acceptodds
Under review as a conference paper at ICLR 2027

PIANO: PATH-INTEGRAL ADAPTIVE NEURAL ODES ON GRAPHS

Abstract

Many natural and engineered spatiotemporal processes evolve through continuous interactions over a network, while the spatial range of relevant interactions can vary with the system state. Existing graph ODEs typically rely on a predetermined hop structure within each vector-field evaluation, preventing the propagation mass assigned to different spatial orders from adapting with the evolving state. We introduce PIANO (Path-Integral Adaptive Neural ODE), whose instantaneous spatial operator is a state-adaptive path-integral propagator: an expected finite-step graph-walk propagator under a state-conditioned distribution over walk lengths. A lightweight topology-aware gate further enables node-wise modulation when structural roles are informative. The adaptive operator is implemented with sparse graph operations whose cost scales linearly with maximum propagation order, edge count, and feature dimension. We show that each realized spatial operator is non-expansive, establish well-posedness of the nonlinear flow, and, for a frozen linear surrogate, characterize how the maximum propagation order and integration horizon jointly determine a sufficient finite-hop approximation radius. Controlled experiments validate the proposed adaptive propagation and topology-aware mechanisms. On ocean and severe-weather forecasting, PIANO achieves strong performance with a compact model and further extends effectively to large-scale sparse graph learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Continuous-Time Graph Learning, Neural ODEs, State-Adaptive Propagation, Multi-Hop Graph Propagation, Graph Dynamics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NixZOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PIANO, a graph neural ODE approach to learn on graph-structured data. For spatial propagation, PIANO considers a learnable convex combination of different-length updates, making it multi-hop and able to go beyond 1-WL tests and MPNN architectures. Node states evolve following a non-linear ODE based on this combination of spatial updates. The flow of the model is well-posed, and the node- and topology-dependent gate is sound for distinguishing topologies around nodes. PIANO is tested on both spatiotemporal and static tasks, with an initial validation of the mechanisms in more synthetic settings.

Critical strengths
  • The paper is well structured and motivated. The theory section is well written and provides useful statements on the motivation and well-posedness of the model, while also highlighting its advantages and how it can distinguish local geometries around nodes.

  • The paper provides enough ablations, technical and implementation details to understand the components of the model and how they work together.

Critical weaknesses

The paper fails to reference or compare with a long list of relevant methods in the field, more specifically, in two categories:

  • Multi-hop and Path GNNs. Since PIANO essentially weights paths and powers of the adjacency matrix, it should be compared to other state-of-the-art models with similar architectures, such as MixHop [1], DReW [2], PathNN [3], and PANConv [4] (needs a deeper comparison, as it is based on Path-integral as well).
  • Differential equation-inspired and Graph ODE models. Similarly, a lot of models are based on latent ODEs and have different propagation guarantees depending on the adopted dynamics, such as PDE-GCN [5], GraphCON [6], ADGN [7], PH-DGN [8], SWAN [9], and SONAR [10].

Furthermore, the authors could compare the models to well-known benchmarks for spatiotemporal propagation (PEMS-Bay and Metr-LA [11], EngRad [12]) or static data (such as LRGB [13]), where the multi-hop and continuous-time propagation dynamics have shown advantages.

Finally, the controlled experiments lack details as of now, especially in the main text. The authors should consider adding more details and interpretation. As of now, it is hard to understand the setting and especially the interpretation of the results.

[1] Abu-el-Haija et al., MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing, ICML 2019.

[2] Gutteridge et al., DRew: Dynamically Rewired Message Passing with Delay, ICML 2023.

[3] Michel et al., Path Neural Networks: Expressive and Accurate Graph Neural Networks, ICML 2023.

[4] Ma et al., "Path Integral Based Convolution and Pooling for Graph Neural Networks", NeurIPS 2020

[4] Eliasof et al., PDE-GCN: Novel Architectures for Graph Neural Networks Motivated by Partial Differential Equations, NeurIPS 2021.

