acceptodds
Under review as a conference paper at ICLR 2027

When to Aggregate: Delay-Aware Message Scheduling with Physics-Motivated Operator Decomposition for Dynamics Prediction

Abstract

Graph neural networks have become a core paradigm for dynamics prediction by explicitly modeling spatial interactions among entities. Most existing approaches rely on synchronous message passing, in which all graph components are updated at every propagation layer. Although effective, this paradigm couples propagation depth with interaction-update frequency: increasing the processor depth not only expands the receptive field but also forces every local interaction to undergo additional transformations. This coupling can be undesirable for autoregressive dynamics, where different interactions may evolve at different rates. In this paper, we propose an asynchronous graph propagation scheme that regulates message aggregation by activating only a subset of edges at each layer. This edge-level mechanism enables information to propagate in a staggered manner, decoupling update frequency from propagation depth without modifying the underlying model architecture. It is model-agnostic and can be integrated into existing message-passing neural networks with minimal implementation effort. We further reinterpret graph propagation as the numerical integration of edge-local and node-local physics-motivated generators and decompose the latent dynamics into complementary operators. Extensive experiments across diverse benchmarks show that asynchronous propagation and operator decomposition consistently improve forecasting accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
Dynamical systems, long-horizon forecasting, asynchronous message passing, multirate learning

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_kCBoOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an asynchronous, edge-level message scheduling scheme for GNN-based autoregressive dynamics prediction. At each intermediate processor layer, only a quantile-selected subset of edges delivers its messages, ranked by a state-dependent priority score τe=de/(∥Fe∥2+ε)\tau_e = d_e / (\lVert F_e \rVert_2 + \varepsilon). The increments of inactive edges are partially retained in a persistent per-edge buffer through a component-wise gate and are released at the next activation. A terminal full refresh then flushes all buffers before decoding. The scheduler is combined with a physics-motivated decomposition of the latent vector field into transport, diffusion, reaction, skew-symmetric coupling, and source-sink forcing, followed by a projection block, with transport and diffusion assigned separate priority clocks and activation ratios. The authors establish several structural properties (synchronous consistency, permutation equivariance, bounded buffers, terminal synchronization, and a conditional rollout bound) and report the lowest RMSE among nine baselines on four benchmarks, supported by an extensive set of ablations.

Critical strengths

1. A clear, simple idea that plugs into existing models. The observation that synchronous message passing ties propagation depth to interaction-update frequency is intuitive, and the question of when a message should be delivered has received comparatively little attention in the learned-simulator literature. The scheduler leaves the graph topology and backbone untouched, and the percentile rule itself introduces no learnable parameters. Table 3 shows that changing only the propagation schedule improves both MGN and MP-PDE on all four datasets, which lends credibility to the claim that the method is model-agnostic.

2. Thoughtful controlled comparisons. The authors clearly anticipated the "this is just sparsification" objection. Under a matched activation budget, they compare the proposed priority against random, static-distance, and magnitude-based policies (Table 6). They also compare the masked buffer against no buffer, latest-message-only, and random or shuffled gates (Table 5), and the full scheduler against node-batch updates, delayed aggregation, and multi-rate gating (Table 8). The staleness diagnostics in Table 7 are a welcome addition.

3. Careful formulation and honest scoping of claims. Several technical details are handled with care, including the orientation consistency of the upwind transport flux, the symmetrization of the diffusivity together with zero-sum incidence scattering, and the role of the tie-breaking rule in permutation equivariance (Remark H.4). I also appreciate that the authors are explicit about what they do not claim. The rollout bound does not imply that asynchronous scheduling is more stable than its synchronous counterpart, explicit Euler integration does not preserve the latent norm, and masking alone does not reduce computation.

Critical weaknesses

1. The central hypothesis is not directly tested, and the source of the gains remains unclear. The paper argues that deferring weak or slowly varying interactions reduces redundant transformations and thereby mitigates error accumulation during autoregressive rollout. The experiments, however, report only aggregate RMSE.

  • There are no plots of error versus rollout step and no estimate of the effective Lipschitz constant KAK_A for the trained synchronous and asynchronous models. Corollary 3.5 holds for any one-step map, so it does not distinguish the proposed method from the baselines.
  • There is no analysis of which edges are deferred, for example whether they lie outside the wake, in low-velocity regions, or where gradients are small. Without such evidence, the "deferring slow interactions" narrative remains a hypothesis.
  • The degree of asynchrony is in fact quite shallow. With L=4L = 4 and ρ=0.8\rho = 0.8, the mean waiting depth is only 1.1 to 1.5 layers, with a maximum of 3. This resembles a structured perturbation of roughly 20% of the edges per layer and is difficult to distinguish from regularization effects such as DropEdge or stochastic depth. A baseline that drops edges at random during training but propagates synchronously at inference is missing.
  • The most direct test of the paper's thesis is absent. If the problem is the coupling of depth and update frequency, the benefit of asynchrony should grow with depth, yet no sweep over LL is reported.
  • Although the title emphasizes "delay-aware" scheduling, none of the benchmarks involves genuinely delayed dependencies, and the authors themselves describe the priority score as a relative ranking rather than a physical delay.

