acceptodds
Under review as a conference paper at ICLR 2027

ProST: Learning to Progressively Refine Spatio-Temporal Forecasts in a Multi-Scale Forecast Space with Adaptive Scale Selection

Abstract

Spatio-temporal forecasting supports real-world applications such as traffic management and air-quality monitoring. Many existing methods model historical temporal dynamics and spatial dependencies to directly predict the future at the target scale. This approach leaves the relationship between coarse future structure and finer detail implicit, without explicit control over how each is revised during forecasting. We introduce ProST (Progressive Spatio-Temporal Forecasting) to progressively refine predictions of the same future in a joint temporal–spatial scale space. ProST controls coarse-scale revision separately from the addition of finer-scale detail, while adaptive scale selection determines the scales at which refinement proceeds. Each selected scale pair defines an explicit forecast state and a scale-aligned target. Refinement expands the preceding forecast and decomposes a learned correction into a weighted coarse-scale update and zero-aggregate detail. A shared scale-conditioned neural forecaster computes these corrections from historical observations and the preceding forecast. A history-conditioned selector chooses a complete scale-selection sequence before forecasting begins. The refinement satisfies an exact cross-scale identity relating the refined forecast's aggregate to the updated coarse prediction. Experiments show positive mean intermediate refinement gains, and frozen-weight interventions on the temporal parent/detail decomposition increase terminal error when either component is removed across all evaluated seeds and datasets. Our code is available in https://anonymous.4open.science/r/ProST-1B6C/

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal Forecasting, Multi-scale Forecasting, Adaptive Scale Selection, Progressive Forecasting, Graph Learning, Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_CKecOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ProST, which progressively refines predictions of the same future in a joint temporal–spatial forecast space. It decomposes each correction into a weighted coarse-scale update and zero-aggregate detail, executed by a shared scale-conditioned forecaster. A history-conditioned selector chooses a complete refinement sequence before forecasting begins. The paper establishes a cross-scale aggregation identity and evaluates the method on three traffic datasets and one air-quality dataset. Fixed-sequence ProST achieves the lowest mean MAE in seven of eight settings. Training across sequences improves average and worst-sequence errors, while frozen-weight interventions support complementary roles for temporal coarse-scale updates and detail. Adaptive selection provides accuracy–computation trade-offs, but does not consistently outperform input-independent random selection.

Critical strengths
  1. ProST explicitly separates coarse-scale updates from zero-aggregate detail and proves that aggregating a refined forecast yields the preceding forecast plus the weighted coarse-scale update. This makes the effects of the correction components precisely describable. Both refinement steps have positive mean gains on all four datasets, and removing either the temporal coarse-scale update or detail increases terminal MAE in every evaluated seed. This is a principal technical argument for acceptance: the mechanism has both an algebraic definition and functional evidence within the trained model, although the direct experimental evidence should be restricted to the temporal decomposition.

  2. Fixed-sequence ProST achieves the lowest mean MAE in seven of eight settings, including a reported 7.73% improvement over STORM on KnowAir at F=12F=12. Moreover, when evaluated on identical test samples and the same 22 sequences, training across sequences reduces both average and worst-sequence errors on all four datasets without changing per-sequence FLOPs. These results provide the strongest empirical support for acceptance: a shared forecaster can support multiple execution sequences and improve its performance across them through training. This conclusion should nevertheless be interpreted alongside the reported seed variability.

Critical weaknesses
  1. The joint temporal–spatial mechanism is not fully tested.
    The method uses joint temporal–spatial decomposition, but the interventions and numerical checks only test temporal decomposition. This leaves the joint mechanism insufficiently validated. Add spatial-only and joint interventions, and check the joint aggregation identity in the actual implementation.

  2. Cross-scale consistency does not guarantee better predictions.
    Proposition 4.4 proves an aggregation identity, but does not guarantee lower prediction error. The paper also shows that larger local gains can lead to worse final predictions. Clearly state the theorem’s scope and report how coarse-scale error, local gains, and final error change together.

  3. Frozen-weight interventions do not establish the benefit of the design after training.
    Removing components from a trained model shows that the model depends on them. It does not show that alternative models would perform worse after retraining. No direct evidence found in the manuscript for retrained comparisons. Retrain models without the coarse update, detail, or damping, and include a standard additive correction model using the same backbone, path, and training budget.

  4. The benefit of training across paths is mixed with extra training.
    Training across paths adds 100 epochs, but there is no control that continues fixed-path training for the same number of epochs. The improvements also vary across seeds. No direct evidence found in the manuscript for an equal-budget control. Add continued fixed-path training with the same optimizer and training budget, and report results for each seed.

  5. Adaptive selection does not consistently beat simple selection.
    Its gains over random selection are very small on PEMS04, and it performs worse on KnowAir. The comparison at matched cost gives positive gains in only two of three seeds. Compare against fixed and random strategies at the same average cost, using more seeds and paired confidence intervals.

  6. The reported cost does not cover the full system.
    FLOPs exclude feature computation and selector execution. End-to-end latency, memory use, and throughput are not reported, so actual inference speedup remains unverified. No direct evidence found in the manuscript for full-system speedup. Measure latency, memory, and throughput with all selection and forecasting steps included, under the same hardware and batch size.

Questions for authors

See Weakness

5 comments

Thank you for the careful review.

W1: The original interventions tested the temporal split used in the released forecaster. The released model predicts at node level, splits corrections along time, and uses the spatial hierarchy in its features. We clarified this in the Abstract and Introduction. The numerical check of the temporal identity in the released implementation had a maximum residual of 4.77×10−74.77\times10^{-7}.

We added a direct joint check in a separate implementation with explicit spatial forecasts. Within each seed, we fixed the checkpoint, path, and raw second-stage correction. We changed only the split at that stage. All later stages used the joint split. The detail aggregate, reconstruction error, and difference between the two aggregation orders were below 10−510^{-5} in all six runs. The joint split gave lower final MAE than either single-axis split in every run. On PEMS04, the gain over the spatial split alone was small.

DatasetTime split minus joint MAESpace split minus joint MAE
KnowAir+0.0190+0.0330
PEMS04+0.2735+0.0025

We separately tested the contribution of changing both scales along the forecast path. We retrained the full, time-only, and space-only schedules with the coarse weight fixed at 0.5. They shared initialization within each seed, trained for 100 epochs, and used the same validation checkpoint rule. Time-only held the spatial feature scale fixed. Space-only kept all 12 forecast steps at every stage.

DatasetTime-only minus full MAESpace-only minus full MAE
KnowAir+0.0532+0.0823
PEMS04+0.0904+0.0325

The full schedule had the lowest mean MAE on both datasets. Both single-axis schedules had lower MAE on KnowAir seed 3, and space-only had lower MAE on PEMS04 seed 2. Mean RMSE favored time-only on KnowAir and space-only on PEMS04. Space-only also used about 47.6% more profiled forecast FLOPs on both datasets. Appendix E.2, Tables 15–17 report the operator and schedule comparisons.