[5] Rusch et al., Graph-Coupled Oscillator Networks, ICML 2022

[6] Gravina et al., Anti-Symmetric DGN: A Stable Architecture for Deep Graph Networks, ICLR 2023.

[7] Heilig et al., Port-Hamiltonian Architectural Bias for Long-Range Propagation in Deep Graph Networks, ICLR 2025.

[8] Gravina et al., On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems, AAAI 2025.,

[9] Trenta et al., SONAR: Long-range propagation through Information waves

[10] Li et al., Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. ICLR 2018

[11] Marisca et al., Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling. ICLR 2024.

[12] Dwivedi et al., Long Range Graph Benchmark, NeurIPS 2022

Questions for authors
  • It seems like the bolding in some tables is wrong or misleading (for example, in Table 2 or Table 3, where PIANO is bold but not the best).
  • Is it possible to extend Theorem 2 to a non-linear model? What are the main differences and advantages?
  • Could the authors compare to other multi-hop and path GNNs, showing advantages and/or main differences?
  • The authors should add more details on the controlled experiments showing the effects of the mechanisms.
11 comments
AuthorsOpenReview ↗

CW1 / Q3 — Related work and architectural differences

Thank you for identifying the missing comparisons and the points needing clarification. We now cite and evaluate all ten requested model families and add EngRAD, METR-LA, PEMS-BAY and Peptides-func comparisons. We also revised the operator comparison, clarified nonlinear theory, and added controlled-experiment explanations. Substantive paper revisions are orange. This response gives the contribution/related work first, forecasting comparisons in Comment 2, direct theory/presentation answers in Comment 3, and LRGB/additional graph results in Comment 4.

Our Central question is: How can a continuous-time graph model redistribute propagation mass across spatial orders as its latent state evolves, while retaining continuous temporal evolution and affordable explicit multi-order propagation?

PIANO makes spatial order an adaptive control variable inside each vector-field evaluation: Mθ(H,t)=∑k=0Lαk(H,t)A^kM_\theta(H,t)=\sum_{k=0}^{L}\alpha_k(H,t)\hat A^k, with α(H,t)\alpha(H,t) on the simplex. Sparse recursion costs O(L∣E∣d)O(L|E|d) for propagation. This combines explicit order control with continuous-time evolution.

Requested multi-hop/path methods. We now cite and compare MixHop (Abu-El-Haija et al., 2019), DReW (Gutteridge et al., 2023), PathNN (Michel et al., 2023) and PAN (Ma et al., 2020). Appendix H explains the distinctions:

ModelSpatial control
MixHopTransformed adjacency powers
DReWDistance-dependent rewiring and delayed messages
PathNNLearned path encodings; shortest-path variant in our comparisons
PANPath-length-weighted convolution and structure-informed pooling
PIANOState-recomputed simplex over explicit walk lengths inside the ODE

Deeper comparison with PAN. PAN learns layer-level path-length weights and uses the operator diagonal for structural pooling. PIANO recomputes its distribution from the evolving state and solver time. Our optional gate shares the closed-walk motivation but modulates node-wise field updates. Our PAN baseline is explicitly a node-preserving convolution-only adaptation, without PANPool; B.3–B.4/H identify this choice.

Requested differential-equation models. Revised Table 1 and H.2 add PDE-GCN (Eliasof et al., 2021), GraphCON (Rusch et al., 2022), A-DGN (Gravina et al., 2023), PH-DGN (Heilig et al., 2025), SWAN (Gravina et al., 2025) and SONAR (Trenta et al., 2025). Their diffusion, oscillator, antisymmetric, Hamiltonian and wave dynamics provide complementary propagation mechanisms. Table 1 now distinguishes explicit spatial reach from temporal derivative order and reach accumulated through integration. It recognizes PH-DGN's possible two-hop energy-gradient reach and GREAD-BS's two-hop reaction.

