acceptodds
Withdrawn from ICLR 2027

StaLiNO: State-Adaptive Latent Neural Operator for Traffic Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate traffic forecasting supports congestion management and route guidance, and its errors are most costly when congestion forms or dissolves. However, road sensors report flow, which does not reveal the traffic regime: the same flow occurs in free flow and in congestion, where disturbances travel in opposite directions. This raises two challenges: (1) identifying the regime from flow that is dominated by recurring demand, and (2) sharing information within a regime without erasing the differences between regimes. To tackle these challenges, we propose the State-Adaptive Latent Neural Operator (StaLiNO), which forecasts the deviation of flow from a learned daily and weekly base field and shares information through latent traffic states. Regimes are identified by the Reference-Relative Regime Assignment module, which assigns sensors to latent states by their deviation trajectories and fixes these assignments across the input window, so that each state tracks the same sensors. Sharing within regimes is handled by the Regime-Separated Latent Evolution module, which evolves every latent state independently and returns it to the sensors through the same assignments. We further prove that such sharing lowers reconstruction risk exactly when its distortion is below the noise it removes, and that mixing latent states removes noise that does not grow with the number of sensors while adding distortion that does. Experiments on four real-world traffic datasets demonstrate that StaLiNO achieves state-of-the-art accuracy and remains stable when sensors lose their history. Latent evolution improves accuracy with tracked states and degrades it with states regrouped at every step, and the learned states separate free-flow from congested traffic without access to occupancy.

Primary area:
learning on time series and dynamical systems
Keywords:
Traffic Prediction, Spatiotemporal Forecasting, Neural Operator

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3iPNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces StaLiNO, a latent neural operator model for traffic flow prediction. It first constructs a calendar-locked base field using learnable daily and weekly tables, and then models the deviation of traffic flows with respect to this base field. Then, the deviation trajectory of each sensor is encoded once by the routing GRU, and the sensors are soft-assigned to latent states that stay fixed across the whole window. Each field layer first performs the local road update and the pole-residue temporal update, and then performs the weighted pooling of the observed sensors in each latent state. Each latent sequence is evolved separately by a GRU with shared parameters and independent hidden states, without state mixing, and then sent back to each sensor through the same assignment. The final prediction is the sum of the future base field, the most recent valid deviation, and the learned correction.

Critical strengths

Based on the traffic physical fact that flow is two-to-one on the fundamental diagram, and that the characteristic speeds of free flow and congestion have opposite signs, this paper points out that operators that regenerate latent states at every step cannot provide a coherent trajectory for the temporal model. Accordingly, a trajectory assignment fixed across the window is designed. The problem formulation is clear and differs from generic latent neural operators.

Critical weaknesses

W1. The paper claims to achieve state-of-the-art performance, but does not compare with widely used baselines such as STAEformer and PDFormer. The MAE values reported in the original STAEformer paper on PEMS04, PEMS07 and PEMS08 are 18.22, 19.14 and 13.46, which are all lower than StaLiNO's 18.27, 19.63 and 13.81. In Table 1, GWNet is worse than STGCN and DCRNN on all four datasets, with 25.51 on PEMS04, which is far worse than the commonly reported value of about 18.8, and the source of the baseline results and the reproduction settings are not stated. Coupled with the fact that the margins on PEMS07 and PEMS08 are only 0.03 and 0.04 MAE with no standard deviation, the main empirical conclusion of the paper does not hold.

W2. In the PEMS08 component experiment, removing the base field increases MAE by 11.4%, removing the whole latent sharing only increases MAE by 0.75%, and the variant without latent sharing, at 13.917, already outperforms all baselines except STWave. Evidently, the performance mainly comes from the daily and weekly base field, which is similar to the idea of CycleNet and HyperD, and the anchored prediction. At the same time, the LWR derivation requires that disturbances propagate in the direction determined by the state, but PS=SDS−1S⊤P_S = S D_S^{-1} S^\top is a symmetric kernel, and the intra-state average does not distinguish between upstream and downstream. The derivation also requires the base state to satisfy the conservation law, whereas the model uses a calendar-averaged flow table. Neither the latent state mechanism nor the physical motivation of the paper can explain the source of its gain.

Questions for authors

Q1. Both the daily and weekly tables are initialized with the slot means and then added directly. How does the initial base field avoid double-counting the mean? How much improvement can the attention-based refinement of the base field bring over directly using the slot means?

Q2. The pooling kernel PSP_S is symmetric, so where does the model represent the direction-dependent propagation, downstream in free flow and upstream in congestion?

Sug1. The lack of recent spatio-temporal forecasting models is the biggest defect in the experiments of this paper, so it is suggested to supplement them.

Sug2. The occupancy of PEMS04 and PEMS08 should be used to report errors by the periods of congestion formation and dissipation, and a k-means control on the base-field residual trajectories should be added to the state alignment experiment.

