acceptodds
Under review as a conference paper at ICLR 2027

EMBER: EVIDENCE-DRIVEN MODAL BOUNDARY EVOLUTION FOR SPATIO-TEMPORAL FORECASTING

Abstract

Spatio-temporal forecasting requires a coordinated transformation from an observed relational field to a future whose spatial organization, regime evolution, temporal memory, and physical scale remain mutually consistent. We introduce EMBER (Evidence-driven Modal Boundary Evolution and Readout), which formulates this task as a coupled boundary-to-future operator. The history and graph first form a coordinated boundary state; this state conditions a node-wise rupture geometry, direct multi-horizon evolution, and a sample-level fractional memory law, while a robust boundary fiber restores local physical units. Our dependency-aligned theoretical analysis establishes stable Galerkin coordination, exact rupture decomposition with a retained orthogonal complement, a quantifiable evidence barrier governing modal reversal and coordinate-wise dynamic regimes, and positive power-law memory with sublinear accumulation and scale-covariant reconstruction. These results show that EMBER preserves boundary information while evolving it through controlled history-dependent dynamics and returning forecasts consistently to local physical units. Mechanism diagnostics further confirm well-conditioned and active Galerkin coordinates together with boundary-dependent fractional orders. Across thirteen graph, grid, and weather tasks, including temperature, humidity, and 12-hour wind-speed forecasting, EMBER achieves the strongest overall performance among the compared methods.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Time Series Forecasting, Dynamical Systems, Multi-Horizon Forecasting, Graph-Based Forecasting, Fractional Memory, Neural Operators

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JopzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EMBER, an operator for graph-indexed spatio-temporal forecasting whose central object is a coupled boundary-to-future map: the observed history and the graph jointly define a "coordinated boundary state", and all forecast horizons are generated directly from that state rather than by autoregressive rollout.

Validation is on six highway datasets (PeMS03/04/07/08, PeMSD7(M)/(L)), two urban-flow datasets (T-Drive, CHIBike, inflow/outflow), and a univariate Weather2K-R cross-domain experiment. The paper also proves two projection/decomposition propositions, a "path dependence and reversal barrier" theorem, a power-law accumulation theorem, and two corollaries, and reports three mechanism diagnostics.

Critical strengths

1. The problem framing is genuinely well-motivated, and one component is a real contribution.

Most spatio-temporal forecasting work assumes the future is a smooth continuation of the recent past, and most STGNNs symmetrize the graph before propagation. This paper takes seriously that (a) relations are directional and signed, (b) observed dynamics can rupture rather than decay, and (c) the same normalized response can correspond to different physical changes at different nodes. The "robust boundary fiber" plus the scale-restoration at readout (Eq. 2 and Eq. 15, with Corollary 2 giving a conditional positive-affine equivariance statement) is the most defensible and most transferable piece of the design: using median/MAD rather than mean/std, and explicitly multiplying back the per-node robust scale, is a principled treatment of the non-stationarity that plagues these benchmarks. This is the part of the paper I would expect others to reuse.

2. Accuracy is competitive across a broad and heterogeneous set of tasks.

EMBER has the best MAE on 10 of 10 traffic tasks and the best RMSE on 6 of 10, spanning sensor graphs and grids, 12-step and 1-step protocols, and speed and flow targets, with a positive cross-domain signal on three meteorological variables. Breadth of this kind is not trivial to achieve, and it is the main reason I am not recommending a clear rejection.

Critical weaknesses

W1. The empirical evidence does not support the paper's causal claims: standard errors are absent, margins are within the baseline spread, and two of the four pillars are never properly ablated.

This is the decisive weakness.

(a) No variance, no seeds. Tables 2–4 report single numbers with no standard deviation, no confidence interval, and no statement of how many runs were averaged. This matters concretely here, because the winning margins are of the same order as the spread among the top baselines. On PeMSD7(M), EMBER's MAE is 2.57 against STGNCDE's 2.68 — a 4.1% gap, while five other baselines (STGODE 2.97, MSTGCN 3.54, ASTGCN 3.61, DCRNN 3.83, STGCN 4.01) are all far worse than that, but the top five methods in that column span only 0.31 MAE. On PeMS03, the gap to TFormer is 3.4%; on PEMS08, 3.4% to STAEformer; on PeMS07, 2.7% to RAST. In the traffic-forecasting literature these gaps are routinely smaller than seed-to-seed variation on identical code. Without std over at least three seeds, it is not possible to tell whether EMBER is better than STAEformer/RAST/PDFormer at all, and every headline claim in the abstract ("strongest overall performance") rests on that.