2. Concerns about experimental rigor and fairness.

  • Several improvements are small. The gain on Gray-Scott is 0.0089 versus 0.0093, and most single-operator ablations differ by only 1 to 2%. Standard deviations are reported only in Table 1, so the significance of the ablation results cannot be assessed.
  • Some numbers are identical across tables, e.g., Delayed Aggregation in Table 8 and Full Accum. in Table 5, and Random Edge Update in Table 8 and Random in Table 6. These may well be reused configurations, but this should be stated explicitly.
  • The comparison between the full model and MGN is confounded. The full model includes a self-attention-based global refinement GattG_{\text{att}}, and removing the projection block P\mathcal{P} causes the largest degradation in the ablation (e.g., from 0.1426 to 0.1579 on CF). A substantial share of the gain may therefore stem from global attention rather than from the physics-motivated decomposition. The parameter count and FLOPs of the full model are not reported, and MGN and CiGNN are not equipped with the same projection block.

3. Limited practical value, with added complexity but no efficiency gains.

  • Despite the "asynchronous" framing, the method yields no computational savings. Training is about 6% slower per epoch and inference is also slower, because all edge messages must still be computed before ranking. The computation-sparse variant is discussed but neither implemented nor benchmarked.
  • The method introduces many components (five operators, two priority clocks, gates, buffers, and a projection block), and ρT\rho_{\mathrm{T}} and ρD\rho_{\mathrm{D}} are set by hand. Table 10 shows that (ρT,ρD)=(0.6,0.6)(\rho_{\mathrm{T}}, \rho_{\mathrm{D}}) = (0.6, 0.6) performs worse than the synchronous model on three of the four datasets, indicating considerable sensitivity to the scheduling ratios.
  • The main text describes the mechanism in terms of increments Δm\Delta m, whereas Appendix C uses a node-message formulation in terms of mm. It is unclear which of the two is actually implemented.
Questions for authors
  1. Is the delivery mask applied during both training and inference? Since the hard quantile mask is non-differentiable, do gradients flow only through the delivered increments, and does detaching the message norm in the priority score affect the results?
  2. At a matched average activation rate, how does a model trained with random edge dropping and evaluated with full synchronous propagation compare with the proposed method? This control seems essential for separating the effect of the scheduling mechanism from that of regularization.
  3. Could you provide direct evidence for the claimed mechanism? In particular, how does the gain over synchronous propagation change for L∈{4,8,16}L \in \{4, 8, 16\}, how does RMSE evolve over the rollout horizon, and do the deferred edges correlate with physical quantities such as local velocity, gradients, or vorticity?
  4. What are the parameter count and FLOPs of the full model, and how much of the gap remains if MGN or CiGNN is equipped with the same GattG_{\text{att}} projection block? For the prior asynchronous and multi-rate GNNs in Table 8 (e.g., GwAC, Co-GNN, DRew, and G²), did you run the official implementations or re-implement them under a common backbone?
  5. Please explain the identical numbers in Tables 5, 6, and 8, and clarify whether the scheduling ratios and retention gates were selected on the validation set separately for each dataset. Could you also report longer rollouts on Black Sea together with physically meaningful metrics such as energy spectra or the drift of conserved quantities?
5 comments
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer kCBo (Part 1)

We sincerely thank you for your detailed review and constructive suggestions. We appreciate your recognition of the simple, portable idea, the controlled comparisons, and the careful scope of the theoretical claims. Below, we address the mechanism, experimental controls, implementation, and practical trade-offs point by point.

Q1: Are delivery masks used during both training and inference? How do gradients reach deferred messages under hard selection, and does detaching the ranking norm matter? Which implementation is used: main-text increments Δm or Appendix C messages m?

Reply: The delivery schedule (selection, buffers and terminal refresh) is used in both training and inference. Deferred messages can receive gradients through later delivery. Once a buffered increment enters a subsequent node update, the loss can backpropagate to it and its retention gate, since the buffer computation graph is retained.

Hard quantile selection does not propagate ranking gradients. The message norm is used only for this selection, so detaching it changes neither the forward ranking nor the gradient path.

The two notations describe two uses of the same buffer recurrence in Eqs. (14)/(S25):

  • Appendix C: directed messages m in generic MPNNs, including MGN and MP-PDE, with MLP-based node updates.
  • Algorithm 1: scaled operator increments Δm, followed by incidence scattering and an additive update.

Q2: Could the gains from asynchrony reflect DropEdge-like regularization? At a matched activation budget, how does random training-time edge dropping followed by synchronous inference compare?

Reply: Table J includes DropEdge-like random edge dropping during training followed by synchronous inference. It is worse than synchronous propagation on all four datasets. Dropping edges permanently discards their information and, when training with dropping but inferring on the full graph, changes the aggregation statistics between training and inference.

Our mechanism defers rather than discards: deferred increments are kept in the buffer and delivered later, and the terminal refresh ensures that every edge contributes within each forecasting step. Training and inference use the same schedule, and Full Accum. loses no increment; the learned gate in Masked Accum. only filters components of deferred increments, and it further improves over Full Accum. on CF, BS and Burgers (Table 5).