Reviewer_9mw3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StaLiNO forecasts traffic flow by grouping sensors according to their recent behaviour. Its motivation is that the same flow can occur in both free-flowing and congested traffic, which follow different dynamics. The model subtracts a learned daily and weekly baseline, groups sensors using their deviation histories, and keeps these assignments fixed within each input window. Each latent group evolves independently before its features are mapped back to sensors. The forecast combines the future baseline, latest valid deviation and a learned correction. A simplified bias–noise analysis explains when grouping reduces noise and when mixing groups adds distortion. On four PeMS datasets, StaLiNO reports the lowest MAE among 15 baselines, with modest gains ranging from 0.15% to 2.3%. Ablations support combining stable group assignments with latent temporal evolution. Further analyses test robustness to missing and perturbed inputs and show that the learned groups align with occupancy-based congestion labels.

Critical strengths

The physical motivation leads to a concrete, testable design. Traffic-regime differences motivate trajectory-based grouping and separate latent evolution. The paper checks the learned groups against occupancy-derived labels that the model never receives, and provides detailed ablations of grouping, evolution and mixing. This helps explain the design’s benefits beyond the headline accuracy.

The robustness results are practically relevant. When 50% of sensors lose their entire history, MAE on the remaining observed sensors rises by only 0.7%. Short interruptions increase MAE by 3.2%, below the compared baselines. These tests provide concrete evidence that the model handles unreliable sensor inputs well.

Critical weaknesses
  1. The comparison misses important baselines. STAEformer and PDFormer are absent from Table 1. STAEformer reports MAE of 18.22, 19.14 and 13.46 on PEMS04/07/08, below StaLiNO’s reported numbers, while also using 12→12 forecasting and 6:2:2 splits. Several included baselines have much weaker results than published implementations, without enough detail to explain why. Consistent reruns and clearer preprocessing and tuning protocols are needed to support the SOTA claim.

  2. The small improvements need uncertainty estimates. StaLiNO beats the best listed baseline by only 0.03 MAE on PEMS07 and 0.04 on PEMS08. Although results average five runs, no standard deviations or confidence intervals are reported. This also affects some mechanistic conclusions based on small ablation gaps. Without per-run variation, it is unclear which rankings are reliable and which could reflect training randomness.

  3. The experiments do not isolate regime identification. The clustering controls use raw flow rather than deviation histories, so they do not show what learned routing adds beyond subtracting the baseline. Removing the base field also changes both the routing input and the forecast anchor, making its 11.4% effect hard to attribute. Despite available occupancy labels, forecasting errors are not broken down by regime or congestion transitions. Moreover, subtracting a calendar baseline does not itself resolve flow ambiguity, so the claim of regime “identifiability” needs qualification.

  4. The no-mixing claim goes beyond the theory and controls. The analysis concerns fixed linear reconstruction under isotropic noise, rather than the learned nonlinear forecaster. The attention ablation tests one near-uniform residual mixing block. Such a block can raise cosine similarity by adding a common component without removing differences between states, so this diagnostic does not establish lost regime information. Gated or zero-initialized attention controls would help distinguish harmful mixing from optimisation or scaling effects.

Questions for authors

Q1. Please report mean ± standard deviation for Table 1, Table 2 and the key (M) settings in Figure 3a, stating the number of independent runs for each. How large are the main differences relative to run-to-run variation? Confidence intervals for the smallest gaps would be helpful.

Q2. Please add STAEformer and PDFormer, ideally rerunning them and a simple baseline such as STID in the same pipeline. Which Table 1 results were reproduced versus taken from prior papers? Do all methods use consistent preprocessing, missing-value handling and metric definitions?

Q3. On PEMS04/08, please report errors separately for free-flow and congested windows, based on the last input step’s occupancy label, and for windows with a regime transition during the forecast horizon. How does StaLiNO compare with strong baselines in these settings, including at 15/30/60 minutes?

Q4. For Table 10, please add k-means on the 12-step deviation window and report the token-sharing variant’s regime agreement. How much do learned trajectory assignments improve AMI and homogeneity over these controls?

Q5. Please change only the routing input from deviations to raw flow, keeping residual inputs elsewhere and the forecast anchor unchanged. How much does MAE change? Also report the anchor-only forecast, (B_{\mathrm{future}}+r_{\mathrm{last}}), without the learned correction.

Q6. Does latent attention still hurt when its output projection is zero-initialized or its residual branch has a gate initialized at zero? Please also report mean-centered cosine similarity or pairwise distances between states, to check whether the 0.44 → 0.96 increase reflects a shared mean component rather than lost state information.

