acceptodds
Under review as a conference paper at ICLR 2027

EviPhys: Evidence-Aware Multimodal Representation Learning with Physics-Guided Graph Modeling

Abstract

Learning reliable representations from structured spatio-temporal data is crucial for traffic flow prediction. However, three major challenges remain: effectively leveraging structural and semantic information from traffic data, improving the robustness of traffic representation under noisy and uncertain observations, and incorporating physical knowledge to enhance prediction reliability. To address these challenges, this paper proposes **EviPhys**, an **Evi**dence-aware multimodal representation learning framework with **Phys**ics-guided graph modeling. First, the Multimodal Representation Fusion Module (MRFM) converts historical traffic measurements into a three-channel image that encodes traffic level, temporal variation, and sensor-axis variation, and aligns the image with a state-aware text prompt through a frozen Vision-Language Model (VLM) to obtain macroscopic traffic semantics. Second, the Dempster-Shafer Evidence Reasoning Module (DSERM) maps each sensor's historical window to a Basic Probability Assignment (BPA) and uses Deng entropy to generate reliability-aware microscopic representations. Third, the Evidence Semantic Graph Learning Module (ESGLM) fuses the macroscopic semantics and microscopic evidence to construct a directed propagation graph. Fourth, the Physics-Guided Prediction Module (PGPM) recursively predicts traffic flow increments and regularizes the predicted flow sequence with a graph-diffusion prior defined on the physical sensor graph, encouraging topology-consistent flow evolution and stable long-horizon prediction. Experiments on four PEMS benchmarks and two large-scale LargeST datasets demonstrate the effectiveness and scalability of EviPhys, achieving average reductions of 2.60%, 1.08%, and 2.17% in MAE, RMSE, and MAPE compared with the second-best results on the PEMS benchmarks. On the SD and GBA datasets, EviPhys achieves the best performance across all reported horizons and metrics, with average reductions of 9.57%, 7.73%, and 11.57% in MAE, RMSE, and MAPE on SD, and 7.80%, 4.41%, and 21.47% on GBA, respectively. Code is available at https://anonymous.4open.science/r/EviPhys-CC32.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal representation learning, Physics-informed neural networks, Dempster-Shafer evidence theory, Multimodal fusion, Graph learning

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_12RFOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EviPhys combines a frozen ViLT-based multimodal representation module, a Dempster–Shafer evidence reasoning module, dynamic graph learning, and a graph-recurrent forecasting model regularized by a graph-diffusion prior. The paper reports results on four PEMS datasets and two LargeST datasets (SD and GBA), with improvements over the listed baselines. Its main claims concern multimodal traffic semantics, uncertainty-aware local representations, adaptive spatial dependencies, and more stable long-horizon prediction.

Critical strengths
  1. Promising reported predictive performance. EviPhys reports competitive results across four PEMS datasets and two larger LargeST datasets. For example, it reports an MAE of 12.99 on PEMS08, compared with 13.56 for the best listed baseline.
  2. Reasonable motivation for free-running trajectory regularization. Applying a graph-diffusion residual to recursively generated trajectories is a plausible way to encourage consistency with a chosen road-graph prior during multi-step prediction. The regularizer is removed at inference, which may limit its direct inference overhead.
Critical weaknesses
  1. Core implementation does not match the stated Deng-entropy formulation. The paper defines Deng entropy with a denominator that depends on the cardinality of each focal set. In the submitted code, compute_deng_entropy uses a constant denominator of 7. When the masses sum to one and clipping is inactive, this computes Shannon entropy plus ln(7), rather than the stated Deng entropy. The implementation also applies an eight-channel Softmax without an explicit mapping that excludes the empty set, despite the text specifying zero mass for the empty set. This mismatch undermines the claim that the reported results validate the proposed Deng-entropy mechanism.
  2. The statistical significance results are not readily reconcilable with the reported summary statistics. For PEMS04, Table 8 reports 18.36 ± 0.091 for PDFormer and 17.95 ± 0.072 for EviPhys, with a paired-test p-value of 0.008 over 10 runs. Under the conventional interpretation of these standard deviations and a two-sided paired t-test, the bound s_d ≤ s_1 + s_2 implies |t| ≥ 7.954 (df = 9), corresponding to a p-value far below 0.008. The authors should provide per-seed results, the pairing procedure, and the exact test code. This is a reporting inconsistency that requires clarification; it is not, by itself, evidence of misconduct.
  3. The claimed stability guarantee is conditional and is not enforced by the parameterization. The appendix’s non-expansiveness argument requires 0 ≤ κΔt ≤ 1. The submitted implementation uses Softplus for the diffusion coefficient, which enforces positivity but does not enforce the required upper bound. This does not invalidate the conditional theorem, but the paper should report whether the trained model satisfies the condition and avoid presenting the theorem as an unconditional guarantee for the implemented model.
  4. Key ablations and robustness comparisons are missing. The main ablation study removes the VLM, evidence, and dynamic-graph components, but does not isolate the effect of the graph-diffusion regularizer itself. The robustness table reports EviPhys under different noise levels without baseline or no-evidence comparisons. As a result, the experiments do not establish how much benefit comes specifically from Deng-entropy gating or the physical regularizer.
  5. The image construction and scalability claims need stronger validation. The sensor-axis difference used to construct the image may depend on the arbitrary ordering of sensors unless that ordering is tied to road topology. The authors should test sensitivity to consistent sensor permutations. In addition, Table 3 reports out-of-memory configurations, which does not invalidate the LargeST results but does qualify the scalability claim; peak memory and throughput across graph sizes would
Questions for authors
  1. Implementation and main results. Which exact code version produced the reported main results? Please clarify the focal-set mapping, explicitly show how the empty set is excluded, and provide results from the implementation that matches the paper’s Deng-entropy equation.
  2. Statistical testing. Please provide the per-seed MAEs for Table 8, the pairing rule, and the statistical test script, and explain the discrepancy between the reported summary statistics and p-values.
  3. Physical prior and robustness. How is κΔt ≤ 1 guaranteed in the trained model? Please provide ablations that separately remove the topology-guided correction and graph-diffusion regularizer, as well as noise evaluations against strong baselines and a model without the evidence branch.
  4. Reproducibility and sensor ordering. Please provide the configurations and evaluation protocol used for SD and GBA, and clarify how sensors are ordered when constructing the traffic images. A permutation-sensitivity test would help determine whether the sensor-axis image channel captures meaningful spatial structure.
12 comments
AuthorsOpenReview ↗