Table J: Scheduling-policy comparison

VariantCFBSGSBurgers
Synchronous0.18920.61470.01930.0118
Random drop, no buffer (random training; synchronous inference)0.19600.66120.02020.0234
Random selection with buffer and mask (Table 6)0.18540.60650.01890.0099
Proposed priority0.17250.56010.01860.0051
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer kCBo (Part 2)

Q3: Please compare gains at L ∈ {4, 8, 16}, rollout error, perturbation test, and deferred edges versus velocity, gradients, or vorticity. Should gains increase with depth? How is “delay-aware” justified without demonstrated physical delays, given that the priority is a relative ranking and the corollary does not distinguish models?

Reply: Table K compares L ∈ {4, 8, 16} for the MGN backbone on Burgers. Asynchronous delivery is better at every depth, with a stable relative reduction of 56.8%, 52.9% and 54.5%. At ρ = 0.8 with terminal refresh, the overall delivery fraction is ρ+(1−ρ)/L, as listed below. Rollout error will be shown in a new figure (Fig. S2) in the revision.

Table K: Depth sensitivity and delivery budgets.

LSynchronous RMSEAsynchronous RMSEActual overall delivery fraction
40.01180.005185.00%
80.01040.004982.50%
160.00990.004581.25%

Regarding the deferred edges, the priority is defined on the latent edge increments, and we do not claim that deferred edges correspond to identifiable physical regions such as low-velocity or low-vorticity zones. Table 7 reports mean waiting depths of 1.12–1.47, a maximum of 3, and pre-terminal coverage of 91.3%–97.7%. Independent random selection of 80% of edges over three intermediate steps would give about 99.2% expected coverage. The lower observed coverage is consistent with persistent selection across layers, rather than independent random perturbations.

Corollary 3.5 is conditional and does not establish smaller sensitivity constants for asynchronous models (Remark H.10). One-step error under input perturbation is reported in Table L with masks recomputed after perturbation.

“Delay-aware” refers to delivery across processor micro-steps within one forecasting step, not a calibrated physical travel time, as stated in Sec. 3.2 and Appendix C. It does not require explicit delay terms in the data equations. CF, GS and Burgers have no such terms; the daily Black Sea data do not establish physical delay dependencies. We will add a one-sentence definition of this usage in the introduction and at the start of Sec. 3.2.

Table L: One-step error under input perturbation on Burgers.

Perturbation (standard deviation)Synchronous one-step RMSEAsynchronous one-step RMSE
0(zero perturbation)3.48e-51.52e-5
1e-43.57e-51.627e-5
1e-24.11e-51.832e-5

Q4: Could global attention explain the full model’s advantage over MGN/CiGNN? Please report full-model parameters and compare backbones equipped with the same G_att. Were the purported prior methods in Table 8 (GwAC, Co-GNN, DRew, G²) run officially or reimplemented under a common backbone?

Reply: Global attention does not account for the scheduling gain: in Table 3, MGN and MP-PDE have no G_att and improve on all four datasets with scheduling alone, and in Tables 2 and 10 the synchronous and asynchronous full models share the same operators and G_att.

To separate attention from the full-model advantage over MGN, Table M compares MGN + G_att with synchronous and asynchronous delivery. For MGN + G_att, asynchronous delivery lowers RMSE on all four datasets, and the full model is lower than MGN + G_att in both modes on all four datasets, so attention alone does not account for its advantage. On CF, adding G_att does not improve synchronous MGN (0.2615 vs 0.1892), while on BS, GS and Burgers it does. Parameter counts are listed in Table M.

Table M: Global-attention controls

ModelParametersCFBSGSBurgers
MGN(4 layers)0.57M0.18920.61470.01930.0118
MGN + G_att, synchronous0.83M0.26150.53450.01020.0083
MGN + G_att, asynchronous0.85M0.19120.52940.00980.0053
Ours (4 layers)0.96M0.14260.51910.00890.0044

Table 8 contains no official implementations of GwAC, Co-GNN, DRew or G²; these methods are discussed in related work only. Table 8 compares update mechanisms implemented by us under a common backbone and a matched update budget (Node-Batch Async and Multi-Rate Gating), together with variants that coincide with rows of Tables 3, 5 and 6 (see Q5).

AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer kCBo (Part 3)

Q5: Why are results repeated across Tables 5, 6, and 8? Were ratios and retention configurations selected separately on each dataset’s validation set? How should sensitivity to manually set ratios be interpreted when (0.6, 0.6) loses to synchronous delivery on three tasks? Can longer Black Sea rollouts include meaningful physical metrics such as spectra or conservation drift?

Reply: Repeated values are reused results, not independent runs. Table 8’s Delayed Aggregation, Random Edge Update, and MGN + Async. correspond to Full Accum. in Table 5, Random in Table 6, and MGN + Async. in Table 3, respectively; MGN + Async. also matches Masked Accum. in Table 5. We will cross-reference these rows in the captions.

Ratios and gates. Hyperparameters, including the scheduling ratios and the retention-gate configuration, are selected on each dataset’s own validation set (Appendix L.4); Table 10 reports test performance of the candidate settings. More deferral is not always better. Table 4 favors ρ = 0.8 over synchronous MGN on all four tasks, while the reported settings with ρ ≤ 0.4 are worse. In Table 10, (0.8, 0.8) is best among the listed settings; (0.6, 0.6) is worse than synchronous delivery on CF, BS, and Burgers but better on GS (0.0094 versus 0.0098).