W2: Proposition 4.4 tells us how much the finer forecast changes the preceding aggregate. It does not guarantee better predictions. For a given raw correction, the coarse weight controls that change and leaves the zero-aggregate detail at full weight.

A smaller coarse update can lower immediate coarse error when the full update moves the aggregate away from its target or overshoots by enough. A large starting error alone tells us little. The direction and size of the update determine where the revised aggregate lands. The next refinement then takes the revised forecast as input and computes another correction. It can preserve, erase, or reverse the immediate advantage.

We compared the separately trained learned-weight and additive models on the same test windows. At 3→63\to6, we grouped windows by the learned model’s coarse MAE using validation cutoffs. Immediate gain is the drop in MAE from each model’s own expanded forecast to its refined forecast at scale 6. The learned model’s mean immediate gain peaked in the middle coarse-error group in all six runs. In the highest coarse-error group, the final comparison changed direction across seeds on both datasets.

DatasetFinal ΔMAE in the highest coarse-error group, seeds 1–3Larger immediate gain and higher final MAE than addition, seeds 1–3
KnowAir−0.0816, +0.1186, −0.4287382/2,333, 617/2,333, 350/2,333
PEMS04+0.2352, −0.0483, −0.03881,027/3,394, 805/3,394, 455/3,394

ΔMAE is learned minus additive. The last column covers all test windows. A larger gain can start from a worse expanded forecast, so even the immediate gain and the refined forecast’s error answer different questions. These paired results show why we evaluate final MAE directly. Appendix E.1, Table 14 reports the results by coarse-error group.

W3: Retaining fine detail helped in every matched retraining pair. Learned coarse attenuation lowered mean MAE on KnowAir and raised it on PEMS04 relative to ordinary addition.

We trained all five rules from scratch for 100 epochs at H=F=12H=F=12. Within each dataset and seed, we matched the backbone, initialization, 3→6→123\to6\to12 path, supervision, loss, and optimizer. We selected checkpoints by validation MAE. The table gives mean-per-batch test MAE averaged over three seeds.

Update ruleKnowAirPEMS04
Ordinary addition13.964918.1183
Learned coarse weight13.936218.1449
Fixed coarse weight 0.513.934818.0975
No coarse update14.414718.1218
No detail14.095718.3071

Learned attenuation beat ordinary addition in two of three KnowAir seeds and one of three PEMS04 seeds. A fixed weight of 0.5 beat addition in two of three seeds on each dataset. Removing detail raised MAE relative to the learned rule in all six pairs. Removing the coarse update raised MAE in every KnowAir seed and lowered it in every PEMS04 seed. Appendix E.1, Table 13 gives the per-seed results.

W4: With the same number of extra optimizer steps, mixed-path continuation lowered both mean and worst-path MAE across the 22 evaluation paths in all six pairs. The predefined path changed little on average. The main gain is in how well the shared forecaster handles alternative paths.

Both arms started from the same checkpoint and used the same data, hierarchy, optimizer settings, and validation rule. Both trained for 100 extra epochs, with 21,900 optimizer steps on KnowAir and 31,900 on PEMS04. The fixed arm used the predefined path on all batches. The mixed arm used it on 60% of batches and sampled a legal path on the other 40%.

Dataset / seedPredefined path ΔMAEMean ΔMAE over 22 pathsWorst-path ΔMAE
KnowAir / 1−0.0010−0.3878−1.0945
KnowAir / 2+0.0003−0.4040−1.2660
KnowAir / 3+0.0369−0.3138−1.4725
PEMS04 / 1−0.0097−4.3022−23.9329
PEMS04 / 2−0.0064−3.5041−19.7317
PEMS04 / 3−0.0180−3.5894−16.0370

Entries are mixed minus fixed test MAE. On PEMS04 seeds 1 and 3, the validation rule selected the starting checkpoint after the fixed arm completed its full budget. This is an equal-step comparison of the two training procedures. Auxiliary losses, updates to individual parameters, and training FLOPs can differ. See Appendix E.5, Table 20.

W5: Evaluated strictly on average accuracy at matched expected FLOPs, history-based selection yields small and inconsistent MAE gains over a validation-defined fixed mixture. However, statistical mixtures only obey computational constraints in expectation across a dataset, leaving deployment systems vulnerable to latency spikes on computationally heavy instances. The structural contribution of our selector is strict, instance-level budget adherence, guaranteeing a hard FLOP ceiling for every single prediction.

To rigorously isolate the effect of assigning paths to particular inputs, we introduced two controls under the same per-window FLOP cap, utilizing the same forecaster and test windows:

  1. Fixed-sequence mixture: Weights matched the selector’s observed mean test forecast FLOPs in expectation.
  2. Shuffled assignments: We uniformly reassigned the selector’s chosen paths across the same test windows, strictly preserving the path use counts and the full forecast FLOP distribution.
DatasetBudgetForecast GFLOPs/windowΔMAE vs. fixed mixture (wins / 5)ΔMAE vs. shuffle (wins / 5)
KnowAirTight4.622−0.00064 (2/5)−0.00488 (2/5)
KnowAirLoose4.653−0.02744 (2/5)−0.03675 (3/5)
PEMS04Tight8.173−0.00967 (4/5)−0.01147 (4/5)
PEMS04Loose8.379−0.00449 (3/5)−0.01001 (4/5)

(ΔMAE is Adaptive minus the comparator. Negative favors Adaptive).

Against the fixed mixture, 18 of 20 paired 95% time-block intervals included zero. Against the shuffle, 16 of 20 intervals included zero (Appendix F.1).

While the selector can learn useful input-path pairings under certain budgets (e.g., lowering MAE by 0.03675 against the shuffle on KnowAir Loose), its universal utility is not an unbounded accuracy boost. Instead, it ensures the operational reliability of dynamic routing without violating strict latency constraints. We have explicitly revised the manuscript to reflect this boundary condition.

W6: We measured the full path from CPU history input to GPU forecast output on an RTX 4090 (seed 1). This encompasses feature computation, transfers, selector execution, grouping, and forecasting. We timed five predefined validation batches per size with ten repetitions, excluding only data loading and final CPU copies. Single-thread tree evaluation substantially reduced selection overhead while preserving all path choices.

ExecutionKnowAir batch 1 (ms)KnowAir batch 32 (ms / samples/s / MiB)PEMS04 batch 1 (ms)PEMS04 batch 32 (ms / samples/s / MiB)
Direct4.4319.84 / 1613 / 1179.64.3932.56 / 983 / 1957.2
Train-best fixed8.9631.85 / 1005 / 1223.98.8553.73 / 596 / 2031.1
Validation fixed, Tight6.7024.17 / 1324 / 1201.76.6341.97 / 762 / 2006.1
Validation-random, Tight4.5323.71 / 1349 / 997.94.4939.03 / 820 / 1351.9
Adaptive–Tight, original74.99113.77 / 281 / 1201.773.93113.17 / 283 / 1981.9
Adaptive–Tight, single thread17.2335.13 / 911 / 1201.716.3350.41 / 635 / 1981.9

Although single-thread evaluation reduces latency, both adaptive settings remain slower than direct prediction. The practical justification for this overhead is that the selector functions as an automated triage mechanism.