Thank you for your insightful comments. We clarify the distinction between Deng-entropy-inspired evidence modeling and strict Deng entropy, verify ten-seed statistical significance, explain the conditional stability and role of graph-diffusion regularization, and address the physical interpretation and scalability of Hovmöller representations.

W1/Q1. Deng entropy motivates our uncertainty-aware local representation learning, but DSERM does not implement strict focal-set-wise Deng entropy. Instead, it maps each sensor's history into eight learnable latent evidence channels and applies Softmax to obtain a normalized distribution. These channels do not explicitly correspond to the eight power-set subsets of a three-element frame of discernment, nor is empty-set mass explicitly set to zero. We use a fixed factor K−1=7K-1=7 to construct a Deng-entropy-inspired surrogate, equivalent to Shannon entropy plus log⁡7\log 7 when clipping is inactive. This surrogate measures evidence-channel dispersion and adaptively regulates local representations for dynamic graph construction through a learnable gate, rather than distinguishing composite propositions of different cardinalities. We have clarified these distinctions, the actual formulation, and its relationship to Shannon entropy in Section 3.3. All reported results follow this implementation, ensuring consistency between the experiments and the implemented method.

W2/Q2. The 10 independent runs assess forecasting consistency across random initializations rather than relying on a single favorable result. Means and standard deviations quantify performance variability, while significance tests evaluate the reliability of observed differences. We acknowledge that the previously reported PEMS04 p-value is inconsistent with the paired t-test result recalculated from the per-seed observations. We have rechecked the per-seed results, pairing procedure, and statistical calculations. Corrected statistics and per-seed results are presented in Tables R3-1 and R3-2 of the second comment, respectively.

W3/Q3. The graph-diffusion regularizer constrains recursive predictions using physical road topology rather than enforcing exact diffusion dynamics. Appendix A shows that when 0≤κΔt≤10\leq\kappa\Delta t\leq1 and Rt=0R_t=0, the graph-diffusion transition is non-expansive; for nonzero residuals, deviations from ideal diffusion are bounded by accumulated residuals. In our implementation, Softplus ensures κ>0\kappa>0, but κΔt≤1\kappa\Delta t\leq1 is not explicitly enforced. Thus, the theorem establishes conditional stability rather than an unconditional guarantee for the trained model. Nevertheless, penalizing graph-diffusion residuals remains a physically motivated objective encouraging topology-consistent evolution during free-running prediction, thereby discouraging deviations from topology-guided traffic propagation. We have clarified this distinction and added a λpinn=0\lambda_{\mathrm{pinn}}=0 ablation to isolate the physics loss contribution, with results presented in Table R3-3 of the third comment.

W4/Q3. EviPhys incorporates physical information through two complementary mechanisms. Topology-guided correction directly adjusts predicted traffic flow increments using state differences between physically connected sensors, while graph-diffusion regularization penalizes topology-inconsistent trajectories during training without additional residual computation at inference. Together, they balance nonlinear forecasting with physics-inspired spatial consistency. Existing baseline comparisons demonstrate competitive forecasting accuracy, while noise perturbation experiments assess stability under increasing input disturbances. These complementary evaluations support predictive effectiveness and noise tolerance. To strengthen the evidence, we have added a λpinn=0\lambda_{\mathrm{pinn}}=0 ablation and noise robustness comparisons against representative baselines and EviPhys without evidence modeling. The results are presented in Tables R3-3 and R3-4 of the third comment.

W5/Q4. Hovmöller images encode spatiotemporal traffic variations for visual representation learning rather than treating sensor ordering as physical spatial coordinates. The three channels represent traffic flow states, temporal variations, and sensor-dimensional variations. Temporal variations capture traffic dynamics, while sensor-dimensional variations reflect adjacent observations under a given sensor ordering, not physical spatial gradients. Actual road topology is explicitly incorporated through subsequent graph modeling. Thus, Hovmöller images complement rather than replace topology-based spatial modeling. To evaluate scalability, we further assess EviPhys on SD and GBA, containing 716 and 2,352 sensors, respectively. Since dynamic graph construction involves pairwise node relationships, computational and memory costs increase with network size. We therefore do not claim linear complexity or unrestricted scalability. The results support effectiveness at the evaluated network scales without implying size-independent computational costs. Additional results on trainable parameters, inference time, and peak GPU memory usage on SD and GBA are presented in Table R3-5 of the third comment.

AuthorsOpenReview ↗

Thank you for identifying the statistical inconsistencies. We recheck the ten-seed results, correct the paired t-test calculations, and provide updated significance results with per-seed records.

Table R3-1. Statistical Significance Across Ten Random Seeds

DatasetPDFormer MAE (Mean ± Std)EviPhys MAE (Mean ± Std)Two-sided Paired p-value
PEMS0314.9398 ± 0.08305514.3100 ± 0.0607536.29 × 10⁻⁸
PEMS0418.3601 ± 0.09103317.9496 ± 0.0720779.84 × 10⁻⁷
PEMS0719.9701 ± 0.11202018.6598 ± 0.0860056.19 × 10⁻¹⁰
PEMS0813.5800 ± 0.07395512.9900 ± 0.0590358.85 × 10⁻⁹

Note: Each dataset contains ten paired observations using random seeds 42–51. Standard deviations are calculated with ddof=1, and p-values are obtained using two-sided paired t-tests with 9 degrees of freedom.