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

This work is motivated by fundamental principles of traffic forecasting: traffic sensors (such as induction loops) measure flow, which serves as an unidentifiable projection of the underlying traffic state. The authors argue that traffic flow alone is insufficient to reliably distinguish between free-flow and congested regimes, as both can exhibit identical flow values ​​while corresponding to different traffic propagation directions. To address this, the authors propose StaLiNO (State-Adaptive Latent Neural Operator for traffic flow forecasting). StaLiNO first learns baseline traffic flow patterns based on daily and weekly periodicities; it then utilizes the deviation of actual flow from these baselines to assign sensors to specific latent states—an assignment that remains fixed throughout the input window. Subsequently, the model evolves each latent state independently over time and maps the predicted information back to the corresponding sensors. Experiments conducted on four PeMS datasets demonstrate that StaLiNO achieves the lowest Mean Absolute Error (MAE) across all benchmarks. The authors further analyze the model’s behavior by examining state assignments, latent state mixing, the identification of congestion regimes, and performance under scenarios with missing sensor data.

Critical strengths

1- The motivation is derived from the Lighthill-Whitham-Richards (LWR) model and the fundamental diagram of traffic flow, rather than merely being an incremental improvement over existing baselines. The observation that flow is an unidentifiable projection of the state that combined with the fact that the characteristic velocities in the two regimes (free-flow vs. congestion) have opposite signs that leads directly to the conclusion that the "spatial kernel" must be state-dependent. The model uses historical trajectories to generate fixed state assignments, ensuring that sensors within the same latent state remain relatively consistent throughout the input window. Table 2 provides valuable insight: while "trajectory assignment" alone performs no better than "token assignment," the combination with independent temporal evolution yields significant improvements. This highlights the effectiveness of the synergy between these two modules, rather than just a simple stacking of network components.

2- Beyond the results on the four PeMS datasets, the authors investigate the correlation between latent states and traffic congestion using occupancy data (which is not included in the model's input). They also analyze the impact of the number of latent states, cross-state mixing, and the presence of missing sensor data.

Critical weaknesses

1- The performance gains appear quite marginal. When the sources of improvement are decomposed, the contribution of the proposed new mechanism is somewhat diminished. A large portion of the model's accuracy stems from the daily/weekly periodic baseline components (a concept shared with models like CycleNet and HyperD).

2- The evaluation is limited exclusively to the PeMS family of datasets. It would be beneficial to see results on other, similar types of traffic datasets.

3-Using 3−M=163-M=16 states for a two-regime (free-flow/congested) model is clearly over-parameterized so it was suspected that this work may be an artifact of hyperparameter tuning.

Questions for authors

refer to weaknesses and these two points.

  1. In Table 1, please include the mean ± standard deviation across five random seeds, and indicate whether the performance gap relative to the strongest baseline is statistically significant (based on a paired significance test).
  2. Please conduct an experiment on the METR-LA dataset (or a similar non-PeMS dataset) to verify the generalizability of the "regime" hypothesis.
2 comments

Dear Reviewer hyaC,

Thank you for reading our paper. We acknowledge that our paper is not yet as solid as we would like, and that our research experience is limited. We appreciate that you recognize the physical motivation from the LWR model and the synergy between trajectory assignment and independent temporal evolution. We will improve the presentation to make the method easier to follow. However, limited experience does not mean that our results are unreliable.

We firmly disagree with the concern that our results may be an artifact of hyperparameter tuning. The review raises this suspicion without pointing to any specific evidence in the paper. We did not tune hyperparameters to obtain the reported results. All hyperparameters, including the number of latent states K, were selected on the validation set, and the test set was used only for the final evaluation.

Our code, configuration files, and trained checkpoints are included in the supplementary material. Every result in Table 1 can be reproduced directly with the released code, for example on AutoDL. We invite the reviewer and the AC to verify our results. If the reviewer finds any specific step in our pipeline that suggests test-set tuning, we would be glad to respond to it directly.

We also note that the review flags our paper for ethics review without stating an ethical concern. Our study uses only the public PeMS traffic flow data released by Caltrans. The data contain no personal information and involve no human subjects. A methodological suspicion is not an ethical issue, and we respectfully ask that the flag be reconsidered. A suspicion of tuning-driven results without supporting evidence is unfair to our work.

We have decided to withdraw this submission. We will use the feedback from all reviewers to improve the paper for a future version. We sincerely thank the reviewer, the AC, and the program committee for their time and effort in reviewing our work.

Sincerely, The Authors

Reviewer_hyaCOpenReview ↗

Reply to the author

Thank you for the clarification and for your response.

I would like to clarify that I have not identified any ethical concerns or evidence of research misconduct in this submission. My previous ethics flag was made out of an abundance of caution regarding the experimental methodology and potential sensitivity to hyperparameter choices. Upon reconsideration, these concerns are methodological rather than ethical, and should not have warranted an ethics flag. I apologize for any confusion this may have caused.

My original assessment was focused solely on the technical contributions, experimental robustness, and generalizability of the proposed method. I appreciate the authors' clarification and their efforts in making the work reproducible, and wish them success in their future research.

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

Related papers

Discussion (0)

Sign in to comment.