(b) The paper's own diagnostics undercut two of its four pillars. Figure 2(c) reports the learned fractional order as having "mean 0.094 and range 0.087–0.138" across 3,567 validation windows. A standard deviation of roughly 0.012 on a quantity that is described throughout — in the abstract ("sample-level fractional memory law"), in Section 3.7 ("sample-conditioned fractional order"), and in the conclusion ("long-memory aggregation") — as sample-conditioned is essentially a constant. Worse, for α=0.094\alpha = 0.094 the kernel wℓ(α)=Γ(ℓ+α)/(Γ(α)Γ(ℓ+1))w_\ell(\alpha) = \Gamma(\ell+\alpha)/(\Gamma(\alpha)\Gamma(\ell+1)) gives w1=1w_1 = 1, w2≈0.047w_2 \approx 0.047, w3≈0.017w_3 \approx 0.017, so the readout places roughly 78% of its total weight on the single most recent step (total mass ≈1.29\approx 1.29 at h=12h=12), versus 8.3% under uniform weighting. The learned operator is therefore close to a one-step readout with a thin tail, not a "power-law long-memory aggregation". Note also that α\alpha sits near the hard floor αmin⁡\alpha_{\min} imposed by the parameterization in Eq. 23, which suggests the model is actively pushing the fractional mechanism toward the identity. The same issue applies to the "Galerkin chart": Figure 2(b) shows the retained subspace captures a projected energy ratio of only ~0.44–0.47, i.e. less than half the latent variation, which the text nevertheless describes as "substantial". And Figure 2(a) is circular as a diagnostic: Proposition 1 proves the whitened Gram matrix equals IrI_r exactly, so the reported "after" condition number of 1.000011.00001 is guaranteed by construction and is not evidence of anything.

Crucially, no ablation isolates these claims. "w/o Fractional" is vacuous on T-Drive Inflow (the authors' own footnote: H=1H=1 makes the operator an identity), so the fractional readout is ablated on exactly one dataset, where it moves MAE by 0.22. There is no variant with a single learned scalar α\alpha, and no variant with linear (uniform) accumulation — precisely the comparisons that would determine whether the "fractional" story is doing work.

(c) Direct multi-horizon decoding is never ablated, despite being Challenge II in the introduction and one of the four contribution bullets. Every competing method in Tables 2–3 also forecasts all horizons, so the paper's central contrast against autoregressive decoding is never tested against an autoregressive version of EMBER itself.

(d) The ablations confound mechanism removal with capacity and input removal. "w/o Galerkin" deletes an entire residual block (Eq. 22), not just the whitening or the low-rank projection; the reported degradation of 0.65 MAE therefore mixes "the Galerkin mechanism matters" with "one fewer block matters". "w/o Boundary Fiber" removes an input pathway, changing the input dimensionality. Both need capacity-matched replacements (e.g., a standard MLP block of equal width/depth; raw windowed features instead of fiber statistics).

W2. The theoretical results are properties of the chosen algebraic form, not statements about learning; they cannot adjudicate the design choices, and one of them is quantitatively vacuous at the horizons tested.

Propositions 1–2 and Theorem 1–2 are proved correctly — I checked the proofs in Appendix B and the unrolling and Gamma-ratio arguments are fine. But their content is definitional:

  • Prop. 1 says an orthogonal projector is an orthogonal projector with spectral norm 1 and that a full-column-rank whitening gives Gram II and condition number 1. True, and true of any whitening.
  • Prop. 2 says an orthonormal frame plus its orthogonal complement reconstructs a vector losslessly. This is the Pythagorean theorem, applied to a frame the model was constrained to construct via modified Gram–Schmidt (Eq. 8). The "exactness" is enforced by the parameterization, not learned.
  • Theorem 1's reversal barrier mrev≥⌈∣Sh∣/u⌉+1m_{\mathrm{rev}} \ge \lceil |S_h|/u\rceil + 1 is a statement about the accumulator χ\chi, and the proof shows only that two different preceding sums yield two different growth values — which is immediate from χ\chi being a sum. The bound follows from Sh+m≥Sh−muS_{h+m} \ge S_h - mu and monotonicity of tanh. It quantifies the arithmetic of a running mean, not any behavior of the trained network.
  • Theorem 2 states that generalized-binomial weights decay and that their partial sums are O(hα)O(h^\alpha). This is the standard asymptotic of the fractional-difference kernel; it is a property of the kernel the authors chose, not a discovered property of the model. Moreover, with the measured α≈0.09\alpha \approx 0.09 and h≤12h \le 12 (and h=1h=1 for the urban tasks, where the mechanism is the identity by the paper's own footnote), hα≈1.26h^\alpha \approx 1.26; the "sublinear accumulation" is barely distinguishable from linear over the horizons actually evaluated.