Analysis: EviPhys consistently achieves lower mean MAE and smaller standard deviations than PDFormer across all four PEMS datasets. All corrected p-values are below 0.001, providing strong statistical evidence for the observed performance differences. These results demonstrate that the improvements are consistent across different random initializations rather than being driven by a single favorable run.

Table R3-2. Per-Seed MAE Results for PDFormer and EviPhys

To ensure reproducibility and clarify the relationship between the reported single-run and multi-run results, we provide the individual MAE values obtained using ten random seeds.

PEMS03 and PEMS04

SeedPEMS03 PDFormerPEMS03 EviPhysPEMS04 PDFormerPEMS04 EviPhys
4214.93814.31118.36117.947
4314.86314.30518.37718.059
4414.93014.30018.43617.919
4514.94914.38318.46717.958
4614.81814.37518.26518.040
4714.98414.23518.20517.838
4814.84614.38818.33717.980
4914.94814.23018.28417.881
5015.08314.33318.37717.997
5115.03914.24018.49217.877

PEMS07 and PEMS08

SeedPEMS07 PDFormerPEMS07 EviPhysPEMS08 PDFormerPEMS08 EviPhys
4219.97018.65813.58212.990
4319.90418.70913.55312.962
4419.73718.71513.62612.999
4519.85918.57913.66813.053
4620.13518.63613.60712.954
4720.04918.56413.49012.979
4820.02018.57613.71412.941
4920.03818.77013.53713.038
5020.00418.58813.52913.093
5119.98518.80313.49412.891

Note: All MAE values are reported to three decimal places. PDFormer and EviPhys results are aligned by identical random seed indices.

Analysis: EviPhys achieves lower MAE than PDFormer for every reported seed across all four datasets, indicating that its performance advantage is not limited to a particular random initialization. The per-seed results also provide the underlying observations required to reproduce the reported means, standard deviations, and statistical significance tests.

Pairing Rule and Reproducible Test Script

For each dataset, we pair the PDFormer and EviPhys MAEs obtained using the same random seed, from seed 42 through seed 51. We then apply a two-sided paired t-test to the ten within-seed differences. This ensures that the statistical comparison follows a consistent pairing rule.

import numpy as np
from scipy.stats import ttest_rel

# Values are ordered by seeds 42, 43, ..., 51.
scores = {
    "PEMS03": {
        "pd": [14.938, 14.863, 14.930, 14.949, 14.818,
               14.984, 14.846, 14.948, 15.083, 15.039],
        "evi": [14.311, 14.305, 14.300, 14.383, 14.375,
                14.235, 14.388, 14.230, 14.333, 14.240],
    },
    "PEMS04": {
        "pd": [18.361, 18.377, 18.436, 18.467, 18.265,
               18.205, 18.337, 18.284, 18.377, 18.492],
        "evi": [17.947, 18.059, 17.919, 17.958, 18.040,
                17.838, 17.980, 17.881, 17.997, 17.877],
    },
    "PEMS07": {
        "pd": [19.970, 19.904, 19.737, 19.859, 20.135,
               20.049, 20.020, 20.038, 20.004, 19.985],
        "evi": [18.658, 18.709, 18.715, 18.579, 18.636,
                18.564, 18.576, 18.770, 18.588, 18.803],
    },
    "PEMS08": {
        "pd": [13.582, 13.553, 13.626, 13.668, 13.607,
               13.490, 13.714, 13.537, 13.529, 13.494],
        "evi": [12.990, 12.962, 12.999, 13.053, 12.954,
                12.979, 12.941, 13.038, 13.093, 12.891],
    },
}

for dataset, values in scores.items():
    pd = np.asarray(values["pd"], dtype=float)
    evi = np.asarray(values["evi"], dtype=float)

    result = ttest_rel(evi, pd, alternative="two-sided")

    print(
        dataset,
        f"PDFormer={pd.mean():.6f} ± {pd.std(ddof=1):.6f}",
        f"EviPhys={evi.mean():.6f} ± {evi.std(ddof=1):.6f}",
        f"t(9)={result.statistic:.6f}",
        f"p={result.pvalue:.12g}",
    )

Statistical Clarification: The reported means and sample standard deviations can be reproduced directly from the per-seed observations in Table R3-2.

AuthorsOpenReview ↗

Thank you for the valuable suggestions. We add physics-loss ablations, noise robustness comparisons, and large-scale computational efficiency evaluations.

To further strengthen our responses to W3–W5 and Q3–Q4, we conducted additional experiments covering three aspects: (1) physics regularization ablation across six datasets to isolate the contribution of the graph-diffusion loss; (2) comparative noise robustness evaluation to assess prediction stability and the contribution of evidence modeling; and (3) computational efficiency analysis on SD and GBA to examine practical scalability. The detailed results and analyses are presented in Tables R3-3, R3-4, and R3-5, respectively.

Table R3-3. Ablation Study of Physics Regularization

To quantify the independent contribution of physics regularization, we compare the full EviPhys model with a variant using λpinn=0\lambda_{\mathrm{pinn}}=0 across six datasets, while retaining the topology-guided correction mechanism.

DatasetFull EviPhysλpinn=0\lambda_{\mathrm{pinn}}=0ΔMAE
PEMS0314.3114.43+0.12
PEMS0417.9518.12+0.17
PEMS0718.6618.89+0.23
PEMS0812.9913.10+0.11
SD16.6316.86+0.23
GBA19.0319.31+0.28

Note: ΔMAE denotes the increase in MAE after removing physics regularization.