Black Sea. Existing results use a common 20-step rollout protocol for all methods. Atmospheric forcing and data assimilation mean that surface kinetic energy is not a conserved quantity here, so we do not report conservation drift for this dataset.

Q6: What practical benefit offsets the added complexity when all edge messages are still computed, training is about 6% slower, inference is slower, and the computation-sparse variant is unimplemented?

Reply: The practical benefit is improved forecasting with modest runtime overhead, not acceleration. Table 9 within more broader and larger architecture reports aggregation reduced from 100% to 82.5%, while training rises from 44.2 to 46.8 seconds/epoch and inference from 0.0079 to 0.0085 seconds/step because all candidate messages are still computed.

Applying the scheduler to an existing backbone does not require the five heads or dual priorities. It adds the retention gate (about 0.025M parameters for the 4-layer configuration and 0.049M for the 8-layer configuration) and the buffering mechanism. Table 3 demonstrates the resulting accuracy benefit on MGN and MP-PDE. Computation-sparse execution remains future work.

Concluding remark: Thank you again for your constructive feedback. We hope these clarifications address your concerns and will incorporate the completed results and implementation details into the revision.

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

The paper studies message-delivery schedules for autoregressive dynamics prediction on graphs. It proposes state-dependent edge activation, componentwise retention of deferred messages, persistent edge buffers, and a terminal full refresh within each forecasting step. The full model combines this mechanism with transport, diffusion, reaction, skew-coupling, and source operators, using separate transport and diffusion priorities and a post-update refinement/projection block. Experiments cover four dynamical systems, comparisons with forecasting baselines, and controlled studies of scheduling and accumulation. The reported results support consistent forecasting improvements from deferred aggregation.

Critical strengths
  1. The paper provides meaningful controlled evidence for deferred aggregation. Table 3 improves both MGN and MP-PDE on all four datasets. Tables 5–8 examine buffering, retention, alternative priorities, and other partial-update mechanisms. In particular, full accumulation without learned masking already improves over the no-buffer variant, so the buffering benefit cannot be attributed entirely to additional gate parameters. These experiments make the scheduling contribution more convincing than comparisons with unrelated architectures alone.

  2. The overall forecasting results are consistent, and several important limitations are reported explicitly. Table 1 reports the best mean RMSE on all four benchmarks, with reductions of 4.3%–33.3% relative to the next-best model. The evaluation uses autoregressive rollouts and trajectory-level splits. The authors also distinguish aggregation sparsity from computational savings: Table 9 reports that all edges are still computed and that the current asynchronous implementation is slightly slower. Similarly, the theoretical discussion appropriately qualifies the rollout bound rather than claiming guaranteed stability superiority.

Critical weaknesses
  1. The distinctive operator-specific scheduling contribution is less directly established than generic deferred aggregation. Tables 6 and 8 evaluate the generic MGN-based asynchronous mechanism; Appendix M.2 specifies its priority as distance divided by message magnitude. Table 10 varies transport and diffusion activation ratios, but does not compare the proposed separate clocks against a shared rule or magnitude-only priorities within the same full architecture. Consequently, these results do not isolate the benefit of assigning transport and diffusion different distance scalings.

    The full model also includes additional operator heads and global attention refinement on every dataset (Table S3). Removing the refinement/projection block produces the largest single-block degradation in Table 2. The existing ablations therefore support the complete architecture, but leave uncertainty about the contribution of the specific physical parameterization. This distinction matters because graph advection–diffusion–reaction decomposition and operator splitting already appear in [ADR-GNN, Eliasof et al., AAAI 2024]. That work does not duplicate the proposed buffered scheduler, but it makes the operator-dependent scheduling mechanism an especially important contribution to isolate.

  2. Some numerical and operator interpretations are stronger than the stated construction supports. Equation (5) guarantees zero-sum edge scattering, but its arbitrary learned transformation does not guarantee the claimed smoothing. For two scalar nodes with κ=1\kappa=1 and ϕD(h)=−h\phi_D(h)=-h, the isolated diffusion contribution gives

    ddt(hi−hj)=2(hi−hj),\frac{d}{dt}(h_i-h_j)=2(h_i-h_j),

    so both the original and transformed contrasts grow.

    Likewise, the “first-order multirate” interpretation needs conditions beyond terminal synchronization. Consider a constant edge flux of one over a unit interval, with Δt=1/L\Delta t=1/L, an edge deferred until the final micro-step, and retention z=q<1z=q<1. Its total delivered increment is q+(1−q)/Lq+(1-q)/L, approaching qq, whereas integrating the unmodified flux gives one. Thus, flushing the retained buffer does not establish consistency with Eq. (3). These issues concern the stated interpretation of the architecture; they do not by themselves invalidate the reported forecasting results. The claims should be narrowed or the necessary restrictions made explicit.