PIANO's distinction is the state-conditioned redistribution among explicit spatial orders within the continuous vector field.

AuthorsOpenReview ↗

CW2 / Q3 — Weather and traffic comparisons

Main-paper results: transfer across tasks

Tables 4–5 give PIANO the best displayed means on Ocean (0.5245 ± 0.0013 Mean ACC; 43.1K parameters), SEVIR (all three metrics; 69.9K parameters) and PubMed (89.645 ± 0.534%). Ocean uses PIANO's separately selected recipe alongside original baseline recipes; B.4/E.6 disclose their different budgets.

EngRAD: forecasting across weather stations

EngRAD (Marisca et al., 2024) uses 487 locations, five channels, 24 observed/24 predicted hours and chronological 2018/2019/2020 splits, with train-only normalization. Mean ± sample SD over ten seeds:

MethodParamsStd. MAE ↓Std. MSE ↓
MixHop55.9K0.4252 ± 0.00110.5231 ± 0.0027
PAN convolution55.9K0.4268 ± 0.00200.5246 ± 0.0027
DReW89.1K0.4289 ± 0.00150.5311 ± 0.0066
PathNN-SP80.8K0.4285 ± 0.00130.5269 ± 0.0038
PDE-GCN55.7K0.4225 ± 0.00110.5173 ± 0.0025
GraphCON55.9K0.4262 ± 0.00080.5250 ± 0.0043
A-DGN55.8K0.4249 ± 0.00130.5215 ± 0.0029
PH-DGN-C51.7K0.4273 ± 0.00110.5242 ± 0.0023
SWAN59.9K0.4252 ± 0.00140.5225 ± 0.0033
SONAR80.8K0.4253 ± 0.00100.5208 ± 0.0018
GRAND55.7K0.4224 ± 0.00120.5169 ± 0.0023
GREAD-BS47.5K0.4240 ± 0.00120.5186 ± 0.0026
PIANO56.2K0.4228 ± 0.00140.5197 ± 0.0027

PIANO's MAE is close to GRAND/PDE-GCN and below the other displayed architectures, including MixHop, DReW and PathNN-SP. This demonstrates competitive transfer of explicit spatial-order control to a new weather environment.

PEMS-BAY: accuracy and model size on a road graph

PEMS-BAY (Li et al., 2018) predicts twelve five-minute steps from twelve observations on a road graph, with chronological 70/10/20 splits and train-only normalization. MAE is in native units; Mean MAE pools all twelve leads.

MethodParamsMAE₁₅ ↓MAE₃₀ ↓MAE₆₀ ↓Mean MAE ↓
MixHop36.4K1.383 ± 0.0011.813 ± 0.0012.345 ± 0.0031.778 ± 0.001
PAN convolution36.4K2.505 ± 0.3892.755 ± 0.3113.139 ± 0.2292.764 ± 0.323
DReW136.3K1.352 ± 0.0031.730 ± 0.0042.140 ± 0.0051.679 ± 0.003
PDE-GCN36.2K1.402 ± 0.0011.879 ± 0.0022.512 ± 0.0041.856 ± 0.002
GraphCON28.1K1.391 ± 0.0011.863 ± 0.0022.481 ± 0.0041.838 ± 0.002
A-DGN28.0K1.387 ± 0.0011.840 ± 0.0012.413 ± 0.0041.809 ± 0.001
PH-DGN-C23.9K1.420 ± 0.0151.898 ± 0.0282.523 ± 0.0571.872 ± 0.032
SWAN32.1K1.388 ± 0.0011.838 ± 0.0012.415 ± 0.0031.808 ± 0.002
SONAR*53.0K1.410 ± 0.0011.894 ± 0.0042.546 ± 0.0101.874 ± 0.005
GRAND28.0K1.432 ± 0.0021.937 ± 0.0042.635 ± 0.0091.922 ± 0.004
GREAD-BS19.8K1.437 ± 0.0001.950 ± 0.0012.652 ± 0.0021.933 ± 0.001
PIANO28.4K1.376‾\underline{1.376} ± 0.0011.799‾\underline{1.799} ± 0.0032.303‾\underline{2.303} ± 0.0061.758‾\underline{1.758} ± 0.003