Analysis: Removing physics regularization consistently increases MAE across all six datasets, with increases ranging from 0.11 to 0.28. These results demonstrate that graph-diffusion regularization provides additional predictive benefits beyond the topology-guided correction retained in the model. The consistent improvements across PEMS, SD, and GBA further support its effectiveness across different traffic network scales.

Table R3-4. Noise Robustness Comparison Under Different Perturbation Levels

To further evaluate noise robustness, we compare EviPhys with HimNet (2024) and EviPhys without the evidence modeling branch under four noise levels on PEMS04.

Noise LevelModelMAERMSEMAPE (%)
0%HimNet (2024)18.1729.9212.03
EviPhys w/o Evidence18.2730.1812.14
EviPhys17.9529.8711.91
5%HimNet (2024)18.5730.4812.31
EviPhys w/o Evidence18.6230.6412.37
EviPhys18.2130.2412.08
10%HimNet (2024)19.1231.2812.70
EviPhys w/o Evidence19.0831.2112.73
EviPhys18.5630.7112.31
20%HimNet (2024)20.0532.6813.49
EviPhys w/o Evidence19.8332.4413.35
EviPhys19.0431.5212.74

Note: All experiments are conducted on PEMS04.

Analysis: EviPhys consistently achieves the lowest MAE, RMSE, and MAPE across all four noise levels. At 20% noise, it reduces MAE by 5.04% compared with HimNet and 3.98% compared with EviPhys w/o Evidence. Moreover, its MAE increases by only 6.07% from 0% to 20% noise, compared with 10.35% for HimNet and 8.54% for EviPhys w/o Evidence. These results demonstrate stronger noise tolerance under the evaluated perturbations and suggest that evidence modeling contributes to maintaining forecasting accuracy under noisy conditions.

Table R3-5. Computational Efficiency and Scalability on Large-Scale Datasets

To further evaluate computational efficiency and scalability, we report the number of trainable parameters, inference time, and peak GPU memory usage of EviPhys on the SD and GBA datasets.

DatasetNodesTrainable ParamsInference Time (s)Peak GPU Memory (MiB)
SD716883,7440.0381529.6
GBA2,352909,9200.0341772.4

Note: Inference time is measured with a batch size of 1. Peak GPU memory usage is reported in MiB.

Analysis: EviPhys maintains a relatively stable parameter count across the two datasets. As the number of sensors increases from 716 to 2,352, trainable parameters increase by only 2.96%, while peak GPU memory usage increases from 529.6 MiB to 772.4 MiB. The measured inference times are 0.0381 s and 0.0341 s, respectively. These results demonstrate the practical feasibility of EviPhys on the evaluated large-scale traffic networks. Nevertheless, dynamic graph construction involves pairwise node relationships, and its computational and memory requirements depend on network size. Therefore, these findings support scalability within the evaluated settings without implying linear complexity or unrestricted scalability.

Reviewer_5ze8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EviPhys combines vision-language representations, evidential modeling, dynamic graph learning, and graph-diffusion regularization for traffic flow forecasting. It is evaluated on four PEMS benchmarks and two LargeST datasets, achieving competitive results against the reported baselines.

Critical strengths

The framework has a clear overall structure and combines global traffic context with local sensor representations. Results are strong across multiple networks, particularly on LargeST. Component ablations and efficiency comparisons provide useful supporting evidence.

Critical weaknesses

My concerns are mainly on the justification of module selection and experiment results - see my questions below.

Questions for authors
  • The method introduces a lengthy pipeline involving a pretrained VLM, evidence modeling, dynamic graph learning, and recurrent decoding for traffic flow forecasting. Given the need for timely predictions and resource-efficient deployment, what justifies this level of complexity? The paper reports improved accuracy but does not sufficiently establish whether these gains outweigh the additional computational, memory, and deployment burdens, or under what practical conditions this approach would be preferable to simpler forecasting models.

  • The overly complex pipeline requires extensive ablations to validate each component. The paper includes branch-removal studies and comparisons between physical priors but lacks controlled comparisons with simpler feature encoders and with no physical regularization or ordinary smoothing. The necessity of these complex designs therefore remains unclear.

  • All six datasets concern California traffic flow forecasting, limiting geographic and task diversity. The main text also does not separately evaluate unusual conditions such as accidents or road closures. Could evidence gating or diffusion regularization suppress abrupt but valid changes?

  • LargeST shows substantial improvements but includes fewer baselines than PEMS, omitting methods such as HimNet and PDG2Seq that perform strongly in Table 1. Were these exclusions due to adaptation difficulties or computational constraints? Please explain the selection and, where feasible, include representative stronger or scalable baselines.

8 comments
AuthorsOpenReview ↗

Thank you for your constructive comments. We clarify the distinct roles and theoretical motivations of EviPhys modules, their ability to preserve abrupt traffic variations, the computational costs of ViLT-based multimodal learning, and the selection of competitive baselines for large-scale traffic forecasting.

Q1. The various modules of EviPhys are designed to address different modeling challenges rather than merely increasing the complexity of the model. MRFM captures global traffic patterns. DSERM characterizes local uncertainty using learnable evidence channels and a Deng-entropy-inspired surrogate. ESGLM integrates global and local evidence to learn dynamic spatial dependencies. PGPM incorporates road topology constraints to improve recursive forecasting. Together, these components form a complementary and interpretable framework. Moreover, the ViLT backbone remains frozen, and graph-diffusion regularization is applied only during training, reducing additional computational overhead. As reported in Section 5, EviPhys contains only 877K trainable parameters and achieves inference times comparable to existing models on PEMS04, while Appendix B further analyzes its computational complexity. We acknowledge that lightweight models may be preferable in resource-constrained scenarios. To further evaluate practical deployment costs, we have added an analysis of GPU memory usage on PEMS04, with the results presented in Table R2-1 of the next comment.

