pages 23–24) of the revised manuscript.


We thank the reviewer again for the detailed feedback. We hope these responses address the concerns, and we would be glad to discuss any remaining points.

References

[R1] Rusch et al. Graph-coupled oscillator networks. ICML 2022.

[R2] Yue et al. Graph wave networks. WWW 2025.

[R5] Lam et al. Learning skillful medium-range global weather forecasting. Science 2023.

[R6] Einizade et al. Continuous product graph neural networks. NeurIPS 2024.

[R7] Pareja et al. EvolveGCN: Evolving graph convolutional networks for dynamic graphs. AAAI 2020.

[R8] Li et al. State space models on temporal graphs: A first-principles study. NeurIPS 2024.

Reviewer_B7C8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

While traditional spatiotemporal graph models excel at extracting information from spatial structures and historical sequences, this paper argues that for propagating spatial patterns, explicitly combining the current pattern with its trend of change can be more advantageous for predicting future evolution. This paper proposes WaveGSSM, a graph wave state-space model that explicitly models both the current spatial pattern and its rate of change to predict future evolution. The two states are coupled through graph-based second-order dynamics, allowing the model to capture the propagation of spatiotemporal patterns. Theoretical analysis supports the proposed formulation, while experiments show that explicitly maintaining motion state can improve multi-step prediction, particularly in aligning large-scale weather patterns.

Critical strengths
  1. The paper provides a clear motivation and a rigorous derivation of the proposed model. The authors formulate the coupled pattern and motion states using second-order graph dynamics and derive the corresponding discrete update rules, making the method well motivated and mathematically clear.
  2. The method is specifically designed for weather forecasting, where spatial patterns and their temporal evolution are both important. By explicitly modeling the motion state and its interaction with the pattern state, the model aims to better capture the propagation of spatiotemporal patterns and improve multi-step prediction. The experiments on both temporal graph datasets and global weather forecasting provide encouraging evidence for its effectiveness.
Critical weaknesses
  1. The novelty of the proposed method is not sufficiently clear. Combining a state with its rate of change is a common idea in dynamical systems, second-order neural networks, and physics-inspired models. Although the paper applies this idea to spatiotemporal graph prediction, it does not clearly explain what fundamentally distinguishes the proposed approach from existing methods.
  2. The paper lacks direct comparisons with existing graph-wave models, making it difficult to assess its advantages over closely related work. Without these comparisons, the paper's methodological contribution remains insufficiently justified.
Questions for authors
  1. How does the proposed method differ from existing second-order graph-wave models? Please include direct comparisons with relevant baselines to demonstrate whether the proposed approach offers substantial advantages over existing methods.
  2. Does the explicit motion state truly outperform standard temporal modeling? Please provide more detailed ablation studies comparing the proposed motion state with simpler alternatives, such as using historical differences or conventional temporal state updates, to isolate its contribution to the performance gains.
2 comments
AuthorsOpenReview ↗

Response to Reviewer B7C8

We sincerely thank the reviewer for recognizing the motivation and derivation of our model, and for the constructive suggestions. We address each point below.


W1–W2 & Q1. How does the proposed method differ from existing second-order graph-wave models? Please include direct comparisons with relevant baselines to demonstrate whether the proposed approach offers substantial advantages over existing methods.

A1. Second-order dynamics are indeed well studied, and we do not claim the idea itself. Our contribution is to model the motion of the observed field as a state that is updated through the graph. In a first-order update U˙=−LUK+F\dot{\mathbf{U}}=-\mathbf{L}\mathbf{U}\mathbf{K}+\mathbf{F}, the symmetric graph term can only smooth a pattern. In WaveGSSM, the graph force changes the motion (V˙=−LUK−VΓ+F\dot{\mathbf{V}}=-\mathbf{L}\mathbf{U}\mathbf{K}-\mathbf{V}\mathbf{\Gamma}+\mathbf{F}, U˙=V\dot{\mathbf{U}}=\mathbf{V}), giving oscillatory graph modes (Eq. 8), so patterns can travel along edges. Second-order graph models such as GraphCON [R1] and Graph Wave Networks (GWN) [R2] run the dynamics over network depth on a static graph. Second-order sequence models such as coRNN [R3] and LinOSS [R4] run them over time, but without graph coupling. WaveGSSM runs them over observed time, with graph coupling inside the transition.