By allocating extra refinement only to selected inputs, the Tight budget reliably reduces mean test MAE compared to static direct prediction (e.g., from 18.4956 to 18.3635 on PEMS04). Furthermore, compared to the Train-best fixed sequence at batch size 32 on PEMS04, the single-thread Tight setting is actually faster (50.41 vs. 53.73 ms) while maintaining its strict per-window ceiling. These measured accuracy–latency trade-offs demonstrate that the system actively balances resources. Rather than requiring practitioners to manually grid-search and deploy different fixed-sequence models for varying hardware constraints, this input-dependent routing provides a unified, out-of-the-box solution (Appendix F.3).

Reviewer_rGjVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProST, a spatiotemporal forecasting framework that progressively refines predictions of the same future across temporal and spatial scales. Each refinement decomposes a learned correction into a weighted coarse-scale update and zero-aggregate detail. A shared scale-conditioned forecaster performs these updates, while a history-conditioned selector chooses a complete forecasting sequence before execution. Experiments evaluate forecasting accuracy, intermediate refinement behavior, and accuracy–computation trade-offs.

Critical strengths
  1. The framework explicitly represents intermediate forecasts at different temporal and spatial scales, enabling scale-aligned supervision and subsequent refinement.
  2. The correction decomposition provides a clear algebraic relationship between coarse-scale revision and finer-scale detail.
  3. Experiments cover traffic and air-quality forecasting, with repeated runs, refinement diagnostics, and comparisons across forecasting sequences.
Critical weaknesses
  1. The necessity of explicit correction decomposition is insufficiently established. An ordinary additive residual can already revise the coarse aggregate and introduce finer-scale variation. Since the proposed components are derived from the same predicted residual, the decomposition mainly introduces selective attenuation of its coarse-scale component.
  2. The selector cannot respond to the evolving refinement process. It chooses a complete sequence using historical information before forecasting begins. Consequently, it cannot adjust subsequent scales when intermediate predictions reveal different refinement needs, such as unresolved spatial variation or an ineffective preceding update.
  3. Frozen-weight interventions do not establish superiority over alternative update rules. Removing the coarse update, removing detail, or disabling attenuation after training changes the mechanism to which the model has adapted. The resulting degradation demonstrates reliance on the trained configuration, but does not show that an ordinary residual model trained from scratch would perform worse.
Questions for authors

Under what conditions should attenuating coarse-scale corrections improve forecasting compared with an ordinary additive residual? Can the authors compare both update rules after training from scratch with identical backbones, sequences, supervision, and training budgets? How sensitive is the preselected sequence to intermediate forecasting errors? Would a selector conditioned on the current forecast provide measurable benefits over selecting the complete sequence from history alone? Why are sequence-specific mean losses not restored when ranking predicted residuals? How does this approach compare with absolute-loss prediction, or with adding the means back before selection, under the same computational constraints?

5 comments

Thank you for the careful review.

W1. A finer forecast can capture local variation well and still propose too large a change to the preceding coarse prediction of the same future. Ordinary addition can revise the coarse average and add detail, and it applies the proposed coarse revision in full.

ProST gives the coarse revision its own weight. For the same predicted correction, λ\lambda controls how much the finer forecast changes the preceding average when aggregated back to its scale. The zero-aggregate detail keeps its full weight. This lets the finer forecast add useful local variation even with a smaller revision to the coarse prediction. Setting λ=1\lambda=1 recovers ordinary addition. Q1 explains when a smaller coarse revision can help and why a local gain may not carry through to the final forecast.

W2. We choose the full sequence from history before forecasting so we can compare short and long paths in one decision. We consider each path’s predicted terminal-loss score and total forecast cost. The current forecast still guides the next prediction at the selected scale. The remaining scales stay fixed. We revised Section 4.3 to describe the history-only selection used in Section 5.2. Q3 measures how errors in the first forecast affect final MAE along a fixed path. Q4 tests one local choice after the first forecast. On unperturbed forecasts, the local rule changed some continuations, gave almost the same MAE, and used slightly more forecast FLOPs.

W3. The original interventions kept the trained weights fixed. They measured how much the model relied on the update rule it was trained with. We tested the alternative rules by training each one from scratch with the same backbone, path, supervision, and training budget. Removing detail raised test MAE in all six paired runs. Learned coarse attenuation lowered mean MAE relative to ordinary addition on KnowAir and raised it on PEMS04. Q2 gives the full matched comparison.

Q1. For the same predicted correction, a smaller coarse update can help in two cases. The full update can move the coarse forecast away from its target. It can also overshoot by enough for the smaller update to land closer to the target. The direction and size of the update both matter. A large initial coarse error alone gives little guidance.

An immediate improvement can still lead to higher final MAE. The next refinement takes the revised forecast as input and computes a new correction. That step can preserve, erase, or reverse the earlier advantage.

We tested this relationship on the same windows with separately trained learned-weight and additive models. At 3→63\to6, we used validation cutoffs to group windows by the learned model’s coarse MAE. Final MAE favored different rules across seeds even in the highest coarse-error group. Across the full test set, the learned rule often improved on its own expanded forecast at scale 6 and still finished with higher MAE than ordinary addition.

DatasetFinal ΔMAE in the highest coarse-error group, seeds 1–3Positive immediate gain and higher final MAE than addition, seeds 1–3
KnowAir−0.0816, +0.1186, −0.42871,133/2,333, 1,004/2,333, 1,060/2,333
PEMS04+0.2352, −0.0483, −0.03882,069/3,394, 1,895/3,394, 1,255/3,394

Final ΔMAE is learned minus additive. The last column counts windows across the full test set. See Appendix E.1, Table 14.

Q2. We trained both rules from scratch under matched settings. Learned attenuation lowered mean test MAE from 13.9649 to 13.9362 on KnowAir and raised it from 18.1183 to 18.1449 on PEMS04. It beat ordinary addition in two of three KnowAir seeds and one of three PEMS04 seeds.

Within each dataset and seed, all runs used the same backbone, initialization, 3→6→123\to6\to12 sequence, supervision, loss, and optimizer. We used H=F=12H=F=12 and trained for 100 epochs. We selected checkpoints by validation MAE. Table 13 reports the mean-per-batch test MAE for each seed.

Update ruleKnowAir S1S2S3PEMS04 S1S2S3
Ordinary addition13.895313.918814.080618.037018.169018.1489
Learned coarse weight13.892613.940413.975718.162218.190418.0821
Fixed coarse weight 0.513.856813.924514.023118.090118.144318.0580
No coarse update14.427814.394614.421718.161518.142718.0611
No detail14.038314.151114.097618.337018.271418.3129

A fixed weight of 0.5 beat ordinary addition in two of three seeds on each dataset. Removing detail raised MAE relative to the learned rule in all six pairs. Removing the coarse update raised MAE in every KnowAir seed and lowered it in every PEMS04 seed. These controls support retaining fine detail and show that the effect of the coarse update depends on the dataset.