Bold marks the best displayed means; underlining marks second-best means. SONAR* uses five seeds at the paper's reporting cutoff; other rows use ten. Protocols and completion details: B.4; I.3; I.5.

PIANO offers a strong accuracy–capacity trade-off. Its separately validation-selected setting achieves 1.7583 ± 0.0035 Mean MAE, second among the completed ten-seed methods listed and below MixHop's 1.7782 ± 0.0014. PIANO uses 28.4K parameters, only 20.9% of DReW's 136.3K; DReW achieves 1.6794 ± 0.0028. The table retains original PIANO; the additional selection budget is disclosed in I.3.

METR-LA: transfer to a second traffic dataset

PIANO's separately selected setting achieves MAE 3.4473 ± 0.0049, below fixed-recipe GRAND/GREAD-BS 3.7455 ± 0.0028 / 3.7610 ± 0.0068. DReW and the completed ten-seed PathNN-SP give 3.2860 ± 0.0118 / 3.3085 ± 0.0158, respectively (Table 41).

Ocean: measured computational efficiency

PIANO uses 43.1K versus PathNN-SP's 67.7K parameters. The isolated audit measures 0.790 versus 11.763 s/epoch and 0.415 versus 6.976 ms per nominal sample (Table 28). These timings use original LR0.001 checkpoints; the selected LR0.003 accuracy setting is separate. Historical ConvLSTM reaches higher overall Mean ACC, 0.547 ± 0.004; full cohorts are in Appendix E.

Together, PIANO combines competitive forecasting with a distinct architectural contribution: efficient redistribution of propagation mass across explicit spatial orders as the latent state evolves within a continuous-time graph model.

AuthorsOpenReview ↗

Q1, Q2 & Q4 — Highlighting, nonlinear theory and controlled experiments

Q1 — Corrected table highlighting

We corrected bolding and clarified the comparison sets. Table 2 now marks CSG-ODE as best on Wave (3.26 versus PIANO's 3.79, in 10510^5 MSE). Full reference sets, including Graph WaveNet and ConvLSTM, remain in Appendices E–F.

PIANO has the best displayed main-table means on Ocean, SEVIR and PubMed, and is competitive with the strongest EngRAD baselines. Together with compact models and measured Ocean efficiency, these results support a favorable accuracy–resource trade-off (Comment 2).

Q2 — Scope of the nonlinear theory

Theorem 2 bounds frozen linear transport, isolating the roles of spatial order LL and integration horizon TT. In the full model, derivatives also pass through the globally pooled controller, adding coefficient sensitivity beyond local feature transport. Extending the radius analysis requires separate bounds on these effects. Theorem 1 already establishes well-posedness of the full nonlinear flow, whose spatial-order weights adapt to the evolving state. Revised §4.2 and Appendix C clarify this distinction.

CW3 / Q4 — Controlled settings and interpretation

We expanded §5.1 and Appendix D, with direct links to Figures 2 and 4.

Wave—adaptive propagation range. A fixed 64-node ring connects first/second neighbors. With the gate off, models observe ten displacement/velocity frames and predict twenty; hidden ρ∈{1,2,4}\rho\in\{1,2,4\} varies interaction range. Learned effective radius tracks ρ\rho (Spearman 0.943 over 200 matched families for the diagnostic checkpoint). Swapping only walk-length schedules at matched solver times increases error by orders of magnitude, demonstrating functional use of the learned weights.