The table summarizes the difference from related second-order models:

GraphCONGWNcoRNNLinOSSOur WaveGSSM
Second-order state✓✓✓✓✓
Graph coupling in the transition✓✓✗✗✓
Evolves over observed time✗✗✓✓✓
Evolving graph across snapshots✗✗✗✗✓
Propagation along edges across snapshots✗✗✗✗✓

WaveGSSM is thus a state-space model over the observed sequence, not a second-order GNN layer.

GraphCON and GWN are static-graph models and are not designed for temporal graphs. coRNN and LinOSS are designed for sequences without graph structure, so we do not include them as graph baselines. For a direct comparison, we paired GraphCON and GWN each with an LSTM over snapshots and evaluated them on the four temporal-graph benchmarks with the same splits and protocol (Micro-F1, mean ± std):

DatasetGraphCON+LSTMGWN-sym+LSTMOur WaveGSSM
DBLP-386.20 ± 2.5386.10 ± 2.2386.34 ± 1.54
Brain91.32 ± 1.0691.44 ± 0.4393.36 ± 0.66
Reddit41.01 ± 1.6545.73 ± 1.3647.47 ± 1.68
DBLP-1072.96 ± 0.8573.35 ± 0.3976.38 ± 0.40

WaveGSSM has the highest mean on all four datasets and is clearly better on Brain, Reddit, and DBLP-10. We have added these baselines and the comparison to Appendix E.2 (pages 24–25) of the revised manuscript, with a summary in Section 4.2 (page 7), and discussed coRNN and LinOSS in Section 2 (page 3).


Q2. Does the explicit motion state truly outperform standard temporal modeling? Please provide more detailed ablation studies comparing the proposed motion state with simpler alternatives, such as using historical differences or conventional temporal state updates, to isolate its contribution to the performance gains.

A2. We address the two alternatives you mention in turn. For historical differences, we retrained a no-carry variant of WaveGSSM. It is identical to the full model, including the parameter count, except that the motion state is not passed between steps. Both variants start from the same difference-based V0\mathbf{V}_0 and keep the instantaneous graph response and terminal readout.

DatasetParams (each)Without motion carryOur WaveGSSM
DBLP-3376,55384.51 ± 1.4186.34 ± 1.54
Reddit2,486,65045.40 ± 0.5947.47 ± 1.68

Micro-F1 (%), mean ± SD (Table 2).

With the same difference cue, carrying the motion state gives higher Micro-F1 on both datasets, so the gain goes beyond supplying historical differences to a first-order graph-coupled update. For conventional temporal updates, our response to Reviewer 2rMi (Q2) compares LSTM and GraphSSM [R8] within a shared backbone on the Figure 1 task, where WaveGSSM has the lowest MSE. The no-carry design is described in Appendix E.1 (page 24), and the shared-backbone comparison has been added to Appendix C (page 20).


We hope these clarifications and additional comparisons address the reviewer's questions, and we are happy to provide further details if needed.

References

[R1] Rusch et al. Graph-coupled oscillator networks. ICML 2022.

[R2] Yue et al. Graph wave networks. WWW 2025.

[R3] Rusch et al. Coupled oscillatory recurrent neural network (coRNN): An accurate and (gradient) stable architecture for learning long time dependencies. ICLR 2021.

[R4] Rusch et al. Oscillatory state-space models. ICLR 2025.

[R8] Li et al. State space models on temporal graphs: A first-principles study. NeurIPS 2024.

AuthorsOpenReview ↗

Summary of Our Responses and a Kind Request for Feedback

Thank you again for your review, for your positive rating, and for your kind words on the motivation and the derivation. We truly appreciate your positive support. We summarize how we addressed your two questions.

  • Difference from existing second-order graph models (W1–W2, Q1). GraphCON and GWN run second-order dynamics across network layers on a static graph. WaveGSSM makes the motion of the observed field a state that evolves over time and is updated through the graph. A first-order graph update can only smooth a pattern, while this transition lets the pattern travel along edges across snapshots. WaveGSSM outperforms GraphCON and GWN, each paired with an LSTM, on all four temporal-graph benchmarks.
  • Motion state vs. simpler alternatives (Q2). With the same difference-based V0\mathbf{V}_0 and parameter count, removing the carry lowers Micro-F1 on DBLP-3 and Reddit. Explicit frame differences help GCN-LSTM and GCN-SSM, but both stay below WaveGSSM. Under a shared backbone, WaveGSSM also has lower MSE than LSTM and GraphSSM.