Q2. Each component is motivated by a clear theoretical rationale. MRFM jointly encodes three-channel traffic images and statistical text prompts to capture temporal variations, inter-sensor differences, and global traffic conditions that cannot be fully represented by two statistical descriptors alone. DSERM characterizes uncertainty through learnable evidence-channel distributions and entropy-inspired gating rather than relying solely on deterministic features. PGPM employs a graph-diffusion residual that couples temporal evolution with physical road connectivity, distinguishing it from generic smoothing constraints. Appendix A further provides theoretical justification through conditional non-expansiveness and residual-based deviation bounds. The ablation studies in Section 4.4 show that removing the VLM, evidential modeling, or dynamic graph learning degrades forecasting performance. Section 4.5 and Appendix H further demonstrate that graph diffusion generally outperforms Lorenz-, wave-, and Burgers-based physical priors. We agree with the reviewer and have therefore added controlled comparisons with an MLP-based statistical feature encoder and without physics regularization. The experimental results are presented in Table R2-2 of the next comment.

Q3. Although all six datasets originate from California, they cover diverse traffic network structures and scales, ranging from 170 to 2,352 sensors, enabling evaluation across different network configurations. Importantly, EviPhys does not inherently enforce smooth predictions. DSERM adaptively regulates local evidence through learned uncertainty-aware gating rather than directly suppressing abrupt variations. PGPM retains a data-driven increment branch to capture nonlinear traffic changes, while graph-diffusion regularization serves as a soft training constraint rather than a mandatory diffusion transition. Furthermore, Appendix F demonstrates that EviPhys effectively tracks substantial traffic fluctuations, while Appendix I confirms its robustness under Gaussian observation noise. Together, these existing results provide complementary empirical evidence of the model's adaptability to varying traffic conditions and its ability to preserve meaningful temporal variations without imposing excessive smoothing.

Q4. The LargeST experiments aim to evaluate forecasting performance on larger-scale traffic networks. The existing baselines cover diverse modeling paradigms, including graph-based forecasting, recurrent neural networks, scalable MLP architectures, and recent retrieval-augmented methods such as RAST. They also span different publication years, including recent studies from 2026, ensuring methodological diversity and research relevance. Given the large scale of LargeST and its substantial computational demands, we initially selected representative baselines within the available experimental time and resources. This selection was intended to balance methodological coverage and computational feasibility, rather than to avoid comparisons with competitive methods. Nevertheless, we agree with the reviewer and have added comparisons with HimNet and PDG2Seq under consistent data preprocessing, dataset splits, and evaluation metrics. The results are presented in Table R2-3 of the next comment.

AuthorsOpenReview ↗

Thank you for your valuable suggestions. We provide additional experiments on computational efficiency, simplified model variants, and competitive LargeST baselines.

To strengthen the empirical evaluation of EviPhys, we added three analyses: (1) computational efficiency and GPU memory usage on PEMS04; (2) ablations assessing multimodal learning and physics regularization; and (3) comparisons with HimNet and PDG2Seq on SD and GBA across forecasting horizons. Results are presented in Tables R2-1–R2-3.

Table R2-1. Computational Efficiency and GPU Memory Usage on PEMS04

To further evaluate the practical deployment cost of EviPhys, we report its parameter count, end-to-end inference time, and GPU memory usage on PEMS04.

ConfigurationTrainable ParamsFrozen ParamsTotal ParamsInference TimePeak GPU Memory
EviPhys + ViLT0.877M111.595M112.472M18.4 s1,246.5 MiB
EviPhys + MLP0.927M0M0.927M4.9 s812.5 MiB

Note: Measurements are conducted on an NVIDIA H100 GPU using the PEMS04 test set (batch size 32). Inference time covers the full test set, including input preprocessing.

Analysis: Although ViLT remains frozen during training, it introduces additional storage and inference costs. The complete EviPhys model contains 112.472M parameters, of which only 0.877M are trainable. Its end-to-end inference time is 18.4 s, compared with 4.9 s for the lightweight MLP alternative. These results clarify the computational trade-off associated with multimodal representation learning. As shown in Table R2-2, the ViLT-based configuration consistently achieves higher forecasting accuracy, supporting its use when predictive performance is prioritized.

Table R2-2. Ablation Study with Simpler Alternatives

To further evaluate the necessity of multimodal representation learning and physics-guided regularization, we compare the full EviPhys model with an MLP-based statistical encoder and a variant without physics regularization across six datasets.

DatasetFull EviPhysMLP EncoderNo Physics
PEMS0314.3114.7014.43
PEMS0417.9518.3818.12
PEMS0718.6618.8918.89
PEMS0812.9913.1713.10
SD16.6317.0416.86
GBA19.0319.4419.31

Note: MLP Encoder replaces the ViLT-based representation with a two-statistic MLP, while No Physics sets λpinn=0\lambda_{\mathrm{pinn}}=0, retaining the topology-guided decoder correction.

Analysis: Replacing ViLT with the two-statistic MLP consistently increases MAE by 0.18–0.43 (1.23%–2.73%) across the six datasets, demonstrating the additional predictive value of multimodal representation learning. Similarly, removing physics regularization increases MAE by 0.11–0.28, confirming that graph-diffusion regularization contributes beyond the topology-guided correction retained in the decoder. These consistent improvements across PEMS, SD, and GBA support the complementary effectiveness of the proposed components across different traffic network scales.

Table R2-3. Additional Comparison with HimNet and PDG2Seq on LargeST

To strengthen the evaluation on large-scale traffic networks, we additionally compare EviPhys with HimNet and PDG2Seq on SD and GBA under consistent data preprocessing, dataset splits, and evaluation metrics. We report MAE, RMSE, and MAPE at forecasting horizons 3, 6, and 12, together with the average performance.