Questions for authors
  1. Operator-specific contribution. With the full architecture, refinement block, retention mechanism, and aggregation budgets fixed, how do shared-clock and magnitude-only scheduling compare with the separate transport/diffusion clocks? This would directly test the additional contribution beyond generic deferred aggregation.

  2. Operator and integration properties. Are restrictions on ϕD\phi_D, retention gates, or physical waiting times omitted from the description that would justify smoothing and consistency with Eq. (3)? Otherwise, would the authors narrow these claims to a learned difference-based flux and an Euler-style residual architecture?

  3. Retention controls and capacity. Table 3 states that capacity is unchanged, while its asynchronous results match the learned masked-accumulation variant in Table 5 and Eq. (S24) introduces a learnable gate. Please specify the gate implementation and exact parameter counts for these comparisons. Also, the random/shuffled gates are substituted only during evaluation; their degradation can reflect distribution shift. Were alternatives trained under their own retention policies evaluated?

  4. Executed operator configuration. Appendix D states that QQ is disabled without observable forcing or controls, while Table S4 lists no forcing term for any dataset and removing QQ changes all four results in Table 2. Is QQ actually used as an unconditional or state-conditioned correction? Please clarify the per-dataset configuration, since it affects the interpretation of the operator ablations.

3 comments
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer usyo (Part 1)

We sincerely thank you for your thoughtful review and precise technical comments. We appreciate your recognition of our scheduling mechanism. Below, we address the operator-specific contribution, numerical interpretation, and implementation details.

Q1: What do separate T/D priorities contribute? With the full architecture, refinement, retention, and budgets fixed, how do shared-clock and magnitude-only policies compare? How are operator-specific gains isolated from extra heads and attention?

Reply: Separate T/D priorities are defined in Eq. (12). Table F compares them with shared-clock and magnitude-only scheduling within the full architecture, keeping retention, depth, terminal refresh, and per-operator delivery budgets fixed. Separate T/D rankings give the lowest RMSE on all four datasets.

Table F: Operator-specific scheduling controls (ρT=ρD=0.8)

ArchitectureCFBSGSBurgers
Separate T/D rankings0.14260.51910.00890.0044
Shared clock0.15000.53140.00970.0054
Separate ranking by magnitude0.14820.53890.00960.0052

Table 10 already shows that asynchronous delivery improves over synchronous delivery with the same operators and global attention. Table F addresses the narrower question of what separate T/D rankings add within the full architecture. The results will be included in the revision.

Q2: What restrictions justify smoothing in Eq. (5) and consistency with Eq. (3)? The counterexamples φ_D(h)=−h and constant flux with retention q yield growing contrasts and total delivery q+(1−q)/L.

Reply: Thanks for your comments. We will make two changes to the numerical interpretation:

  • Eq. (5): replace “smoothing” with “a learned difference-based flux with zero-sum scattering.” Incidence scattering applies each edge flux with opposite signs at its endpoints. Its contribution to the derivative of half the squared latent norm is −κ⟨hi−hj, φD(hi)−φD(hj)⟩. Monotonicity of φD makes this non-positive, but our model does not impose that restriction.
  • Eq. (15) and Appendix E: use “latent residual update.” For the constant-flux example, total delivery is q+(1−q)/L. Terminal refresh releases retained history; it cannot recover discarded components and does not establish first-order consistency with Eq. (3).

Proposition 3.1 remains unchanged: activating all masks recovers the synchronous discrete update. We will also align the abstract's numerical-integration wording with these statements.

Q3: How can Table 3 claim unchanged capacity when its asynchronous variant includes a learned retention gate? Please provide gate details and exact parameter counts. Since evaluation-only random/shuffled gates introduce distribution shift, were alternative retention policies also trained and evaluated consistently?

Reply: In Table 3, the backbone, depth, hidden dimension and training hyperparameters are unchanged. The asynchronous variant adds the learned retention gate z of Eq. (S24), whose parameters are listed in Table G; quantile selection itself has no trainable parameters. We will state this in the revised Table 3 description.

Table G: Backbone and retention-gate parameter counts

BackboneSynchronous Param.Asynchronous Param.Gate Param.Gate inputsϕz\phi_z
MGN (4 layers)0.57M0.60M0.025Me$\parallel$b in Eq.(S24)MLP (2 layers)

Full Accum. uses z ≡ 1 and retains the delivery mask, so it adds no gate parameters. It outperforms No Buffer on all four datasets, so the buffer benefit is not due to the gate. Learned retention further improves CF, BS, and Burgers, while GS ties at the reported precision (Table H).

Table H: Buffer and retention ablations

Retention variant (Table 5)CFBSGSBurgers
No Buffer0.18450.60480.01890.0100
Full Accum.0.17820.57630.01860.0061
Masked Accum.0.17250.56010.01860.0051

Table I adds Random and Shuffled gates used during both training and evaluation. The learned gate is lower than every random or shuffled alternative on all four datasets.

Table I: Random and Shuffled mask control

Policy during trainingPolicy during evaluationCFBSGSBurgers
Gate by Eq.(S24)Gate by Eq.(S24)0.17250.56010.01860.0051
Gate by Eq.(S24)Random Gate0.18290.59600.01910.0091
Random GateRandom Gate0.18400.60510.02240.0086
Gate by Eq.(S24)Shuffled Gate0.18140.59770.01890.0087
Shuffled GateShuffled Gate0.18300.60200.02060.0073
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer usyo (Part 2)

Q4: Is Q actually used, and in what role, given that Appendix D says it is disabled without observable forcing, Table S4 lists no forcing term, and removing Q changes all four results in Table 2?