We would really value your view on whether these responses address your concerns. If you feel they do, we would be grateful if this could be reflected in your final assessment. We are also happy to clarify anything further.

Reviewer_2rMiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WaveGSSM models spatio-temporal graphs using coupled pattern and motion states. It is evaluated on four temporal node-classification benchmarks and global weather forecasting, with improvements over the reported baselines.

Critical strengths

The method has a clear motivation and an interpretable architecture. Experiments cover multiple tasks, and the retrained motion-carry ablation provides useful evidence of its effectiveness.

Critical weaknesses

My concerns are mainly about the results and interpretations - see the questions below.

Questions for authors
  • Removing motion carry increases weather RMSE from 30.72 to 31.53 m, whereas the full model reports a 20.2% average improvement over GraphCast*. Although these percentages are not directly comparable, the modest ablation gap raises questions about the source of the overall gains. Could the authors clarify how much improvement comes from motion carry versus other design choices?

  • Comparisons with LSTM and GraphSSM involve multiple architectural differences. Could the authors isolate the proposed motion update through a controlled comparison using the same inputs, spatial processing, decoder, and comparable model budgets?

  • GraphCON already uses coupled states and second-order graph dynamics. Beyond applying these dynamics along observed time, what are the essential methodological advances of WaveGSSM?

  • Could persistent motion memory hurt predictions when abrupt events invalidate past trends? Please clarify how the model handles such changes and provide supporting evidence, if available.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2rMi (Part 1/2)

We sincerely thank the reviewer for the time and constructive comments, which have helped us improve the paper. We address each point below.


Q1. Removing motion carry increases weather RMSE from 30.72 to 31.53 m, whereas the full model reports a 20.2% average improvement over GraphCast*. Although these percentages are not directly comparable, the modest ablation gap raises questions about the source of the overall gains. Could the authors clarify how much improvement comes from motion carry versus other design choices?

A1. Thank you for raising this point about the source of the gains. On a common scale (mean geopotential RMSE over six levels and days 1–5), GraphCast* [R5] obtains 37.21 m, WaveGSSM without motion-state carry 31.53 m, and the full WaveGSSM 30.72 m.

With h=Δth=\Delta t and D=(Id+hΓ)−1\mathbf{D}=(\mathbf{I}_d+h\mathbf{\Gamma})^{-1}, eliminating V\mathbf{V} gives, for t≥2t\ge 2,

Full:Ut=Ut−1+(Ut−1−Ut−2)D+h2(−LtUt−1K+Ft)D,\text{Full:}\quad \mathbf{U}_t=\mathbf{U}_{t-1}+(\mathbf{U}_{t-1}-\mathbf{U}_{t-2})\mathbf{D}+h^2(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\mathbf{D},

No motion-state carry:Ut=Ut−1+h2(−LtUt−1K+Ft)D.\text{No motion-state carry:}\quad \mathbf{U}_t=\mathbf{U}_{t-1}+h^2(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\mathbf{D}.

The ablation removes only the momentum term (Ut−1−Ut−2)D(\mathbf{U}_{t-1}-\mathbf{U}_{t-2})\mathbf{D}. The graph-wave term h2(−LtUt−1K+Ft)Dh^2(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\mathbf{D} stays the same. The ablated model thus still has the graph-wave transition itself, which accounts for the 5.68 m gain over GraphCast*. Carry adds a consistent further 0.81 m, with lower z200 error at every reported lead.

The same ablation also lowers Micro-F1 and Macro-F1 by 1.81%–3.36% on three temporal-graph benchmarks. In the frozen ERA5 model, suppressing the motion pathway raises the standardized error by 51.1%.

We have made this decomposition explicit in Section 4.4 (page 9) and Appendix F.2 (pages 26–27) of the revised manuscript.


Q2. Comparisons with LSTM and GraphSSM involve multiple architectural differences. Could the authors isolate the proposed motion update through a controlled comparison using the same inputs, spatial processing, decoder, and comparable model budgets?