Springs—structural modulation. Ten particles use the same 10-to-20-frame task, with a supplied graph resampled per trajectory and stiffness κ=0.1\kappa=0.1 or 1.01.0. At weak coupling, removing/shuffling structural information increases MSE from 5.28 × 10⁻⁴ to 8.85–9.12 × 10⁻⁴. Effects are small at strong coupling. This identifies a useful regime for correctly aligned structural information and motivates the gate's optional, task-dependent role.

Q4 — Matched controllers and ongoing feedback

H0H_0-frozen fixes weights predicted from the initial state; H0,tH_0,t-conditioned uses the initial state and solver time; full PIANO uses the evolving state and time. Within each task, architecture, parameter count and field-evaluation count are matched. Results are mean ± sample SD over ten paired seeds (Table 3; D.11):

ControllerSource-disjoint Wave: 10510^5 MSE ↓Bounded SEVIR: 10310^3 MSE ↓Ocean: Mean ACC ↑
H0H_0-frozen4.087 ± 0.6866.6199 ± 0.10020.523110 ± 0.000975
H0,tH_0,t-conditioned4.144 ± 0.8236.6147 ± 0.06220.523152 ± 0.000969
Full evolving-state PIANO3.874 ± 0.6206.6036 ± 0.05720.524505 ± 0.001302

Full has the best displayed mean in all three settings. Ocean improves in all ten pairs against each control, supporting a practical benefit from ongoing feedback. Wave/SEVIR mean improvements remain uncertain in the paired analysis. Wave uses the source-disjoint companion; bounded SEVIR uses a separate event-disjoint protocol.

Reviewer_Q2oiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors think existing graph neural ODEs employ a relatively fixed spatial propagation structure when computing state changes, limiting their ability to adapt effectively to the system state. Then, this paper introduce PIANO (Path-Integral Adaptive Neural ODE), a graph neural ODE that adapts the mixture of propagation orders to the evolving latent state, allowing information to propagate over different spatial ranges. Two spatial controls (state-adaptive path-integral propagator and selective topology-aware node modulation) were proposed to address the problem identified by the authors. The author also provides a substantial amount of theoretical proof in the paper. Finally, experiments on synthetic dynamics, ocean and severe-weather forecasting, and node classification demonstrate competitive performance and scalability.

Critical strengths

1.Clear presentation and coherent argumentation. The paper develops a logical progression from problem motivation and method design to theoretical analysis and experimental validation. This makes the rationale, contributions, and scope of the claims easy to follow. 2.Good experimental performance. Controlled experiments, ablations, and inference-time interventions provide evidence that state-adaptive propagation is useful. Results across multiple tasks further support the method’s practical effectiveness.

Critical weaknesses

The practical benefits of the adaptive components in node classification tasks are not obvious. As shown in Table 21, the fixed-weight version and the full model both achieve a score of 0.895 on PubMed, while on ogbn-arxiv, the performance improves only marginally from 0.664 to 0.666. Table 23 reveals that shuffling or removing topological signals has negligible impact on predictions, indicating that the model does not effectively utilize this structural information. Although synthetic experiments offer mechanistic evidence, results from node classification tasks fail to clearly demonstrate which scenarios actually necessitate state-adaptive mechanisms and topological gating. The paper needs to more explicitly distinguish between the benefits derived from standard multi-hop propagation and those attributable to the newly introduced modules, as well as analyze the conditions under which these modules are applicable.

Questions for authors

Please provide a further explanation of the reasons based on the critical weakness.

1 comment
AuthorsOpenReview ↗

Thank you for your constructive review: component attribution and applicability

Comment 1 — Paper revisions

Thank you. We now separate multi-order representation, ongoing state feedback and structural gating, with matched controls and explicit applicability discussion (Table 3; D.11–D.12; G.2–G.3). PIANO's architectural contribution is an explicit, state-conditioned distribution over spatial orders inside the continuous-time vector field.