Reply: We do not assume knowledge of the governing equations or of any forcing in the learned datasets. Table S4 lists “–” due to unknown forcing or control signal. Q does not depend on how the data were generated, which is why removing Q changes all four results in Table 2. The statement in Appendix D that Q is disabled without observable forcing was misleading; we have revised it in Appendix D.

Concluding remark: Thank you again for your constructive feedback. We hope these clarifications address your concerns and will incorporate the completed results and implementation details into the revision.

Dear Reviewer,

As the rebuttal period is coming to an end, I would like to kindly ask you to read the authors' response and reply as soon as possible.

Thank you for your prompt attention.

Best regards,

Area Chair

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

This paper proposes an asynchronous graph neural network framework for autoregressive dynamics prediction. The method selectively activates edges based on state-dependent priorities, accumulates deferred messages in buffers, and performs a terminal refresh. It also decomposes latent dynamics into five physics-motivated operators—transport, diffusion, reaction, skew-symmetric coupling, and source–sink forcing—with separate scheduling priorities for transport and diffusion. The authors establish theoretical properties concerning permutation equivariance, bounded accumulation, terminal synchronization, and autoregressive error propagation. Experiments are conducted across various scenarios.

Critical strengths

Methodological design

  • Modular asynchronous propagation mechanism. The method combines state-dependent edge priorities, selective retention of deferred increments, and terminal refresh. It can be applied independently of the operator decomposition, making it a potentially useful modification to existing message-passing simulators.



Empirical performance and evaluation

  • Consistently favorable forecasting results. The full model achieves the lowest reported mean RMSE among the compared baselines on all four datasets, covering regular and irregular discretizations and two- and three-dimensional dynamics. 


  • Benefits beyond the full proposed architecture. Table 3 reports improvements when the asynchronous mechanism is added to both MGN and MP-PDE across all four datasets. This supports the usefulness of the generic asynchronous component independently of the five-operator architecture.


  • Substantial component-level analysis. The experiments examine operator decomposition, asynchronous propagation, deferred retention, scheduling policies, and alternative update mechanisms. These are meaningful attempts to separate contributions.
Critical weaknesses

Architectural motivation and novelty

  • The particular five-way decomposition is insufficiently justified. The blocks have reasonable structural motivations, but the paper does not clearly establish why five separately parameterized components are preferable to a simpler decomposition. The relevant gap is the lack of evidence that this particular separation offers advantages beyond general modeling capacity.


  • Positioning relative to related decomposed architectures needs strengthening. The distinction from ADR-GNN[1] and FINN[2] should be made clearer. In particular, the paper should distinguish the contribution of its operator parameterizations from that of coupling those operators with deferred, operator-specific message delivery.

Theoretical analysis

  • The stated results do not establish comparative improvements. Proposition 3.3 bounds buffer contents using bounded increments, while Proposition 3.4 follows directly from terminal refresh. Corollary 3.5 applies to any forecasting architecture satisfying the same Lipschitz and one-step approximation assumptions. Neither KAK_A nor εA\varepsilon_A is derived or compared with baseline constants, so these results do not explain the observed performance advantage.


  • The Lipschitz assumption needs justification for hard scheduling. State-dependent edge selection can change discontinuously at score crossings unless additional conditions control the switching behavior. This does not invalidate the conditional corollary, but it leaves open whether a useful finite Lipschitz constant holds on the intended rollout domain, before considering whether it is smaller than a baseline's.



Ablation design and attribution

  • Table 2 supports the overall operator-based configuration more clearly than five distinct blocks. Individual removals yield modest, similar changes on Cylinder Flow and Black Sea, and negligible changes on Gray–Scott. For example, removing reaction or skew coupling leaves Gray–Scott RMSE at 0.0089, compared with 0.0186 without operator-specific modeling. Thus, the necessity of each block remains unclear, although Burgers shows larger individual effects.


  • Block deletion does not isolate the benefit of physical structure from capacity. Removing a block eliminates both its constraint and its computational pathway. Generic replacements with comparable parameter counts, or a simpler decomposition such as T+D+L\mathcal T+\mathcal D+\mathcal L with one capacity-matched local operator, would test the value of the proposed structure more directly. Such controls are not present in Table 2.



Statistical reliability and reproducibility

  • Uncertainty and implementation details limit interpretation. Tables 2 and 8 lack standard deviations or paired per-seed differences, making the reproducibility of small improvements unclear.

  1. ADR-GNN: Eliasof, M., Haber, E., & Treister, E. (2024). Feature Transportation Improves Graph Neural Networks, AAAI.

  2. FINN: Karlbauer, M., et al. (2022). Composing Partial Differential Equations with Physics-Aware Neural Networks. ICML.

Questions for authors
  • Why are five separate operators needed? Can the full decomposition be compared with transport, diffusion, and one capacity-matched local operator? What demonstrates complementary functionality among reaction, skew coupling, and source–sink heads?


