acceptodds
Under review as a conference paper at ICLR 2027

SLiDNet: Scale-Separated Liquid Dynamics for Spatio-Temporal Forecasting

Abstract

Spatio-temporal forecasting requires modeling both the state of each node and the rate at which that state evolves. However, existing models often rely on shared latent representations for state formation and temporal evolution. This coupling may mix signal-level differences with temporal variation and make it harder to distinguish a node's own evolution from the influence of its neighbors. To address this issue, we propose the Scale-Separated Liquid Dynamics Network (SLiDNet), which separates state formation from evolution-rate modeling. SLiDNet forms node states from scale-stable temporal content and compact window-level signal context. Robustly normalized first- and second-order temporal differences, together with recurrent context, condition node-specific Liquid time constants to adapt the evolution rate. Spatial dependencies are captured through prior-guided graph diffusion. Experiments across six real-world spatio-temporal datasets and multiple forecasting horizons show consistent improvements over competitive baselines. Controlled ablations and sensitivity analyses further support the effectiveness of the proposed scale-separated design.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Liquid Dynamics Networks, Scale-Separated Representation, Adaptive Time Constants, Traffic Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_s9MvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether spatio-temporal forecasting should separate a node's signal level from its temporal variation. SLiDNet uses level information to form the recurrent candidate and variation information to control a bounded Liquid time constant. Experiments on six traffic benchmarks compare it with seven baselines and include ablations and efficiency analysis.

Critical strengths

The method is clearly specified and evaluated across six standard traffic benchmarks against several baseline families.

Critical weaknesses

Regarding motivation, Figure 1 shows that signal level and temporal variation are not identical, but it does not establish that modelling them with a shared representation is harmful or that the proposed separation is necessary. The ablations did not compare separated with shared conditioning or remove the level-context path.

Regarding experiments, Table 1 does not include some commonly-used baselines, e.g., DCRNN, or RLSTG [a], the nearest Liquid graph model. The six datasets are all medium size datastes, missing large datasets like LargeST [b] for scalability. In addition, the T=HT=H baseline settings are insufficiently described, and no standard deviation across repeated runs is reported.

Questions for authors
  1. How does adding the window's last value and log-scale to the candidate state differ from re-injecting statistics removed by normalisation through learned de-stationary factors, as in [c]?
  2. Were hyperparameters selected and the sensitivity analysis in Figure 4 conducted on validation or test data?

References

[a] Riemannian Liquid Spatio-Temporal Graph Network. WWW 2026.

[b] LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting. NeurIPS 2023.

[c] Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting. NeurIPS 2022.

3 comments

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than 1 days remaining, I wanted to ensure we have addressed all your concerns satisfactorily.

If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

Thank you for your time and effort in reviewing our paper.

Thank you for the effort during the rebuttal and the additional experiments. My main concern remains the central claim: the new ablation shows that the separated and shared designs perform about the same, so the results do not yet show that separating signal level from temporal variation is what drives the gains. I will therefore keep my original score.

Dear Reviewer,

Thank you for your further consideration. We respectfully acknowledge that the shared and separated variants exhibit comparable forecasting accuracy, but we believe this result should be interpreted in the context of the intended contribution of SLiDNet.

Our motivation is not to establish that separated conditioning universally outperforms shared conditioning in terms of prediction error, but to introduce a principled, role-specific inductive bias that assigns signal-level information to candidate-state formation and scale-normalised temporal variation to evolution-rate control. Importantly, the shared variant introduces additional cross-conditioning paths and parameters without demonstrating consistent performance gains, whereas the proposed design preserves this functional distinction with a simpler architecture. The complementary ablation further shows that level context benefits METR-LA and PEMS08, although its contribution is dataset-dependent. We therefore agree that the current results do not establish separation as the sole source of SLiDNet's forecasting gains, and have refined our claims accordingly. Nevertheless, we believe the explicit separation of state formation and evolution-rate control constitutes a meaningful architectural contribution, supported by the mechanistic motivation and comprehensive empirical evaluation, rather than merely an accuracy-driven design choice.

We respectfully hope you will reconsider the contribution in this broader context.

Reviewer_pMBPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SLiDNet is a spatio-temporal forecaster that separates what state a node forms from how fast that state evolves. Inside a Liquid relaxation update, the target state comes from normalized content plus explicit signal-level context, while node-bounded time constants are driven by scale-normalized temporal variation. Graph diffusion with an adaptive adjacency, anchored to a correlation prior, adds spatial dependencies, and the model leads on six traffic benchmarks.

Critical strengths
  1. The model design is specified in reasonable detail. The equations, the algorithm, and the appendices make the main components and their interactions understandable to the reader, although full reproducibility is limited without released code.
  2. The method shows consistent empirical gains over the compared baselines in most of the evaluated datasets and forecasting horizons, with competitive efficiency.
Critical weaknesses
  1. The "Liquid" part is basically a standard GRU gate. All datasets are sampled every 5 minutes, so the time step never changes, and the update is a learned gate that mixes the old state with a new candidate; the “liquid dynamics” framing needs to be more explained and justified. It is also unclear what an "evolution rate" means when the input window has only 3 or 6 time steps, since the variation statistics are computed from very few points.
  2. The central claim, that scale separation is the source of the gains, is not directly tested. The ablations remove reference shrinkage, node-specific tau ranges, dynamic tau, and rate-controlled interpolation. None of them tests the contribution of the separation part itself.
  3. There is no analysis of the learned time constants. The paper claims τ adapts to how fast each node evolves, but it is not clear that the learned τ values correlate with observed node volatility or change during sudden events.
  4. All baselines use a common 50-epoch budget. The comparison may reflect undertrained baselines, not a better model. Some baselines require more training epochs.
  5. The results reporting is unclear and incomplete. The baseline numbers differ noticeably from known published values. It is unclear whether errors are averaged over the horizon or reported at the final step.
  6. All datasets are traffic datasets despite broader claims. Dataset-specific κ is set without a stated selection procedure.
Questions for authors
  1. Can you report results with each baseline trained under its official schedule, and report mean and std over seeds?
  2. How were the dataset-specific k values selected?
1 comment

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than 1 days remaining, I wanted to ensure we have addressed all your concerns satisfactorily.

If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

Thank you for your time and effort in reviewing our paper.

Reviewer_sWfDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a spatiotemporal prediction model SLiDNet, which separates the construction of candidate states from the control of evolution rate: states are constructed through contraction normalization and signal context, node time constants are adjusted using normalized temporal differences and cyclic states, and spatial dependencies are modeled through prior guided graph diffusion.

Critical strengths

Among the 18 datasets and prediction length configurations, MAE and RMSE achieved 17 and 16 optimal results, respectively, indicating that the benefits are not limited to individual settings.

Critical weaknesses
  1. Figure 2 is full of garbled characters and the format is chaotic.
  2. This article is difficult to read.
Questions for authors

See Weaknesses

1 comment

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than 1 days remaining, I wanted to ensure we have addressed all your concerns satisfactorily.

If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

We hope the above clarifications and the additional experiments in the revised draft sufficiently addressed your concerns. If you are satisfied, we kindly request you to consider updating the score to reflect the newly added results and discussion. We remain committed to addressing any remaining points you may have during the discussion phase.

Thank you for your time and effort in reviewing our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.