The abstract promises that the analysis "establishes stable Galerkin coordination, exact rupture decomposition, a quantifiable evidence barrier ..., and positive power-law memory". In each case the property is built into the functional form by construction. None of the four results distinguishes EMBER from a simpler model with the same constraints, and none connects to the experiments. Together with W1(b), the theory and the diagnostics are largely restating the architecture back to the reader.

W3. Specific result overclaiming, including one contradiction with the authors' own table.

Section 5.4 states that "Tables 2 and 3 show consistent predictive gains" and the abstract claims "the strongest overall performance among the compared methods". However, in Table 3 EMBER is not best on three of four urban-flow RMSE columns:

SettingEMBER RMSEBest baseline RMSE
T-Drive Outflow30.7830.61 (PDFormer)
CHIBike Inflow5.345.31 (PDFormer)
CHIBike Outflow5.295.27 (STAEformer)

The text carefully says "lowest MAE on all four" — which is true — but the surrounding claim of consistent gains, and the RMSE framing in PEMS03 ("the 2.46-point reduction ... is particularly clear", which is an RMSE figure positioned immediately after a list of MAE figures), create an impression of uniform superiority that the table does not support. On the six highway datasets, the text also says EMBER "obtains the lowest MAE and RMSE in every dataset", while PEMS04 MAE is a 18.06/18.06 tie with STAEformer (acknowledged in the same sentence, then contradicted by it).

W4. Baseline comparability is unspecified, and the cross-domain experiment uses a materially weaker baseline pool.

The paper does not state whether baseline numbers were re-run under EMBER's splits and preprocessing or copied from the original papers. This is not a pedantic point for these benchmarks: PeMS03/04/07/08 have documented inconsistencies in split ratios and evaluation masks across the literature, and TFormer/PDFormer/STAEformer numbers differ by several MAE points depending on protocol. Without a per-baseline statement, the margins in W1(a) cannot be attributed to the method.

Relatedly, the Weather2K-R comparison (Appendix D, Table 6) uses DCRNN, STGCN, MSTGCN, TGCN, and CLCLR — none of the strong baselines from Table 2 (STAEformer, PDFormer, RAST, STGNCDE, MFormer). The claimed "cross-domain transfer" is therefore confounded with a weaker comparison set, and the univariate, distance-adjacency setup is considerably easier than the multivariate traffic setting it is being used to support.

Questions for authors
  1. Variance. How many seeds were run for Tables 2–4? Please report mean ± std over at least three seeds, or state explicitly that the numbers are single runs. Given the 2.7–4.2% MAE margins and the spread among top baselines, this determines whether the results support any of the paper's comparative claims.
  2. Baseline protocol. For each baseline in Tables 2, 3, and 6, were numbers re-run under your splits, preprocessing, and mask, or copied from the original papers? Please add a table stating this per method and per dataset (especially split ratios for PeMS03/04/07/08).
  3. Table 3 RMSE. EMBER's RMSE is second or third on T-Drive Outflow, CHIBike Inflow, and CHIBike Outflow. How do you reconcile this with "consistent predictive gains" in §5.4 and "strongest overall performance" in the abstract?
  4. Fractional order. Figure 2(c) gives α\alpha a mean of 0.094 and a std of about 0.012, near the parameterization floor. What does the boundary-conditioned α\alpha buy over a single learned scalar (or over uniform linear accumulation)? Please provide both ablations, on a dataset with H>1H > 1.
  5. Direct multi-horizon decoding. Please provide an ablation in which EMBER's readout is replaced by autoregressive rollout. This is Challenge II in the introduction and is never tested.
  6. Rupture validity. No diagnostic shows the rupture/hysteresis mechanism triggering at real events. Can you show, on PEMS08 or T-Drive, that the accumulated charge and the modal growth coefficient change sign or magnitude at congestion onset and clearance, as the motivation in §1 predicts? A synthetic system with known regime switches would also serve.
  7. Weather baselines. Why is the Weather2K-R comparison restricted to four older models plus CLCLR, omitting every strong method used in Table 2?
  8. Capacity-matched ablations. Can "w/o Galerkin" and "w/o Boundary Fiber" be replaced with capacity-matched and input-matched controls (a standard residual block of equal width; raw windowed features instead of fiber statistics)?
  9. Please provide D,r,K,LD, r, K, L, learning rate, batch size, epochs, parameter counts, and training/inference wall-clock relative to STAEformer on PEMS08, and release code.
1 comment

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

Reviewer_4ZKBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper propose EMBER, a spotio temporal forecaster built as a sequence of explicitly parameterized stages.