Q3: Large errors in the first forecast increased final MAE along the fixed path. At magnitude 0.20, all three perturbations raised final MAE in every seed on both datasets. Smaller perturbations gave mixed results on KnowAir.

We used the forecaster trained over multiple paths and kept the path fixed at 3→123\to12. The baseline uses the original first forecast at scale 3, followed directly by scale 12. We added positive bias, negative bias, or zero-mean Gaussian noise to the first forecast. Magnitudes are multiples of the training-set standard deviation of the normalized first forecasts. The table reports final test MAE averaged over windows and three seeds. The main benchmark uses 3→6→123\to6\to12 and averages MAE over batches.

Change to the first forecastMagnitudeKnowAirPEMS04
No added error0.0013.971618.1833
Positive bias0.0514.178918.4883
Positive bias0.1014.523019.3918
Positive bias0.2015.575822.5620
Negative bias0.0513.909118.7975
Negative bias0.1013.993520.1094
Negative bias0.2014.575524.3978
Zero-mean noise0.0513.996518.6169
Zero-mean noise0.1014.068619.6778
Zero-mean noise0.2014.322623.3436

A better continuation could recover part of the loss. For the 0.20 negative bias on PEMS04, we kept the perturbed first forecast fixed. An oracle used the true future to pick the best of three continuations for each window. It lowered final MAE from 24.3978 to 22.7555. The same fixed path scored 18.1833 without added error. This gives the room for recovery in hindsight for that setting. See Appendix G.1, Table 28.

Q4: In our test, one local choice after the first forecast changed some paths, with almost no change in final MAE and slightly higher forecast FLOPs.

We compared with the released history-only Adaptive–Loose selector. Both policies used the same frozen forecaster and started from the same first scale selected by the released policy. The local rule used history and features from that first forecast to choose among feasible continuations. On validation, we required it to change choices and use multiple continuations. We also required its mean forecast FLOPs to stay within 1% of the released policy over all windows and within 2% over eligible windows. We then froze the policy and tested it on unperturbed forecasts.

DatasetTest windowsMAE (released → local)Forecast GFLOPs/window (released → local)Changed choices
KnowAirAll14.72982 → 14.730014.65660 → 4.6585211/6,999
KnowAirEligible21.02425 → 21.032155.33064 → 5.4131811/163
PEMS04All18.31982 → 18.319618.38462 → 8.3948551/10,182
PEMS04Eligible23.13077 → 23.130208.89654 → 8.9242351/3,761

The table pools three seeds. Eligible windows had more than one feasible continuation after the first scale. Every changed choice added a refinement stage. Even on eligible windows, the changes produced little net effect on MAE. Whether a local policy can recover the injected errors in Q3 remains open. See Appendix G.2, Tables 29–30.

Q5: We leave out the means so the score favors paths expected to perform especially well on the current input relative to their own usual error. We subtract each path’s mean training loss from its per-window loss and train the selector to predict that difference from history. The selector ranks the predicted differences plus a forecast-cost penalty.

Adding the mean back gives a path credit for its usual accuracy on every input. The resulting score ranks predicted absolute final MAE plus forecast cost. These two objectives can rank the same paths differently. Q6 tests both choices.

Q6: Both alternatives lowered test MAE at Loose on both datasets. At Tight, they lowered MAE on KnowAir and slightly raised it on PEMS04. Mean test forecast cost increased in every setting.

We compared the released centered score with two alternatives. One adds the training means back to the released predictions. The other trains the same type of predictor directly on absolute MAE. All three used the same frozen forecasters, candidate paths, and per-window FLOP ceilings. On validation, we chose a mixture of cost-penalty policies for each alternative under the released selector’s mean validation cost. We froze the mixtures before testing.

DatasetBudgetReleased MAE / normalized forecast costMean-restored or absolute MAE / cost
KnowAirTight14.7936 / 0.021914.7592 / 0.0296
KnowAirLoose14.7298 / 0.033114.6879 / 0.0371
PEMS04Tight18.3635 / 0.120518.3643 / 0.1230
PEMS04Loose18.3198 / 0.170218.3045 / 0.1775

The table gives three-seed means. The two alternatives gave the same expected MAE and cost to the reported precision. Both were calibrated to the same validation cost target, and both used more forecast FLOPs on test. Which score gives lower MAE at equal mean test FLOPs remains open. See Appendix F.2, Table 24.

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

This paper proposes ProST, a progressive spatio-temporal forecasting framework that refines predictions across temporal and spatial scales. It separates forecast corrections into coarse-scale updates and finer-scale details, and introduces a history-conditioned selector to determine refinement sequences. Experiments on four benchmark datasets demonstrate competitive forecasting performance and examine the effectiveness of the proposed refinement mechanism.

Critical strengths
  1. Meaningful research idea. The paper presents a reasonable framework for progressive spatio-temporal forecasting by separating coarse-scale updates from finer-scale details.

  2. Substantial experimental work. Experiments on four benchmark datasets demonstrate competitive performance, with additional analyses supporting the effectiveness of the proposed framework.

Critical weaknesses
  1. The research motivation needs further justification. The Introduction discusses several existing multi-scale forecasting methods, including Scaleformer, STORM, and NeST, but does not sufficiently analyze their specific limitations, particularly in handling coarse-scale prediction errors and finer-scale details. As a result, the necessity of the proposed method is not adequately justified, and its distinction from existing approaches remains unclear.

  2. The methodological novelty needs further clarification. Section 4.2 decomposes the forecast correction into a coarse-scale update and zero-aggregate detail. However, the cross-scale consistency property largely follows from the aggregation and expansion operations themselves. The authors should provide a more convincing explanation of the advantages of this decomposition over direct residual correction and why it is necessary.

  3. The methodological descriptions are not fully consistent. Section 4.3 discusses selecting the next scale based on the current forecast state, whereas the actual implementation in Section 5.2 determines the entire scale-selection sequence from historical information before forecasting begins. The authors should clarify the distinction and relationship between these two approaches to avoid confusion.

  4. Some of the claimed contributions lack direct validation. The Abstract and Introduction emphasize joint temporal–spatial refinement. However, the key intervention experiments in the appendix mainly focus on the temporal dimension. The current evidence is therefore insufficient to establish the independent contribution of spatial refinement and the specific advantages of joint refinement.

Questions for authors

Overall, I consider this a promising study with a substantial amount of work. However, the manuscript does not sufficiently justify the research motivation or the necessity of its key methodological designs, particularly in terms of how the proposed approach differs from existing multi-scale forecasting methods. If this were a journal submission, I would be inclined to recommend a major revision to give the authors sufficient room for improvement. However, given the limited scope for substantial revisions during the ICLR review process, I currently lean toward a weak reject, unless the authors can provide convincing responses to the four concerns raised above.

4 comments

Thank you for the careful review.

W1: We fully agree with your assessment that our original Introduction did not sufficiently analyze the specific limitations of prior methods in handling coarse-scale errors versus fine-scale details. To clearly justify the necessity of ProST, we have heavily rewritten the Introduction. The revised text now explicitly points out that methods like Scaleformer, STORM, and NeST couple coarse and fine adjustments into a single step. A residual update large enough to correct a biased coarse average will inevitably distort fine-grained local patterns.