  • What exactly changes in “w/o Ope.” and “w/o P\mathcal P”? Are global refinement, boundary handling, processor depth, and parameter budgets retained? Can refinement and explicit constraint handling be ablated separately?


  • What connects the structural propositions to improved forecasting? Can the authors derive or measure differences in one-step approximation error and perturbation amplification between matched synchronous and asynchronous models? What conditions justify the Lipschitz assumption?
4 comments
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer sxLH (Part 1)

We sincerely thank you for your careful review and constructive suggestions. We appreciate your recognition of the scheduling mechanism and its benefits across existing backbones. Below, we clarify the contribution, the scope of the operator ablations, and the relationship between the theoretical results and forecasting performance.

Q1: Why are five separate operators needed? Can the full decomposition be compared with transport, diffusion, and one capacity-matched local operator? What demonstrates complementary functionality among reaction, skew coupling, and source–sink heads? How does the contribution differ from ADR-GNN and FINN, and which benefits arise from operator parameterization versus deferred delivery?

Reply: Our work proposes delivery Scheduling within a physics-motivated operator framework. The five heads mentioned by the reviewer introduce different physical inductive biases; they are not intended to be indispensable on every task. T and D model spatial interactions along edges, whereas R, S, and Q provide distinct node-local updates. This decomposition supplies structure for learning without assuming that the underlying system contains exactly five independent mechanisms.

Table 2 shows task-dependent marginal effects: removing individual heads has a larger impact on Burgers, smaller effects on CF and BS, and leaves GS at 0.0089 when R or S is removed. We therefore evaluate the operator-structured configuration as a whole rather than claiming that every head is necessary, and we will revise Sec. 4.2 accordingly, replacing the “complementary benefits” sentence with this description.

Relation to ADR-GNN and FINN. ADR-GNN combines advection, diffusion, and reaction in spectral graph updates, following an operator-composition principle related to Eq. (S28); FINN composes physical terms within a finite-volume structure. These structured updates do not include the feature-level, edge-wise buffered delivery mechanism studied here. Our distinction concerns when different edge increments of an operator enter the node update. T and D have separate state-dependent rankings and edge buffers; deferred increments are selectively retained for later activation, with a terminal refresh. This mechanism can be incorporated into the sequential composition in Eq. (S28), the additive update in Algorithm 1 / Eq. (S27), and existing MGN and MP-PDE backbones (Table 3). It makes edge-delivery frequency within an operator an independently adjustable design variable.

Two sets of experiments support the scheduling contribution. Within the full architecture, retaining the same operators, asynchronous delivery outperforms full-edge synchronous delivery on all four tasks (Tables 2 and 10). Adding the mechanism to MGN and MP-PDE also improves forecasting on all four tasks (Table 3). Thus, the scheduling benefit does not depend on the particular five-head instantiation.

To examine the local branches, we keep T, D, scheduling, and input information fixed, replacing only R/S/Q with one capacity-matched local MLP head. Table A compares this structural separation with a generic local mapping of comparable capacity. The comparison will be included in the revision.

Table A: local-operator comparison

Local structureCFBSGSBurgers
R+S+Q (Table 2)0.14260.51910.00890.0044
local MLP head0.15030.53260.01010.0064

Q2: What changes in “w/o Ope.” and “w/o P”? Can refinement and explicit constraints be ablated separately?

Reply: The "w/o P" removes both the learnable refinement G and the explicit constraint operator Π; for example, the CF change from 0.1426 to 0.1579 therefore measures their combined effect. The "w/o Ope." removes T, D, R, S, Q, and P, retaining the MGN backbone and asynchronous scheduler. It is the MGN+Async configuration in Table 3, which explains the identical values. Table B separates G and Π and will be reported in the revision. On BS, only the identity mapping is applied for Π because its boundary conditions are unknown.

Table B: Ablation of G & Π

ConfigurationFive operatorsGΠCFBSGSBurgers
Full (Table 2)RetainedRetainedTask-specific0.14260.51910.00890.0044
w/o Ope.RemovedRemovedRemoved0.17250.56010.01860.0051
w/o PRetainedRemovedRemoved0.15790.57000.00930.0062
w/o G onlyRetainedRemovedRetained0.15380.57000.00900.0050
w/o Π onlyRetainedRetainedRemoved0.15210.51910.00910.0057
AuthorsOpenReview ↗

Rebuttal to the comments by Reviewer sxLH (Part 2)

Q3: How do the structural results explain improved forecasting? Can matched models be compared through one-step error and perturbation amplification? What justifies Lipschitz regularity under discontinuous hard selection?

Reply: Our propositions characterize asynchronous updates; matched experiments establish whether they improve forecasting. Proposition 3.1 recovers the synchronous update when all masks are active. Theorem 3.2 establishes permutation equivariance under conditions including equivariant tie handling. Propositions 3.3–3.4 characterize bounded buffers and terminal clearing. These properties explain how the scheduler handles deferred information.

Specifically, Corollary 3.5 relates rollout error to one-step approximation error and the sensitivity of the forecasting map. It does not establish smaller constants for asynchronous models, as already stated in Remark H.10. Our evidence for improved accuracy comes from the synchronous/asynchronous comparisons in Tables 2, 3, and 10, rather than a claim of universal superiority derived from the corollary.