  1. Relational coordinates : Fixed signed node coordinates come from a truncated SVD of two hop directed affinity matrix
  2. Boundart fiber : Each node gets median (or mean) absolute deviation normalization + slope and curvature features
  3. Whitened Galerkin blocks : Build a sample conditioned basis from hidden states + relational coordinates, orthonormalized by QR and apply low rank projection
  4. Rupture stage : Each node terminal state is split exactly into K modeal coordinates and orthogonal residual. Modal coordinates evolve across horizons through a diagonal affine recdurrence whose growth rate is tanh of 1/sqrt(h) scaled cumulative sum of evidence

Experiments cover six PeMS datasets, T-Drive and CHIBike in/outflow, and three Weather2K-R variables. EMBER reports the best or tied-best MAE on all tasks. There is a four-component ablation on PEMS08 and T-Drive, and mechanism diagnostics on PEMS08 covering Gram conditioning, projected energy, and the learned alpha

Critical strengths

Coherent, fully specified foward pass with non_autorefressive multi horizon generation. The architecture is described precisely enough to reimplement the forward computation. The appendix gives the exact parameterizations of every gate, bound, and scale. The design choice of anchoring a sample adaptive, QR whitened low rank basis to fixed signed SVD coordinates of the directed graph is a reasonable and somewhat distinctive alternative to row-stochastic diffusion and to free adaptive adjacency. The Galerkin ablation suggests it carries real signal: it is the largest drop on both ablated datasets.

Critical weaknesses

I'm not sure about empirical advantage at the level of rigor the margins require. The difference reported in results are small. For example, PEMS08 gain 0.47 MAE compare to baseline.

Also, detail is kind of missing. Paper isn't analyzing where the baseline come from such as settings. Additionally, multiople seeds may need in this task. Two baselines the paper itself discusses, Graph WaveNet and AGCRN, are absent from the tables. Per-horizon results (3/6/12 steps) are not reported either, even though multi-horizon evolution is the paper’s central claim.

Questions for authors
  1. Where do the baseline numbers in Tables 2, 3, and 6 come from? If they were rerun, under what protocol, hyperparameter search, and number of seeds? Can you report the original published numbers for STAEformer and PDFormer alongside yours, and give mean ± std over at least 3–5 seeds for EMBER and the strongest baselines?
  2. Please give all architectural and training hyperparameters, plus runtime, memory, and parameter counts relative to STAEformer and PDFormer. Per-sample QR and per-node MGS may be costly on PeMS07 and PeMSD7(L).
1 comment

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

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

The paper proposes EMBER, a non-autoregressive spatio-temporal forecasting architecture that treats prediction as a coupled boundary-to-future operator. From an observed history and a directed graph, EMBER constructs a coordinated boundary state using a sample-conditioned, whitened Galerkin chart; evolves this boundary via a rupture–residual decomposition with hysteretic evidence accumulation; and performs a boundary-relative, sample-conditioned fractional readout that restores node-specific physical scale. The authors provide aligned theoretical guarantees for stability and path dependence, diagnostic analyses of internal mechanisms, and strong empirical results on 10 public spatio-temporal benchmarks in the main text.

Critical strengths

·The formulation of forecasting as a single boundary-to-future operator is conceptually fresh, explicitly tying relational coordination (Galerkin whitening), regime-sensitive latent dynamics (rupture with hysteresis), and physical-scale readout (fractional, boundary-relative accumulation) through a common boundary state. ·The sample-conditioned whitened Galerkin basis that mixes graph-anchored signed coordinates with current sample features is a principled alternative to diffusion- or attention-only spatial coordination, with a clear energy-preserving projection property. ·The rupture–residual decomposition with exact, orthogonal initialization and bounded, hysteretic modal growth is a well-motivated structural mechanism for regime persistence, decay, and reversal without error-compounding autoregression. ·The fractional readout with sample-level order α learned from the boundary mean, coupled with robust per-node scale restoration, is a neat way to achieve path-dependent accumulation while keeping outputs covariant to local physical units. ·The traffic results support the architecture across two forecasting settings. Table 2 reports the best or joint-best MAE on all six highway tasks. On the four single-step urban-flow tasks, EMBER has the lowest MAE throughout and the lowest RMSE on T-Drive Inflow.