Revised Table 1 and Appendix H distinguish spatial operators, explicit order and order-weight adaptivity. Expanded comparisons include A-DGN (Gravina et al., 2023), SWAN (Gravina et al., 2025), MixHop (Abu-El-Haija et al., 2019) and DReW (Gutteridge et al., 2023), plus EngRAD forecasting (Marisca et al., 2024; Table 5, Appendix I).

Comment 2 — Component attribution and applicable regimes

Q1: What comes from multi-hop propagation, and what comes from feedback?

The original component study (G.2, Table 35) reports one-hop / learned-fixed multi-order / full accuracy of 89.1 / 89.5 / 89.5% on PubMed and 66.1 / 66.4 / 66.6% on arxiv. These descriptive means support multi-order representations; they do not establish that evolving-state feedback or structural gating is necessary for classification.

We therefore added gate-off controls matched in architecture, parameters and NFE. Both initial-state controls retain input-conditioned multi-order propagation, so this comparison isolates recomputing weights from the evolving state. Mean ± sample SD over ten paired seeds (D.11):

ControllerPubMed: accuracy (%) ↑ogbn-arxiv: accuracy (%) ↑
H0H_0-frozen89.632 ± 0.68467.385 ± 0.460
H0,tH_0,t-conditioned89.670 ± 0.50067.199 ± 0.658
Full evolving-state PIANO89.640 ± 0.55067.300 ± 0.573

The additional feedback gain remains uncertain on these static tasks. Initial graph/features may already provide sufficient conditioning information.

Separate architectural comparisons show practical effectiveness: PIANO reaches 89.645 ± 0.534% on PubMed, the highest displayed original-split mean, and ranks first/second on Cora/Citeseer among five newly matched models (G.4–G.5). Dataset references: Planetoid (Yang et al., 2016) and OGB (Hu et al., 2020). Architectural rankings and component attribution are reported separately.

Q2: Where is there evidence for ongoing feedback?

All arms retain the same multi-order architecture, parameters and NFE. Mean ± sample SD over ten paired seeds (Table 3; D.11):

ControllerSource-disjoint Wave: 10510^5 MSE ↓Bounded SEVIR: 10310^3 MSE ↓Ocean: Mean ACC ↑
H0H_0-frozen4.087 ± 0.6866.6199 ± 0.10020.523110 ± 0.000975
H0,tH_0,t-conditioned4.144 ± 0.8236.6147 ± 0.06220.523152 ± 0.000969
Full evolving-state PIANO3.874 ± 0.6206.6036 ± 0.05720.524505 ± 0.001302

Ocean supplies direct real-data evidence: Full improves in all ten pairs against each control. Wave/SEVIR means also favor Full, while their feedback gains remain uncertain. EngRAD's controller means are nearly identical (D.11). Ocean supports ongoing feedback in a real dynamical setting; evolving spatial requirements are a candidate condition for its utility.

Q3: When does structural information add value beyond feature gating?

The full gate combines feature and closed-walk structural signals; feature-only removes the structural term. Mean ± sample SD over ten seeds per arm (D.12; Springs in D.5):

Setting / metricFeature-only gateFull gate
PubMed: accuracy (%) ↑89.432 ± 0.76689.599 ± 0.583
ogbn-arxiv: accuracy (%) ↑67.250 ± 0.47467.076 ± 0.709
Springs, weak coupling: 10410^4 MSE ↓9.12 ± 0.355.28 ± 0.75
Springs, strong coupling: 10310^3 MSE ↓3.79 ± 0.123.79 ± 0.13

Static structural gains remain uncertain. In the original arxiv checkpoints, at least 99% of gates exceed 0.95 (G.3), consistent with weak structural sensitivity. These gate cohorts are separate from Q1's controller cohorts.

Weak-coupling Springs demonstrates structural utility; stronger coupling reduces it. This supports treating the gate as optional. Candidate useful conditions are predictive closed-walk roles beyond features, variation across nodes, and a gate responsive to that variation. Detailed protocols and paired analyses are in the paper.