We then establish ProST’s mechanistic distinction: mathematical decoupling. Our explicit decomposition (J+λL+D\mathbf{J} + \lambda \mathbf{L} + \mathbf{D}) entirely separates the correction into a learnable coarse-scale update (λL\lambda \mathbf{L}) and a zero-aggregate detail component (D\mathbf{D}). This guarantees that the model can independently correct a biased coarse average without washing out sharp local variance. Please see the updated Introduction.

W2: Proposition 4.4 is indeed a straightforward algebraic identity by design, but its true value is functional rather than mathematical. It solves a structural bottleneck inherent to direct residual corrections.

Mathematically, a direct residual correction forces λ=1\lambda = 1, which permanently binds the coarse and fine updates together. Consequently, if the model must apply a large update to fix a heavily biased coarse average, this identical large update is forced onto the local details, inevitably distorting fine-grained patterns.

Our decomposition breaks this bind by introducing an independent degree of freedom. It allows the model to attenuate the coarse-scale correction independently while keeping the zero-aggregate detail fully intact. To empirically prove this functional necessity, we ran a new sensitivity analysis by retraining the model from scratch across fixed λ\lambda values. We have included these results in Appendix E.7 of the revised manuscript.

λ\lambdaKnowAirPEMS04
0.000.00 (No coarse update)14.4147±0.017714.4147 \pm 0.017718.1618±0.053418.1618 \pm 0.0534
0.250.2513.9795±0.083213.9795 \pm 0.083218.1174±0.033518.1174 \pm 0.0335
0.500.5013.9348±0.0837\textbf{13.9348} \pm 0.083718.1375±0.043618.1375 \pm 0.0436
0.750.7513.9357±0.057613.9357 \pm 0.057618.1076±0.0208\textbf{18.1076} \pm 0.0208
1.001.00 (Direct residual)13.9649±0.100913.9649 \pm 0.100918.1583±0.071118.1583 \pm 0.0711
Learned weight13.9362±0.041713.9362 \pm 0.041718.1449±0.056218.1449 \pm 0.0562

The data exhibits a clear U-shaped trend. The direct residual (λ=1.00\lambda = 1.00) yields suboptimal results because the unattenuated coarse update interferes with local details, whereas completely removing the coarse update (λ=0.00\lambda = 0.00) performs worst since macroscopic biases remain uncorrected. The fact that intermediate attenuation levels (0.50 or 0.75) consistently achieve the best performance confirms our core hypothesis: decoupling the coarse update for independent attenuation is structurally necessary for optimal forecasting accuracy.

Importantly, this ablation highlights the value of our learned adaptive weight. While exhaustively tuning a fixed λ\lambda yields dataset-specific optima (0.50 for KnowAir, 0.75 for PEMS04), our learned weight automatically secures highly competitive accuracy across different data distributions out-of-the-box.

W3: The discrepancy between Section 4.3 and Section 5.2 in our initial manuscript was confusing. As you correctly noted, our implemented system determines the entire sequence upfront from historical information. To fix this inconsistency, we have completely rewritten Section 4.3. We removed the local step-by-step formulation entirely. The updated text now formally defines scale selection as a global, history-conditioned decision made before any forecasting begins. Sections 4.3 and 5.2 are now fully consistent. We appreciate you catching this disconnect.

W4. The original interventions tested the temporal split alone. We added a direct test of what the spatial axis contributes to the correction. This test uses a separate implementation with explicit spatial forecasts. Within each seed, we kept the checkpoint, forecast path, and raw second-stage correction fixed. We changed only how that correction was split. The joint split applies the coarse weight to the average correction over a time interval and a spatial group. It keeps the deviations within that block at full weight. Each updated forecast then enters the final stage with the same joint split.

The table shows the mean reduction in final test MAE over three seeds.

DatasetTime split → joint splitSpace split → joint split
KnowAir0.01900.0330
PEMS040.27350.0025

Both comparisons favored the joint split in every seed on both datasets. The first column measures the gain from including space in the split. The second shows the additional gain over the spatial split alone. On PEMS04, the spatial split alone captured almost all of the gain over the temporal split. The joint split's added benefit therefore depends on the single-axis comparison. These results cover one refinement step with frozen weights.

We also revised the Abstract and Introduction to state what the released model implements. It forecasts at node level, splits corrections along time, and uses the spatial hierarchy in its features. In a separate retraining comparison, we held the temporal path fixed and let the spatial feature scale change by stage. This reduced mean MAE by 0.0532 on KnowAir and 0.0904 on PEMS04. Some seeds and RMSE results favored single-axis schedules. Appendix E.2 reports both experiments.

Thank you for helping us clarify ProST’s departure from existing multi-scale methods.

The revised Introduction now focuses on a specific bottleneck. Standard residual connections apply coarse-scale corrections directly. A massive coarse update inevitably washes out fine-grained local patterns. ProST mathematically decouples this with J+λL+D\mathbf{J}+\lambda \mathbf{L}+\mathbf{D}. The coarse revision gets its own weight λ\lambda. The zero-aggregate detail keeps its full weight. This independent degree of freedom corrects macroscopic biases without destroying local variance.

Our from-scratch retraining supports this mechanism. We retrained the model across fixed λ\lambda values for 100 epochs (Appendix E.7). The performance curve is clearly U-shaped. Unattenuated updates damaged local details. Removing the coarse update left massive biases uncorrected. Furthermore, removing detail degraded MAE in all six matched pairs (Table 13).

We also aligned the selection description in Sections 4.3 and 5.2. The released model splits corrections along time and uses the spatial hierarchy in its features. The revised manuscript states this explicitly. For spatial validation, we implemented explicit spatial forecasts and changed only the second-stage split. The joint operator won in all six runs (Appendix E.2).

We hope these additions address your concerns regarding our motivation and mechanism.

Thank you to the authors for their response. The revised manuscript has improved in terms of clarity and overall readability, addressing some of my concerns. Therefore, I have decided to raise my score accordingly. However, as I still have some limitations in my understanding and assessment of certain technical details, I am unable to increase my confidence in this review at this stage.

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

This paper proposes ProST, a progressive spatio-temporal forecasting framework that refines predictions across temporal and spatial scales. A shared scale-conditioned forecaster predicts corrections, separated into a weighted coarse-scale update and zero-aggregate detail. A history-conditioned selector chooses a complete scale sequence before forecasting. Experiments on three traffic datasets and one air-quality dataset show competitive fixed-sequence accuracy and positive average intermediate gains. Additional experiments examine the correction components and the accuracy-computation trade-offs of scale selection.

Critical strengths
  1. The method description is clear. The definitions of forecast states, scale-aligned targets, and the relationship between coarse updates and finer detail are explicit, making the intermediate forecasts easier to inspect.

  2. The fixed-sequence results are competitive. ProST achieves the lowest mean MAE in seven of eight settings in Table 1 and improves over STORM in all eight settings. The authors report negative refinement cases and acknowledge the limitations of history-conditioned selection, which is convincing and comprehensive.