A2. We thank the reviewer for raising this concern. We ran the suggested controlled comparison on the Figure 1 data. All models use the same inputs, encoder, spatial processing, and decoder, with the same number of training updates and similar parameter counts. Only the temporal update differs. WaveGSSM w/o carry keeps the initial motion state V0\mathbf{V}_0 and the readout, but does not pass the motion state on to later steps.

Temporal updateParamsMSE ×1000 ↓Reduction by WaveGSSM
LSTM31,8671.0494 ± 0.00644.72%
GraphSSM [R8]31,4051.2880 ± 0.063122.37%
WaveGSSM w/o carry32,3891.0682 ± 0.11786.39%
WaveGSSM32,3890.9999 ± 0.0126—

WaveGSSM has the lowest mean MSE under the shared backbone, so the gain is not explained by differences in spatial processing or decoding. The no-carry variant has the same parameter count, inputs, and initial motion cue as the full model, and carrying the motion state across steps lowers MSE by 6.39%. We have added this comparison to Appendix C (page 20) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 2rMi (Part 2/2)

Q3. GraphCON already uses coupled states and second-order graph dynamics. Beyond applying these dynamics along observed time, what are the essential methodological advances of WaveGSSM?

A3. Thank you for this question on the relation to GraphCON [R1]. Although both start from a second-order ODE, WaveGSSM and GraphCON are fundamentally different models:

GraphCON (nn = layer index, static graph): Yn=Yn−1+Δt[σ(Fθ(Xn−1))−γXn−1−αYn−1],Xn=Xn−1+ΔtYn\mathbf{Y}^{n}=\mathbf{Y}^{n-1}+\Delta t\big[\sigma(\mathcal{F}_\theta (\mathbf{X}^{n-1}))-\gamma \mathbf{X}^{n-1}-\alpha \mathbf{Y}^{n-1}\big],\qquad \mathbf{X}^{n}=\mathbf{X}^{n-1}+\Delta t\mathbf{Y}^{n}

WaveGSSM (tt = observed time step): Vt=[Vt−1+Δt(−LtUt−1K+Ft)](Id+ΔtΓ)−1,Ut=Ut−1+ΔtVt\mathbf{V}_t=\big[\mathbf{V}_{t-1}+\Delta t(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\big](\mathbf{I}_d+\Delta t\mathbf{\Gamma})^{-1},\qquad \mathbf{U}_t=\mathbf{U}_{t-1}+\Delta t\mathbf{V}_t

  1. Different state. In GraphCON, Y\mathbf{Y} is the change of features across layers. In WaveGSSM, Vt\mathbf{V}_t is the rate of change of the observed field across snapshots, i.e., the motion of the pattern.
  2. Different role of data. GraphCON is autonomous, and the input only sets X0\mathbf{X}^0. WaveGSSM is a data-driven state-space model, in which every snapshot enters through Ft\mathbf{F}_t.
  3. Different operator. GraphCON applies a GNN layer on a fixed graph and needs a separate temporal module. In WaveGSSM, LtUt−1K\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K} with a time-varying Laplacian Lt\mathbf{L}_t advances spatial propagation and temporal evolution in one operator.

These are differences in model class, not only in the time index. We also paired GraphCON with an LSTM over snapshots and evaluated it under the same splits and protocol:

DatasetGraphCON+LSTMWaveGSSMRelative improvement
DBLP-386.20 ± 2.5386.34 ± 1.54+0.16%
Brain91.32 ± 1.0693.36 ± 0.66+2.23%
Reddit41.01 ± 1.6547.47 ± 1.68+15.75%
DBLP-1072.96 ± 0.8576.38 ± 0.40+4.69%

WaveGSSM achieves higher mean Micro-F1 on all four datasets, with clear gains on Brain, Reddit, and DBLP-10. We have added this comparison and the side-by-side formulation to Appendix E.2 (pages 24–25) of the revised manuscript and summarized the result in Section 4.2 (page 7).


Q4. Could persistent motion memory hurt predictions when abrupt events invalidate past trends? Please clarify how the model handles such changes and provide supporting evidence, if available.