Reviewer_6wpwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PIANO, a graph neural ODE whose vector field mixes normalized adjacency powers using a state-conditioned distribution over walk lengths. A globally pooled latent state and solver time determine the order weights at every vector-field evaluation. An optional node gate uses feature magnitude and closed-walk statistics. Sparse recursion evaluates the polynomial without constructing dense graph powers. The analysis establishes pointwise non-expansiveness of each realized spatial operator, well-posedness of the nonlinear flow, and a receptive-field approximation bound for a frozen linear surrogate. Experiments include controlled wave and spring systems, ocean and radar forecasting, and node classification.

Critical strengths

The paper is clearly written and well motivated. Table 1 distinguishes state-conditioned order mixing from fixed diffusion, adaptive one-hop edges, and learned graphs. The figures are informative, and the theoretical analysis carefully separates pointwise non-expansiveness, nonlinear well-posedness, and the frozen-linear receptive-field bound.

The controlled experiments provide convincing evidence that the model uses the proposed mechanism. On Graph-wave-Scale, the controller comparisons, scale–radius correlation, and schedule interventions support the usefulness of state-conditioned propagation weights. The paper also reports negative and regime-dependent results across Springs, forecasting, and node classification, which helps establish the method’s practical scope.

Critical weaknesses
  1. The benefit of recomputing the weights throughout integration is not directly isolated. The main novelty lies in updating explicit multi-order weights from the evolving state, rather than in polynomial filtering or path-length weighting themselves. However, the controls do not include weights predicted once from H0H_0 and then frozen. The current evidence therefore establishes the value of state-conditioned mixing more clearly than the additional value of ongoing feedback. The results of α(H)\alpha(H) and α(H,t)\alpha(H,t) shown in Table 8 indicates that explicit time conditioning adds little improvement.

  2. Evidence for the topology-aware gate is narrower than for adaptive order mixing. It helps on weakly coupled Springs but contributes little at stronger coupling, as shown in Table 9. Structural-score interventions have negligible effects on SEVIR and ogbn-arxiv, as shown in Figure 11 and Table 23. The paper acknowledges this limitation and presents the gate as optional; nevertheless, its practical value beyond the controlled setting remains less established.

Questions for authors
  1. Could you compare against an α(H0)\alpha(H_0)-frozen control, predicting the weights once and holding them fixed throughout integration while matching the architecture, parameters, and NFE? This would directly test the contribution of ongoing state feedback.

2。 Is there evidence in the existing real-data experiments that the structural component of the gate provides a reproducible benefit, separately from feature-dependent gating? If not, what properties of a task or graph would make this component useful?

3 comments
AuthorsOpenReview ↗

Thank you for your constructive review: new controls and clarifications

Comment 1 — Paper revisions

Thank you for highlighting the need to isolate the adaptive components. PIANO's central contribution is explicit, state-conditioned redistribution across spatial orders inside a continuous-time vector field. We have strengthened the evidence for this contribution and clarified the optional role of structural gating.

We added independently trained initial-state controls to §5.1, Table 3 and Appendix D.11, including the completed Ocean study. Appendices D.12 and E.6 separate structural information from feature-dependent gating. We also expanded Table 1 and the baseline coverage, added EngRAD [1] weather forecasting and further traffic comparisons, and clarified the experimental environments and interpretations (Tables 4–5; Appendices D–I).