Critical weaknesses
  1. The history-conditioned scale selector does not show consistent advantage over input-independent selection. Section 6.4 and Table 11 report MAE improvements of only 0.0034/0.0032 on PEMS04 over validation-frequency random selection, while KnowAir results are worse by 0.0749/0.0536. Since adaptive scale selection is a main contribution, the added value of conditioning on history is not fully established.

  2. The current ablations do not clearly isolate the benefit of the proposed joint temporal-spatial parent/detail mechanism. Removing components from a frozen checkpoint shows the trained model uses them, but does not prove whether this design outperforms retrained ordinary residual refinement. Since Scaleformer already uses shared-weight multi-scale refinement, stronger controlled comparisons are needed to prove the contribution of joint refinement and correction decomposition.

Questions for authors
  1. NeST is discussed as closely related but not included in experiments. How does ProST compare with NeST under the same setting?

  2. Can you compare the full model with ordinary residual refinement, temporal-only refinement, and spatial-only refinement? These comparisons would help determine which gains come from the parent/detail mechanism versus the joint scale structure.

  3. The adaptive settings save 32.5–36.2% of forecasting FLOPs relative to Train-best fixed, but MAE increases by about 1% on PEMS04 and 6% on KnowAir. On KnowAir, Adaptive–Tight costs only 1.3% more than direct prediction, which suggests that most of the saving comes from choosing direct prediction rather than from the selector itself?Could you clarify the practical value of these operating points?

3 comments

W1. History helps match paths to windows in some Loose runs. The evidence for an MAE gain is mixed. We shuffled the selector's chosen paths across the same test windows with the forecaster fixed. This preserves each path count and the forecast FLOP distribution. All policies obey the same per-window cap.

At Loose, the selector lowered mean MAE against the expected shuffle by 0.03675 on KnowAir (3/5 seeds) and 0.01001 on PEMS04 (4/5). Two KnowAir 95% time-block intervals favored the selector. All PEMS04 Loose intervals included zero. At Tight, two KnowAir intervals favored shuffling. Tables 22–23 in Appendix F.1 report both budgets.

In Table 11, Tight/Loose changed MAE by −0.0034/−0.0032 on PEMS04 and +0.0749/+0.0536 on KnowAir versus validation-frequency random selection. Mean test costs differed. The separate validation-defined mixture matches mean test FLOPs in expectation. Eighteen of 20 paired 95% time-block intervals within seeds included zero. Both exceptions favored the mixture on KnowAir Tight.

W2. Scaleformer also shares weights across scales. We trained each rule from scratch on the same ProST backbone (H=F=12H=F=12, 100 epochs). Per seed, we matched initialization, path, supervision, optimizer, and validation selection. Addition weights both parts at 1. We learn the coarse weight and keep detail at 1. Test MAE averages batches and three seeds.

RuleKnowAirPEMS04
Learned weight13.936218.1449
Ordinary addition13.964918.1583
No coarse update14.414718.1618
No detail14.095618.3071
Fixed weight 0.513.934818.1375

Removing detail raised MAE in all six pairs. Removing the coarse update hurt every KnowAir seed and helped every PEMS04 seed. The learned rule beat addition in two KnowAir seeds and one PEMS04 seed. A fixed weight of 0.75 also beat addition in mean MAE on both datasets in the sweep (Appendix E.7).

Q1. We did not find official NeST code during our runs. We reimplemented its patch autoregression from the paper. Both used H=F=12H=F=12 and the same test windows, masks, and metric reductions. ProST had lower MAE in all six pairs. Entries are three-seed mean MAE / RMSE.

DatasetProSTNeST (our reimplementation)
KnowAir13.9362 / 20.400114.3709 / 20.7356
PEMS0418.1449 / 28.596019.0339 / 28.9656

Historical inputs, training/validation boundaries, optimizers, and stopping rules differed. NeST ran 51–65 epochs vs 100 for ProST.

Q2. W2 covers ordinary addition. We retrained all three schedules at fixed coarse weight 0.5. Per seed, we matched initialization, 100 epochs, and validation selection. Time-only kept spatial modeling at one scale. Explicit space-only kept 12 steps per stage.

DatasetFullTime-onlySpace-only
KnowAir13.9348 / 7.24313.9880 / 7.28814.0171 / 10.694
PEMS0418.0975 / 12.08718.1879 / 12.16818.1300 / 17.845

Cells give test MAE (three-seed mean) / forecast GFLOPs. We counted supported CPU ops for history encoding and forecasting on one shared validation input per dataset. Both controls had lower MAE on KnowAir seed 3. So did space-only on PEMS04 seed 2. Mean RMSE favored time-only on KnowAir and space-only on PEMS04.

Q3. Direct prediction drove most forecast FLOP savings. Tight/Loose used it on 93.17%/90.28% of KnowAir and 58.24%/42.82% of PEMS04 windows. Its shares of net savings vs Train-best fixed were 95.25%/93.37% and 66.21%/51.61%. KnowAir Tight cut MAE by 0.096 at 1.3% more forecast FLOPs vs direct.

Loose spent extra FLOPs on windows with higher direct MAE. Against always predicting directly, refinement cut MAE on those windows from 28.32 to 26.75 on KnowAir and 23.83 to 23.53 on PEMS04. Windows routed to direct prediction had MAE of 13.45 and 11.37. All are three-seed means.

SettingKnowAirPEMS04
Direct14.8897 / 4.568 / 19.8418.4956 / 7.624 / 32.56
Train-best fixed13.9499 / 7.247 / 31.8518.1746 / 12.091 / 53.73
Tight14.7936 / 4.626 / 113.77→35.1318.3635 / 8.162 / 113.17→50.41
Loose14.7298 / 4.657 / 115.01→34.9418.3198 / 8.385 / 115.18→54.98

Cells give three-seed mean MAE / GFLOPs and median ms. We timed five validation batches ten times each on an RTX 4090 (seed 1, batch 32). Timing spans CPU history to GPU output, including feature computation and selection. Arrows show original→single-thread times. One tree thread preserved validation choices and equivalent forecasts. Both remained slower than direct.

Thank you for the honest and detailed response, and for revising the abstract to reflect the new results. I find the new experiments helpful. In particular, the NeST comparison and the retrained detail, time-only and space-only controls address my questions about the refinement design.

After reading them, I still have some concerns:

  1. The history-conditioned selector shows some small gains in the Loose setting, but compared with shuffled or random selection, the advantage is not yet clear overall.
  2. The learned coarse weight performs similarly to a fixed weight of 0.5.
  3. A large part of the FLOP savings comes from choosing direct prediction, especially on KnowAir.

Since adaptive scale selection is an important part of the paper, I would like to keep my current score for now. I think the progressive refinement idea is interesting, and I hope these results are helpful for future improvement.

Thank you for your constructive engagement throughout this rebuttal phase. We deeply appreciate your recognition that our retrained controls and NeST comparisons successfully address your questions regarding the core progressive refinement design.