DatasetModelHorizon 3 (MAE / RMSE / MAPE)Horizon 6 (MAE / RMSE / MAPE)Horizon 12 (MAE / RMSE / MAPE)Average (MAE / RMSE / MAPE)
SDHimNet15.30 / 26.72 / 10.22%18.28 / 31.58 / 12.30%22.94 / 40.11 / 16.30%18.20 / 31.64 / 12.49%
SDPDG2Seq15.13 / 26.29 / 10.06%18.07 / 31.07 / 12.11%22.68 / 39.46 / 16.05%18.00 / 31.14 / 12.29%
SDEviPhys14.21 / 24.07 / 8.78%16.56 / 28.74 / 10.62%20.55 / 36.90 / 13.86%16.63 / 29.49 / 10.78%
GBAHimNet17.30 / 29.71 / 13.84%20.81 / 35.05 / 18.03%25.72 / 43.02 / 24.14%20.64 / 34.88 / 18.01%
GBAPDG2Seq17.11 / 29.24 / 13.63%20.58 / 34.49 / 17.75%25.43 / 42.33 / 23.77%20.40 / 34.33 / 17.73%
GBAEviPhys16.36 / 28.40 / 10.71%19.16 / 32.88 / 12.98%22.92 / 38.93 / 16.60%19.03 / 32.95 / 13.06%

Note: All models are evaluated using the same forecasting horizons and evaluation metrics.

Analysis: EviPhys consistently outperforms both HimNet and PDG2Seq across all forecasting horizons and evaluation metrics on SD and GBA. Compared with the stronger newly added baseline, PDG2Seq, EviPhys reduces average MAE by 7.61% on SD and 6.72% on GBA, while reducing average RMSE by 5.30% and 4.02%, respectively. More notably, the average MAPE decreases by 12.29% on SD and 26.34% on GBA. These improvements demonstrate that EviPhys maintains strong predictive performance across short-, medium-, and long-term forecasting horizons, further validating its scalability on large-scale traffic networks.

AuthorsOpenReview ↗

A Summary of Our Earlier Responses on Model Design, Computational Efficiency, and Large-Scale Forecasting Validation

We sincerely thank the reviewer for the constructive comments regarding the model design, theoretical motivations, computational efficiency, and experimental validation of EviPhys. We would like to briefly summarize the clarifications and additional experimental evidence presented in our earlier responses.

First, we further clarified the distinct roles and theoretical motivations of the proposed modules. MRFM captures global traffic patterns through multimodal representations, DSERM characterizes local evidential uncertainty, ESGLM learns dynamic spatial dependencies by integrating global and local information, and PGPM incorporates road topology and graph-diffusion regularization into recursive forecasting. These components address complementary modeling challenges rather than merely increasing model complexity. We also explained the theoretical justification for graph-diffusion regularization, including conditional non-expansiveness and residual-based deviation bounds.

Second, we conducted additional controlled experiments to evaluate the necessity of multimodal learning and physics regularization. Replacing the ViLT-based representation with a lightweight statistical-feature MLP increases MAE by 0.18–0.43 across six datasets, while removing physics regularization increases MAE by 0.11–0.28. These results provide further evidence that both multimodal representation learning and physics-guided regularization contribute to forecasting performance beyond simpler alternatives.

Third, we extended the computational evaluation by reporting trainable and frozen parameters, end-to-end inference time, and peak GPU memory usage on PEMS04. Although EviPhys contains only 877K trainable parameters, we acknowledge the additional inference and memory costs introduced by the frozen ViLT backbone. The comparison with a lightweight MLP alternative provides a more transparent assessment of the trade-off between predictive performance and computational efficiency.

Fourth, we clarified that EviPhys does not inherently impose excessive smoothing on traffic predictions. DSERM adaptively regulates local evidence, while PGPM retains a data-driven increment branch to capture abrupt temporal variations. Graph-diffusion regularization acts as a soft training constraint rather than a mandatory diffusion transition. The existing fluctuation-tracking and noise-robustness experiments further support the model's adaptability to varying traffic conditions.

Finally, we strengthened the LargeST evaluation by incorporating HimNet and PDG2Seq as additional competitive baselines on SD and GBA. Under consistent preprocessing, dataset splits, and evaluation metrics, EviPhys achieves better performance across all evaluated forecasting horizons and metrics. Compared with PDG2Seq, EviPhys reduces average MAE by 7.61% on SD and 6.72% on GBA, providing additional evidence of its effectiveness on large-scale traffic networks.

We hope these additional explanations and experiments have adequately addressed the reviewer's concerns and provided a clearer understanding of EviPhys's methodological contributions, computational trade-offs, and experimental effectiveness. We would greatly appreciate any further feedback from the reviewer at their earliest convenience, allowing us to promptly address any remaining questions.

We sincerely thank the reviewer and the Area Chair for their valuable time, thoughtful consideration, and continued evaluation of our work.

Reviewer_2GGNOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

EviPhys is a GraphGRU forecaster for traffic flow. It adds a frozen vision-language encoder for a global context vector, an entropy-gated per-sensor branch, and a dynamic graph. A graph Laplacian residual on free-running predictions is the physics regularizer.

Critical strengths

The paper evaluates on six datasets, including two with 716 and 2,352 sensors. It also reports a sensitivity study, a comparison of four physical priors, a noise sweep, a 10-seed significance test, and an anonymous code repository.

Gating per-sensor representations by uncertainty before graph construction is a reasonable place to inject reliability information.

Critical weaknesses

Sec. 4.4 and Fig. 3 ablate the VLM, evidence, and dynamic-graph components, but never the physics loss. No experiment reports what happens with λ_pinn = 0.