A4. Thank you for raising this concern about abrupt changes. In our model, past motion is not kept fixed. In the update

Vt=[Vt−1+Δt(−LtUt−1K+Ft)](Id+ΔtΓ)−1,\mathbf{V}_t=\big[\mathbf{V}_{t-1}+\Delta t(-\mathbf{L}_t\mathbf{U}_{t-1}\mathbf{K}+\mathbf{F}_t)\big](\mathbf{I}_d+\Delta t\mathbf{\Gamma})^{-1},

the current snapshot enters through Ft\mathbf{F}_t before Vt\mathbf{V}_t is used to update Ut\mathbf{U}_t, so a sudden change in the input changes Vt\mathbf{V}_t in the same step. The damping term (Id+ΔtΓ)−1(\mathbf{I}_d+\Delta t\mathbf{\Gamma})^{-1} is learned and reduces the old motion at every step, so an old trend decays if new inputs do not support it.

We also see this in the experiments. Among our datasets, Brain and Reddit change the most between snapshots (TCs\mathrm{TC}_{\mathrm{s}} = 0.024 and 0.030, compared with 0.139 for DBLP-3, Table 5). If carrying motion were harmful under such changes, removing it should help on these two datasets. Instead, F1 drops by 2.07%–3.36% on Brain and Reddit, compared with 1.81%–1.83% on DBLP-3. We have added this observation to Appendix E.1 (page 24) of the revised manuscript.


We hope these responses address the reviewer's concerns, and we would be glad to clarify any remaining questions.


References

[R1] Rusch et al. Graph-coupled oscillator networks. ICML 2022.

[R5] Lam et al. Learning skillful medium-range global weather forecasting. Science 2023.

[R8] Li et al. State space models on temporal graphs: A first-principles study. NeurIPS 2024.

AuthorsOpenReview ↗

Summary of Our Responses and a Kind Request for Feedback

Thank you again for your review. Your questions helped us make the source of our gains much clearer, and we summarize the main points below.

  • Where the gains come from (Q1). On ERA5, the graph-wave transition accounts for 5.68 m of the 6.49 m gain over GraphCast*, and carrying the motion state adds 0.81 m. The new ablation that keeps V0\mathbf{V}_0 and removes only the carry shows that most of this 0.81 m comes from the carry itself, with WaveGSSM better at all 60 level–lead pairs (posted under Reviewer 26hT, Q8).
  • Controlled comparison (Q2). When only the temporal update changes, WaveGSSM has lower MSE than LSTM, GraphSSM, and the no-carry variant.
  • What is new relative to GraphCON (Q3). GraphCON runs second-order dynamics across layers on a static graph. WaveGSSM makes the motion of the observed field a state that evolves over time and is updated through the graph, so patterns can travel along edges between snapshots. WaveGSSM also outperforms GraphCON paired with an LSTM on all four benchmarks.
  • Abrupt changes (Q4). New inputs update the motion state in the same step, and the learned damping lets old trends decay. Removing the carry hurts most on Brain and Reddit, which change the most between snapshots.

We were glad that you found the architecture interpretable. We would really value your view on whether these responses address your concerns. If they do, we would be grateful if you could take them into account in your final assessment, and we are happy to clarify anything further.

AuthorsOpenReview ↗

A Gentle Follow-up

Thank you again for taking the time to review our paper and for your constructive comments. As the discussion is coming to a close, we would greatly appreciate hearing whether our responses address your concerns. Please let us know if you have any other questions.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and the AC

We thank all reviewers for their careful and constructive feedback. We are encouraged that the reviewers found the motivation clear (Reviewers 2rMi, B7C8, 26hT), the derivation rigorous (Reviewer B7C8), the architecture interpretable (Reviewer 2rMi), and the paper well written with clear figures (Reviewer 26hT). Below we summarize how we addressed the three concerns shared across the reviews.


1. Novelty relative to existing second-order models (2rMi, B7C8, 26hT)

Second-order dynamics are well studied, and we do not claim the idea itself. The contribution of WaveGSSM is to let the graph act on the motion of the observed field over time. In a first-order update U˙=−LUK+F\dot{\mathbf{U}}=-\mathbf{L}\mathbf{U}\mathbf{K}+\mathbf{F}, the symmetric graph term can only smooth a pattern. In WaveGSSM, the graph force changes the motion (V˙=−LUK−VΓ+F\dot{\mathbf{V}}=-\mathbf{L}\mathbf{U}\mathbf{K}-\mathbf{V}\mathbf{\Gamma}+\mathbf{F}, U˙=V\dot{\mathbf{U}}=\mathbf{V}), giving oscillatory graph modes (Eq. 8), so patterns travel along edges across snapshots. The table compares WaveGSSM with second-order graph models, GraphCON [R1] and GWN [R2], and second-order sequence models, coRNN [R3] and LinOSS [R4]:

GraphCONGWNcoRNNLinOSSOur WaveGSSM
Second-order state✓✓✓✓✓
Graph coupling in the transition✓✓✗✗✓
Evolves over observed time✗✗✓✓✓
Evolving graph across snapshots✗✗✗✗✓
Propagation along edges across snapshots✗✗✗✗✓

GraphCON and GWN run the dynamics over network depth on a static graph, with the input entering only as the initial state. In WaveGSSM, every snapshot enters through the forcing Ft\mathbf{F}_t, so spatial propagation and temporal evolution advance in one transition. Under the same protocol, WaveGSSM achieves higher mean Micro-F1 than GraphCON+LSTM and GWN-sym+LSTM on all four temporal-graph benchmarks (B7C8 A1, 26hT A6).


2. Separating the motion state from other sources of improvement (2rMi, B7C8, 26hT)

  • Decomposition on ERA5. Mean geopotential RMSE over days 1–5 (lower is better) drops from 37.21 m for GraphCast* [R5] to 31.53 m for WaveGSSM without motion-state carry, and to 30.72 m for the full model. The graph-wave transition thus accounts for most of the gain (5.68 m), and carrying the motion state adds 0.81 m, with lower z200 error at every reported lead (2rMi A1).
  • Same backbone, only the temporal update changed. WaveGSSM has lower MSE than LSTM (by 4.72%), GraphSSM (by 22.37%), and the no-carry variant (by 6.39%) (2rMi A2).
  • Difference cues vs. second-order structure. Frame differences help GCN-LSTM and GCN-SSM, but WaveGSSM remains best. With the same difference-based V0\mathbf{V}_0, removing carry lowers Micro-F1 from 47.47 to 45.40 on Reddit, and by 1.81%–3.36% on three benchmarks (26hT A5, B7C8 A2).
  • Model size. Parameter count does not explain the gains: GraphSSM has more parameters than GraphCast* but a higher error, and the no-carry ablation with identical parameters is worse (26hT A10).
  • Additional ablation. Retaining V0\mathbf{V}_0 and removing only later carry gives 31.31 m vs. 30.72 m for WaveGSSM, which is better at all 60 level–lead pairs (26hT Q8).

3. Additional baselines and evaluation protocol (B7C8, 26hT)

  • CITRUS [R6] on ERA5 under the same data setup: WaveGSSM lowers mean RMSE over days 1–5 by 34.03% (26hT A2).
  • EvolveGCN-H [R7] and GraphSSM [R8], official implementations: on the Figure 1 task, WaveGSSM is best on both metrics, with about 30% lower localization error than GraphSSM (26hT A1).
  • Common protocol. Baselines rerun with the same preprocessing, node splits, and validation-based selection are consistent with Table 1 (26hT A3).
  • Cost. A one-day ERA5 forecast takes under 30 ms on one H100 GPU (26hT A10).

Revisions. We have uploaded a revised manuscript, and all page numbers in our responses refer to it. The main text now presents Observation 1 as motivation and clarifies the weather setting and the source of the gains (Sections 1, 2, 4.1, and 4.4). The new results are in Appendices C, D.2, E.2, E.3, and F.1.

We thank the reviewers and the AC for their time and effort.


References

[R1] Rusch et al. Graph-coupled oscillator networks. ICML 2022.

[R2] Yue et al. Graph wave networks. WWW 2025.

[R3] Rusch et al. Coupled oscillatory recurrent neural network (coRNN): An accurate and (gradient) stable architecture for learning long time dependencies. ICLR 2021.

[R4] Rusch et al. Oscillatory state-space models. ICLR 2025.

[R5] Lam et al. Learning skillful medium-range global weather forecasting. Science 2023.

[R6] Einizade et al. Continuous product graph neural networks. NeurIPS 2024.

[R7] Pareja et al. EvolveGCN: Evolving graph convolutional networks for dynamic graphs. AAAI 2020.

[R8] Li et al. State space models on temporal graphs: A first-principles study. NeurIPS 2024.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.