We completely understand and respect your remaining perspectives on the adaptive scale selector and the learned coarse weight. In fact, your rigorous scrutiny has fundamentally improved this paper.

To summarize our final stance on your three points for the Area Chair and the reviewers:

  1. The Selector's Marginal Gain & Direct Prediction: We fully agree that the history-conditioned selector's accuracy advantage is small, and its FLOP savings heavily rely on routing simple instances to direct prediction. As promised, we have explicitly downgraded the selector's claims in the revised Abstract and Introduction. We now present it not as a magical accuracy booster, but transparently as an operational triage tool that guarantees per-window computational ceilings.
  2. Learned Weight vs. Fixed 0.5: We view the learned weight performing similarly to the optimal fixed weight (0.5 on KnowAir) as a functional success. It demonstrates that the network can automatically approximate the dataset-specific optimal attenuation out-of-the-box, saving practitioners the massive computational overhead of exhaustively grid-searching fixed λ\lambda values across different datasets.

Your review has helped us precisely define the scope of our contributions. With the adaptive routing now properly framed as a secondary operational feature, we are thrilled that the core scientific contribution stands fully validated by the new empirical controls you prompted.

Thank you again for your time and excellent feedback.

Reviewer_K1bZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes ProST to better model the relationship between coarse-scale structures and fine-scale details in spatiotemporal forecasting. ProST uses a shared forecasting network to progressively refine predictions from coarse to fine scales, decomposing each correction into a coarse-scale update and a zero-aggregate fine-scale detail component. It also introduces a separate scale selector that chooses a forecasting scale sequence based on historical observations to balance accuracy and computational cost. Experiments on several spatiotemporal datasets evaluate the forecasting performance and the effects of progressive refinement and scale selection.

Critical strengths
  1. The coarse-to-fine refinement approach is intuitive. By separating corrections into coarse-scale updates and fine-scale details, the model can adjust previous coarse-scale predictions while preserving fine-scale variations.
  2. The paper provides clear mathematical definitions for temporal and spatial aggregation, forecast expansion, and correction decomposition.
  3. The experiments cover several spatiotemporal datasets and forecasting horizons, with additional analyses such as intermediate gains and parent/detail interventions to evaluate the refinement process.
Critical weaknesses
  1. The paper lacks direct qualitative comparisons of how different scale-selection paths affect fine-grained predictions. It would be useful to show how different valid paths influence temporal patterns, spatial details, and local errors for the same input samples.
  2. The ablation study on training with different scale paths is not sufficiently controlled. A comparison between training with a fixed scale path and training with multiple paths under the same training budget would help isolate the benefit of multi-path training. The additional contribution of the scale selector could also be evaluated more clearly.
  3. The existing intervention experiments focu son temporal correction decomposition, leaving the contribution of joint spatiotemporal decomposition less clear. Moreover, the spatial hierarchy relies on clustering, but the effects of different temporal and spatial scale partitioning strategies are not systematically explored.
Questions for authors
  1. The paper states that λ\lambda depends on the scale transition, but its parameterization is unclear. Is it learnable? Does it depend only on the scale pair, or also on the input sample? Could the authors report its values across different transitions and compare fixed and learnable weights?
  2. Could the authors further investigate the relationship between locally optimal scale selection and globally optimal scale sequences, for example by comparing the final forecasting performance of a Greedy Local Oracle and a Global Sequence Oracle?
  3. ProST mainly focuses on forecasting a single fixed future window. Could it be extended to open-loop autoregressive forecasting over multiple windows? If previous predictions are repeatedly fed back as inputs, would this lead to more severe error accumulation?
5 comments

W1: Appendix D compares four legal paths on three test windows per dataset. We chose the windows before inspecting errors. Per window, the paths share a frozen checkpoint, input history, and target. All paths forecast the same 12 steps at all nodes. The temporal scales set the resolution of that future at each stage. The spatial scales set feature levels.

PathTemporal scalesSpatial feature levels
Direct12NN
Predefined3→6→12s1→s2→Ns_1 \to s_2 \to N
Change time only2→4→12s1→s2→Ns_1 \to s_2 \to N
Change space only3→6→12s1→s3→Ns_1 \to s_3 \to N

NN is the original node level. Relative to the predefined path, we change the first two time scales or only the middle spatial feature level.

In KnowAir case 1, changing only the time scales reduced overprediction after a sharp drop at node 1. At step 7, the target was 40 and the prediction fell from 186.37 to 156.32 (Fig. 16). In PEMS04 case 3, the predefined path lowered node 154 MAE from 1.176 to 0.708 and raised node 307 MAE from 4.211 to 4.336 relative to direct prediction (Fig. 23). Changing only the middle spatial level left the maps nearly identical. Across the six windows, the largest prediction change at a valid node and step was about 0.553 on KnowAir and 0.179 on PEMS04 in the original units (Figs. 18–23). The maps use the same node order and color limits within each dataset.

W2: Mixed training lowered mean and worst-path MAE in all six pairs at equal optimizer steps. Both arms shared the starting checkpoint, training data, optimizer settings, and validation rule. Both ran 100 extra epochs (21,900 steps on KnowAir and 31,900 on PEMS04). Fixed training used the predefined path on every batch. Mixed training used it on 60% of batches and sampled a legal path on the other 40%. We evaluated both on the same 22 paths. The table gives mean test MAE changes over three seeds (mixed minus fixed).

DatasetPredefinedMean over 22Worst of 22
KnowAir+0.0121−0.3686−1.2777
PEMS04−0.0114−3.7985−19.9005

Predefined-path MAE changed little on either dataset. The auxiliary losses differ, and training FLOPs can differ (Appendix E.5).

History-based selection gave small and inconsistent MAE gains at matched expected forecast FLOPs. Both policies used the same forecaster and test windows. We chose fixed paths using validation MAE and profiled FLOPs. Their mixture used the same path probabilities for every input. We set weights from the selector's observed mean test forecast FLOPs. Test targets were used only for scoring.

DatasetTightLoose
KnowAir−0.00064 (2/5)−0.02744 (2/5)
PEMS04−0.00967 (4/5)−0.00449 (3/5)

Entries give five-seed mean selector minus mixture MAE and the number of seeds favoring the selector. Of 20 paired 95% time-block intervals over test windows within seeds, 18 included zero. The other two favored the mixture on KnowAir Tight (Appendix F.1).

W3: Joint decomposition: The joint split lowered final test MAE against both single-axis splits in all six pairs. In a separate implementation with explicit spatial forecasts, the splits shared a frozen checkpoint, path, and raw stage-2 correction per seed. Only the second-stage split changed. The final stage used the joint split. The gains over the time and space splits were 0.0190 and 0.0330 on KnowAir, and 0.2735 and 0.0025 on PEMS04. The PEMS04 gain over the space split was small. The revision clarifies that the released model splits corrections along time and uses spatial hierarchy features (Appendix E.2, Table 15).

Scale and grouping sensitivity: We retrained one learned-weight variant per factor. Each pair used the same initialization, 100 epochs, and validation checkpoint rule. The table shows changes from baseline in three-seed mean test MAE and profiled forecast FLOPs.