In Table 6, four different priors give MAE within 0.03 to 0.17 of each other (PEMS04: 17.95, 17.99, 17.99, 18.09), while the seed std in Table 8 is 0.06 to 0.09. The physics term appears to have no measurable effect.

The residual R_t in 3.5.3 is a graph heat equation that pulls each node toward its neighbor average. It encodes no conservation or wave physics, and the Conclusion concedes it “cannot fully characterize” traffic dynamics.

Questions for authors

How were the means in Table 8 obtained, given that all eight match the seed-42 results in Table 1 exactly? Please provide the per-seed MAE values.

Please report results with λ_pinn = 0 on PEMS and SD/GBA. Please also state at which epoch the best validation checkpoint was selected, relative to the 100-epoch warm-up.

Please replace the ViLT feature with an MLP over the same two statistics and report the result. Please also report the parameter count and inference time including the frozen ViLT.

Were the Table 1 baseline numbers re-run under the 70/10/20 split or copied from the original papers? Please state the source for each baseline.

2 comments
AuthorsOpenReview ↗

Thank you for your insightful comments. We clarify the theoretical foundations and independent contribution of physics regularization, the interpretation of physical priors, the reliability of statistical evaluations, and the effectiveness of ViLT-based multimodal encoding.

W1. Although we do not isolate the physics loss, Appendix A provides its theoretical justification. During recursive forecasting, bounded corrections use road topology, while graph diffusion residuals constrain free-running trajectories without teacher forcing during training. Appendix A proves that graph diffusion updates are non-expansive when 0≤κΔt≤10\leq\kappa\Delta t\leq1 and the residual is zero. Otherwise, multi-step state deviations are bounded by accumulated residuals, providing a theoretical basis for suppressing drift in recursive forecasting. To further quantify the independent contribution of regularization, we have added an ablation experiment with λpinn=0\lambda_{\mathrm{pinn}}=0, as shown in Table R1-1 of the next comment.

W2. Table 6 examines the model's sensitivity to different physical priors, rather than comparing performance with and without physics regularization. In sensor-level traffic forecasting, complex dynamics across roads and time periods are difficult to accurately characterize using a single simplified equation. On PEMS04, the four physical priors yield MAEs of 17.95, 17.99, 17.99, and 18.09. These comparable results indicate robustness to the tested physical priors, meaning that predictive performance does not heavily depend on any specific prior. However, this does not imply that physics regularization makes no contribution.

W3. This paper focuses on aggregated traffic flow at sensor locations, with input data incorporating road network connectivity. Rather than fully characterizing traffic dynamics using a single physical equation, we construct an auxiliary evolution prior with a clear mathematical interpretation based on road network topology. Accordingly, Sec. 3.5.3 adopts a coarse-grained propagation prior on the road graph, constraining flow changes between consecutive time steps to be consistent with physically neighboring states, rather than assuming traffic strictly follows the heat equation. Furthermore, the conditional stability and residual bounds established in Appendix A characterize the mathematical properties of this prior, rather than proving complete traffic flow conservation laws.

Q1. Table 1 reports the main experimental results using a fixed random seed of 42, whereas Table 8 summarizes the mean performance and variability across ten runs. The ten-run means for the four datasets are 14.9398, 18.3601, 19.9701, and 13.5800 for PDFormer, and 14.3100, 17.9496, 18.6598, and 12.9900 for EviPhys. Rounded to two decimal places, these means coincide with the corresponding seed-42 results. Therefore, this agreement at the reported precision does not imply that the ten-run means were derived from a single run. To ensure reproducibility, we provide unrounded means, sample standard deviations, two-sided paired t-test results, and per-seed MAE values in the following response. See Tables R1-2 and R1-3 in the next comment.

Q2. The 100-epoch warm-up does not mean that physics regularization is inactive during the first 100 epochs. Instead, its influence gradually increases through a progressive weight schedule. As described in Appendix E.3, the physics regularization coefficient accounts for the relative scales of supervised and physics losses, while min⁡(1,e/100)\min(1,e/100) controls regularization strength during early training. Therefore, physics regularization is generally active during warm-up, although its weight has not reached the full scheduled level. Meanwhile, the best checkpoint is selected based on validation performance rather than warm-up completion, so its effect should be assessed using the actual training epoch and corresponding physics weight. Following the reviewer's request, we report results with λpinn=0\lambda_{\mathrm{pinn}}=0 on PEMS and SD/GBA datasets in Table R1-1 of the next comment.

Q3. We use ViLT to jointly process three-channel images containing traffic state intensity, temporal differences, and sensor-axis differences, together with statistical text prompts derived from global variance and the standard deviation of node-wise temporal means, to extract global contextual representations. Therefore, an MLP using only two statistics cannot directly capture the complete historical state information in the images. To evaluate ViLT's actual contribution, we follow the reviewer's suggestion and replace it with an MLP using the same two statistics. The comparison results are presented in Tables R1-4 and R1-5 of the next comment.

Q4. The PEMS datasets are chronologically split into 70% training, 10% validation, and 20% testing sets, while SD and GBA use a 60%/20%/20% split. All evaluations follow consistent forecasting settings and use MAE, RMSE, and MAPE metrics. These settings are widely adopted experimental protocols in traffic flow forecasting research. Therefore, the baseline results in Table 1 are directly taken from the original papers. Notably, we reran PDFormer with random seed 42 for statistical significance analysis.

AuthorsOpenReview ↗

Thank you for suggesting additional quantitative evaluations. The following tables provide physics ablation results, statistical validation, ViLT comparisons, and computational efficiency analysis.

Table R1-1. Ablation Study of Physics Regularization

To isolate physics regularization's contribution, we compare full EviPhys with its λpinn=0\lambda_{\mathrm{pinn}}=0 variant across six datasets.