For selection, fix the graph and random keys and consider a neighborhood with strict gaps at the selection thresholds and continuous score maps. Sufficiently small perturbations then preserve the selection branch. If the maps within that branch are locally Lipschitz and any hard retention gates do not switch, a local bound follows. This argument applies within a fixed-selection neighborhood; behavior across thresholds is addressed by the empirical measurements below. We will state this scope explicitly.

To connect predictive performance with local sensitivity, Table C reports one-step RMSE on identical Burgers test inputs under input perturbations of the same scale, with the selection masks recomputed after perturbation so that switching effects are included. The asynchronous model has a lower one-step error at every perturbation level. Both models are trained with noise injection, which is why the errors change moderately with the perturbation.

Table C: One-step error under input perturbation on Burgers.

Perturbation (standard deviation)Synchronous one-step RMSEAsynchronous one-step RMSE
0(zero perturbation)3.48e-51.52e-5
1e-43.57e-51.627e-5
1e-24.11e-51.832e-5

Q4: Can Tables 2 and 8 include standard deviations?

Reply: Tables D and E report the mean ± standard deviation over three seeds, which will be considered to add to the supplementary material in the revision.

Table D: Data for Table 2.

VariantCFBSGSBurgers
Full0.1426 ± 0.00110.5191 ± 0.00630.0089 ± 0.00020.0044 ± 0.0003
w/o T0.1459 ± 0.001230.5290 ± 0.007800.0090 ± 0.000220.0047 ± 0.00035
w/o D0.1453 ± 0.001260.5297 ± 0.007160.0090 ± 0.000220.0048 ± 0.00029
w/o R0.1457 ± 0.001230.5281 ± 0.005670.0089 ± 0.000180.0048 ± 0.00030
w/o S0.1452 ± 0.001010.5296 ± 0.006910.0089 ± 0.000230.0048 ± 0.00035
w/o Q0.1453 ± 0.001030.5306 ± 0.006320.0090 ± 0.000180.0047 ± 0.00038
w/o P0.1579 ± 0.001320.5700 ± 0.006850.0093 ± 0.000210.0062 ± 0.00040
w/o Ope.0.1725 ± 0.001330.5601 ± 0.007760.0186 ± 0.000420.0051 ± 0.00038
w/o Async.0.1579 ± 0.001370.5477 ± 0.006950.0098 ± 0.000220.0061 ± 0.00045
w/o Both0.1892 ± 0.00170.6147 ± 0.00580.0193 ± 0.00030.0118 ± 0.0013

Table E: Data for Table 8.

MethodCFBSGSBurgers
Synchronous MPNN0.1892 ± 0.00170.6147 ± 0.00580.0193 ± 0.00030.0118 ± 0.0013
Random Edge Update0.1854 ± 0.002220.6065 ± 0.004340.0189 ± 0.000290.0099 ± 0.00094
Node-Batch Async0.1860 ± 0.001690.6076 ± 0.006590.0191 ± 0.000300.0104 ± 0.00137
Delayed Aggregation0.1782 ± 0.001430.5763 ± 0.005000.0186 ± 0.000440.0061 ± 0.00058
Multi-Rate Gating0.1752 ± 0.001990.5663 ± 0.006220.0186 ± 0.000450.0056 ± 0.00068
MGN + Async.0.1725 ± 0.001330.5601 ± 0.006760.0186 ± 0.000420.0051 ± 0.00052

Concluding remark: Thank you again for your constructive feedback. We hope these clarifications address your concerns and will incorporate the completed results and implementation details into the revision.

Thank you for the additional accounts, which address several of my experimental concerns and strengthen the evidence for the structured local branches. The discussion of ADR-GNN and FINN also clarifies the scheduling contribution, whose practical value is supported by in Tables 3, 5, and 6.

My remaining concern primarily relates to the theoretical interpretation. I agree that Propositions 3.3 and 3.4 provide structural guarantees and that Corollary 3.5 is not intended to establish universal superiority. The fixed-selection argument is reasonable as a local justification of Lipschitz continuity. However, it applies while the selection and hard-retention branches remain unchanged; it does not show a uniform bound across switching boundaries or throughout the rollout domain.

Table C provides useful evidence of lower one-step prediction error on Burgers under the tested perturbations, with selection masks recomputed. However, prediction error under input perturbation also incorporates the clean-input approximation error and therefore does not directly measure perturbation amplification. Table C strengthens the accuracy evidence but does not look satisfactorily showing a smaller sensitivity constant.

AuthorsOpenReview ↗

Thank you for your follow-up

Thank you for your follow-up and for recognizing the experimental evidence.

Our paper does not claim a smaller sensitivity constant for asynchronous models, nor does it attribute its accuracy gains to one. This scope is explicit after Corollary 3.5, in Appendix G, and in Remark H.10. Table C provides additional accuracy evidence under input perturbations; it does not establish KA<KSK_A<K_S. The fixed-selection argument is local; our paper does not claim a uniform stability guarantee across switching boundaries.

We would be grateful if you would take this clarification into account in your final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.