ChangeControlsKnowAir ΔMAEPEMS04 ΔMAEΔFLOPs
3→6→12 to 2→4→12Same spatial levels and grouping rule−0.0190−0.0395−9.53%
More groups at all three coarse levelsSame time path and learned grouping rule−0.0031−0.0280+0.07%
Balanced random groupsSame time path, group counts, and nesting−0.0334+0.00560.00%

Coarse group counts went from 10/19/46 to 12/23/62 on KnowAir and from 16/31/77 to 20/39/103 on PEMS04. The new time scales reduced forecast FLOPs and lowered MAE in five of six pairs. Each spatial change lowered MAE in three of six. Balanced random grouping was competitive in these tests (Appendix E.3, Table 18).

Schedule check: We retrained the schedules with a fixed coarse weight of 0.5. Time-only fixed the spatial level. Explicit space-only predicted 12 steps per stage. The full schedule had the lowest mean MAE on both datasets. Single-axis schedules had lower MAE in some seeds. Mean RMSE favored time-only on KnowAir and space-only on PEMS04 (Appendix E.2, Tables 16–17).

Q1. Yes, λ\lambda is learnable. The released model combines embeddings of the source and destination temporal scales with numerical descriptors of the transition. A linear layer and a sigmoid produce the weight. It starts at 0.5. At a given forecast horizon, each temporal transition shares one weight across input samples and spatial scale pairs.

The table shows the ranges across three released Stage II checkpoints per dataset. Appendix E.4 reports all tested transitions.

TransitionKnowAirPEMS04
2→40.5158–0.52760.5733–0.5900
3→60.1428–0.18050.6675–0.7233
4→120.8286–0.92200.9942–0.9971
6→120.9117–0.94150.9967–0.9983

We also retrained fixed and learned weights from scratch for 100 epochs at H=F=12H=F=12 under matched settings. The table gives mean test MAE over three seeds (Appendix E.7).

Coarse weightKnowAirPEMS04
0.00 (no coarse update)14.414718.1618
0.2513.979518.1174
0.5013.934818.1375
0.7513.935718.1076
1.00 (ordinary addition)13.964918.1583
Learned13.936218.1449

The data exhibits a clear U-shaped trend. The direct residual (λ=1.00\lambda = 1.00) yields suboptimal results because the unattenuated coarse update interferes with local details, whereas completely removing the coarse update (λ=0.00\lambda = 0.00) performs worst since macroscopic biases remain uncorrected. The fact that intermediate attenuation levels (0.50 or 0.75) consistently achieve the best performance confirms our core hypothesis: decoupling the coarse update for independent attenuation is structurally necessary for optimal forecasting accuracy.

Importantly, this ablation highlights the value of our learned adaptive weight. While exhaustively tuning a fixed λ\lambda yields dataset-specific optima (0.50 for KnowAir, 0.75 for PEMS04), our learned weight automatically secures highly competitive accuracy across different data distributions out-of-the-box.

Q2. Greedy selection missed the lowest final MAE in over half the windows on both datasets. The greedy oracle maximizes immediate refinement gain. The global oracle minimizes final MAE. Both use test targets, the same scale-3 start, three candidate paths, and a normalized forecast cost ceiling of 1. We used all test windows for seed 1.

DatasetWindows with higher greedy MAEFinal MAE (greedy / global)Mean forecast cost (greedy / global)
KnowAir1,442 / 2,33313.9707 / 13.89090.4381 / 0.7358
PEMS041,826 / 3,39418.1212 / 18.10310.5589 / 0.7709

The mean MAE gaps were 0.0799 and 0.0181. The global oracle achieved lower MAE and used more forecast computation under the common ceiling (Appendix G.3).

Q3. We chained four 12-step forecasts with a frozen seed-1 model and a fixed path. The open-loop run feeds each forecast into the next input window. The reference updates the history with new observations. Each pair uses the same forecast origins (2,297 on KnowAir and 3,358 on PEMS04).

DatasetOpen-loop MAE by 12-step blockMAE with new observations
KnowAir13.9345 / 18.0181 / 19.0850 / 19.443713.9345 / 13.8944 / 13.8785 / 13.8445
PEMS0418.2649 / 19.8678 / 20.8622 / 21.618918.2649 / 18.2875 / 18.2976 / 18.2963

Open-loop MAE rose in every successive block on both datasets. MAE stayed nearly flat with new observations. Appendix G.4 reports cumulative MAE up to 48 steps.

Thank you for requesting the qualitative path comparisons and the strictly controlled training ablation. These additions strengthen the paper significantly.

Appendix D now includes same-input comparisons. These show the distinct local impacts of different paths. Modifying only the temporal path prevented severe overprediction on specific KnowAir instances. The predefined path on PEMS04 showed clear spatial trade-offs between nodes compared to direct prediction.

Your training-budget question led to our strongest baseline: equal-step continuation. We matched the exact starting checkpoint and optimizer steps. Mixed-path training consistently lowered both mean and worst-path MAE across all 22 evaluation paths (Table 20). The predefined path changed little on average. The performance gains stem strictly from the shared forecaster's improved generalization across alternative paths. Extra training volume is not the cause.

We also tested robustness across temporal scales and grouping rules. Finer time scales lowered MAE in five of six pairs and reduced FLOPs. Balanced random groups were also competitive (Appendix E.3).

We hope these controls fully address your questions about path mechanics and the training gains.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Closing Statement: Summary of Revisions and Core Contributions

Dear Area Chair,

As the discussion period concludes, we respectfully ask you to consider the progressive refinement framework on its own proven merits.

Through our rebuttal revisions, we transparently reported that the history-based path selector yields small and inconsistent gains at matched expected FLOPs. Consequently, we have re-centered the manuscript. The primary evidence for the forecaster now rests firmly on experiments with fixed and mixed paths.

As Reviewer rGjV highlighted, ProST makes explicit predictions of the same future at successive scales. Each step revises a forecast supervised at its own scale. This refinement rule lets the model independently control how much the preceding coarse prediction changes while keeping the fine detail at full weight. Relying solely on this mechanism, fixed-path ProST achieved the lowest mean MAE among the evaluated methods in seven of eight benchmark settings (Table 1).

Furthermore, we conducted two rigorous new controls to definitively address reviewer questions about this design:

  1. From-Scratch Ablations: Removing fine detail raised test MAE in all six matched retraining pairs, proving its functional necessity (Appendix E.1, Table 13).

  2. Equal-Budget Training: When fixed-path and mixed-path training were continued from the same checkpoint for the exact same number of optimizer steps, mixed-path training consistently lowered both mean and worst-path MAE across all 22 candidate paths (Appendix E.5, Table 20).

We believe these comprehensive results make a strong case for acceptance. Explicit progressive refinement delivers state-of-the-art forecasts on a fixed path, and mixed-path training endows a single shared forecaster with robust multi-path execution capabilities. Crucially, both of these core scientific results hold entirely independently of the learned selector at test time.

Thank you for your time, guidance, and consideration of this revised evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.