Critical weaknesses

·Computational overhead may be significant: per-block QR whitening over N×r, per-node modified Gram-Schmidt for K rupture modes, and repeated sample-conditioning could strain scalability for very large N, T, or H. The paper lacks time/memory profiling and big-O analysis. ·The reported wind-speed gain depends on an unresolved protocol difference. All five baseline MAE/RMSE pairs in Table 6 match, at the reported precision, the 12-hour column of Wang et al., Table 2 . For example, DCRNN gives 1.39/1.91 and CLCRN gives 1.26/1.74 in both. Appendix D describes 1,866 Weather2K-R stations, hourly sampling and univariate inputs. Wang et al. use 25 stations, sample every three hours and report the fourth forecast step with five input features. The source and evaluation protocol of each row, including EMBER’s, determine whether the reported improvement over CLCRN is comparable. ·Missing baselines include several competitive or recent direct-forecasting and non-autoregressive approaches (e.g., Graph WaveNet, AGCRN in the tables; spectral-loss methods like FreST as a training plug-in; operator-style or horizon-conditioned models). Given the claims about direct multi-horizon evolution and operator framing, such comparisons would strengthen the case. ·Operator-learning and non-autoregressive spatio-temporal forecasting beyond FNO (e.g., spatio-temporal operator decoders like STONe) and horizon-conditioned emulators (e.g., recent multi-horizon GNNs for physical systems) are not empirically compared though conceptually close. ·The ablation leaves the benefit of accumulating evidence unresolved. Removing the full rupture/hysteresis module raises MAE from 13.40 to 13.82 on PEMS08. It also raises MAE from 16.65 to 17.05 on T-Drive Inflow, where a one-step forecast has no earlier forecast step to accumulate. Both results support retaining the module. The single-step result shows that its other operations can already improve accuracy. On PEMS08, Table 4 does not separate their contribution from the effect of carrying evidence across steps. ·Figure 1 has serious formatting and legibility problems. For example, an arrow in column 3 overlaps the growth-rate symbol gih, and column 4 displays x_Ti where the main text uses xTi. These and similar formatting issues make it difficult to trace the decoder’s computation from the paper’s main architecture figure. The authors should redraw Figure 1 with larger labels, clear separation between arrows and symbols, and mathematical notation consistent with the main text. ·The paper adopts specialized terminology (“boundary fiber”, “rupture geometry”) that, while internally consistent, can be opaque and could benefit from more intuitive bridging to standard ST-forecasting vocabulary.

Questions for authors

Q1 Can you provide training/inference time and GPU memory usage versus top baselines on at least one small (PeMS08) and one large (PeMSD7(L)) dataset? How do r (Galerkin rank) and K (rupture modes) scale performance and cost? Q2 Which wind-speed results share an evaluation protocol? For EMBER and the five baselines, please identify the source of each score, the stations, input variables, sampling interval and test split. If the rows use different protocols, could you report a matched comparison or revise the wind-speed claim? Q3 How much does carrying evidence across forecast steps help? On PEMS08, could you keep the modal frame, additive forcing, growth bound and training setup fixed, but let the gate in Eq. (10) use only the current step’s evidence? Comparing MAE with the full model on the same seeds would isolate the effect that Table 4 leaves open. A separate comparison with a single learned fractional order would test whether adapting that order to each sample adds value. Q4 How were the other baseline results selected? For PEMS04, Table 2 lists RAST RMSE as 30.55, while its published Table 1 gives 29.93 . The latter is below EMBER’s 29.98. The PEMS07 values also differ (32.97 versus 32.73). Was RAST rerun, and if so which split and metric aggregation were used? For weather, please explain whether the benchmark’s MFMGCN baseline can be compared under matched graph inputs. Q5 Please clarify how Figure 1 was prepared and checked, including any AI assistance, and redraw the affected arrows and subscripts at a legible size.

1 comment

Dear Reviewer,

Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.

We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.

Thank you for your time and consideration.

Best regards,
The Authors

Comments on the paper

1 comment

Dear Area Chair and Program Chairs,

Thank you very much for your time and efforts in coordinating the review process.

As the discussion period progresses, we would sincerely appreciate your assistance in encouraging the reviewers to engage with our rebuttal and provide further feedback. We have carefully addressed the reviewers' concerns and would welcome the opportunity to clarify any remaining questions through constructive discussion.

We would be particularly grateful if the reviewers could reconsider their initial assessments in light of our responses and, where appropriate, update their scores accordingly.

Thank you again for your consideration and support in facilitating a fair and constructive discussion process.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.