Comment 2 — Direct answers

Q1: Benefit of ongoing state feedback

We added independently trained H0H_0-frozen and H0,tH_0,t-conditioned controls, matched to the full model in architecture, parameter count. Results are mean ± sample SD over ten paired seeds (Table 3; Appendix D.11):

ControllerSource-disjoint Wave: 10510^5 MSE ↓Bounded SEVIR: 10310^3 MSE ↓Ocean: Mean ACC ↑
H0H_0-frozen4.087 ± 0.6866.6199 ± 0.10020.523110 ± 0.000975
H0,tH_0,t-conditioned4.144 ± 0.8236.6147 ± 0.06220.523152 ± 0.000969
Full evolving-state3.874 ± 0.6206.6036 ± 0.05720.524505 ± 0.001302

On Ocean, the full controller improves Mean ACC in all ten pairs against each control, supporting an additional benefit from ongoing state feedback. The Wave and bounded-SEVIR mean results also favor Full, although these effects remain uncertain; EngRAD's controller means are nearly identical. Appendix D.11 gives the complete paired analyses.

Q2: Structural contribution beyond feature gating

The current real-data comparisons do not yet establish a reproducible additional benefit from the structural term beyond feature-dependent gating (D.12; E.6). We therefore retain the structural gate as an optional component.

Controlled weak-coupling Springs identifies a useful regime: full/feature-only MSE is 5.28 × 10⁻⁴ / 9.12 × 10⁻⁴. At strong coupling, both means round to 3.79 × 10⁻³.

The evidence suggests three candidate conditions for structural utility:

  • Structural roles carry predictive information beyond node features.
  • Closed-walk scores vary across nodes.
  • The learned gate remains responsive to that variation.

These task-dependent findings complement the real-data evidence for PIANO's evolving-state order control.

[1] Marisca, I., Alippi, C., and Bianchi, F. M. (2024). Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling. ICML.

Thanks the responses, which have addressed my questions. I decide to maintain my current rating and confidence level.

Thank you for reading our response and letting us know that it addressed your questions. We appreciate your thoughtful review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to the reviewers

Thank you for the constructive reviews. PIANO's central contribution is state-conditioned redistribution across explicit spatial orders inside a continuous-time graph vector field. It adapts the relative contributions of different hop orders as the latent state evolves, with sparse propagation cost O(L∣E∣d)O(L|E|d).

  • Revisions and expanded coverage. We added references and comparisons covering all ten requested model families, deepened the PAN comparison, and evaluated EngRAD, METR-LA, PEMS-BAY and Peptides-func. Revised Table 1 distinguishes spatial reach and order adaptivity. We clarified controlled settings and interventions, corrected table highlighting, and distinguished nonlinear well-posedness from the frozen-linear transport bound. Substantive changes are orange.

  • Direct evidence for ongoing feedback. New independently trained initial-state and initial-state/time controls match PIANO's architecture, parameter count and vector-field evaluation budget. On Ocean, Full achieves 0.5245 Mean ACC versus 0.5231/0.5232, improving in all ten paired seeds against each control, with positive 97.5% paired intervals (Table 3; D.11). This supports an additional benefit from evolving-state feedback under the selected protocol. Wave interventions further demonstrate functional use of the learned spatial-order weights. Additional feedback gains remain uncertain on several other tasks; structural gating is optional and task dependent.

  • Broader practical effectiveness. Expanded comparisons give PIANO the best displayed main-table means on Ocean, SEVIR and PubMed (Tables 4–5), with Ocean’s distinct selection budgets disclosed. EngRAD performance is close to GRAND/PDE-GCN. On PEMS-BAY, a separately validation-selected PIANO setting achieves 1.7583 ± 0.0035 Mean MAE versus MixHop’s 1.7782 ± 0.0014, using 28.4K versus 36.4K parameters. This gives 1.12% lower mean error with a compact model, supporting the practical value of explicit spatial-order control. The additional validation selection is documented in Appendix I.3.

  • Measured efficiency. An isolated audit of the original Ocean checkpoints measures 14.9× faster training and 16.8× faster inference than PathNN-SP, with 43.1K versus 67.7K parameters (Appendix E). These measurements support the computational practicality of explicit multi-order propagation; their recipe is identified separately from the selected accuracy setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.