DatasetFull EviPhysλpinn=0\lambda_{\mathrm{pinn}}=0ΔMAE
PEMS0314.3114.43+0.12
PEMS0417.9518.12+0.17
PEMS0718.6618.89+0.23
PEMS0812.9913.10+0.11
SD16.6316.86+0.23
GBA19.0319.31+0.28

Analysis: Removing physics regularization increases MAE by 0.11–0.28 across all six datasets, with the largest degradation observed on GBA. These results demonstrate that graph-diffusion regularization provides additional predictive benefits beyond topology-guided correction.

Table R1-2. Statistical Significance Across Ten Random Seeds

We report ten-run means, sample standard deviations, and two-sided paired t-tests to assess forecasting consistency.

DatasetPDFormer (Mean ± Std)EviPhys (Mean ± Std)p-value
PEMS0314.9398 ± 0.08305514.3100 ± 0.0607536.29 × 10⁻⁸
PEMS0418.3601 ± 0.09103317.9496 ± 0.0720779.84 × 10⁻⁷
PEMS0719.9701 ± 0.11202018.6598 ± 0.0860056.19 × 10⁻¹⁰
PEMS0813.5800 ± 0.07395512.9900 ± 0.0590358.85 × 10⁻⁹

Note: Standard deviations use ddof=1, and p-values use two-sided paired t-tests.

Analysis: EviPhys achieves lower mean MAE and smaller standard deviations than PDFormer across all four datasets. The paired t-tests yield p-values below 0.001, supporting the statistical significance of these improvements. Moreover, the ten-run means round to the same two-decimal values as the seed-42 results in Table 1, explaining their agreement with Table 8.

Table R1-3. Per-Seed MAE Results for PDFormer and EviPhys

We provide per-seed MAEs to ensure reproducibility and clarify the relationship between Tables 1 and 8.

SeedPEMS03 PDPEMS03 EviPEMS04 PDPEMS04 Evi
4214.93814.31118.36117.947
4314.86314.30518.37718.059
4414.93014.30018.43617.919
4514.94914.38318.46717.958
4614.81814.37518.26518.040
4714.98414.23518.20517.838
4814.84614.38818.33717.980
4914.94814.23018.28417.881
5015.08314.33318.37717.997
5115.03914.24018.49217.877
SeedPEMS07 PDPEMS07 EviPEMS08 PDPEMS08 Evi
4219.97018.65813.58212.990
4319.90418.70913.55312.962
4419.73718.71513.62612.999
4519.85918.57913.66813.053
4620.13518.63613.60712.954
4720.04918.56413.49012.979
4820.02018.57613.71412.941
4920.03818.77013.53713.038
5020.00418.58813.52913.093
5119.98518.80313.49412.891

Analysis: EviPhys outperforms PDFormer for every random seed across all four datasets, demonstrating consistent improvements across 40 paired comparisons. This indicates that the performance gains are not dependent on a particular random initialization. The complete per-seed records also enable verification of the statistical results in Table R1-2.

Table R1-4. Comparison Between ViLT and a Two-Statistic MLP

We replace ViLT with an MLP using the same two statistics while keeping the remaining framework unchanged.

DatasetEviPhys (ViLT)EviPhys (MLP)ΔMAERelative Increase
PEMS0314.3114.70+0.39+2.73%
PEMS0417.9518.38+0.43+2.40%
PEMS0718.6618.89+0.23+1.23%
PEMS0812.9913.17+0.18+1.39%
SD16.6317.04+0.41+2.47%
GBA19.0319.44+0.41+2.15%

Analysis: Replacing ViLT with the two-statistic MLP increases MAE by 0.18–0.43 across all six datasets, corresponding to relative increases of 1.23%–2.73%. These consistent differences support the predictive value of jointly encoding traffic images and statistical text prompts.

Table R1-5. Parameter Count and End-to-End Inference Time

We compare parameters and end-to-end inference times, including preprocessing, to quantify ViLT's overhead.

ConfigurationTrainable ParamsFrozen ParamsTotal ParamsInference Time
EviPhys + ViLT0.877M111.595M112.472M18.4 s
EviPhys + MLP0.927M00.927M4.9 s

Note: Measurements use an NVIDIA H100 GPU on PEMS04 (batch size 32).

Analysis: Although frozen, ViLT increases total parameters to 112.472M and inference time to 18.4 s, compared with 0.927M and 4.9 s for the MLP. Table R1-4 shows that this cost yields consistent accuracy gains, while the MLP offers a lighter, faster alternative, illustrating the accuracy–efficiency trade-off.

Comments on the paper

2 comments
AuthorsOpenReview ↗

We sincerely thank the reviewers, Program Chairs, Area Chair, and Senior Program Committee members for their careful evaluation of EviPhys and the opportunity to clarify its methodological contributions, theoretical foundations, and experimental reliability.

Our work makes four complementary contributions to evidence-aware and physics-guided traffic forecasting. First, at the global representation stage, we introduce MRFM to jointly encode three-channel Hovmöller images and statistical text prompts using a frozen vision-language backbone, capturing global traffic patterns and temporal variations. Second, at the local uncertainty modeling stage, we develop DSERM to characterize uncertainty through learnable evidence channels and a Deng-entropy-inspired surrogate, adaptively regulating local representations. Third, at the spatial dependency learning stage, we propose ESGLM to integrate global and local evidence for dynamic graph construction, capturing evolving dependencies among traffic sensors. Fourth, at the recursive forecasting stage, we design PGPM to combine data-driven predictions with topology-guided corrections and graph-diffusion regularization, encouraging physically consistent multi-step forecasting.

Our responses address the reviewers' concerns regarding methodological validity, theoretical justification, experimental reliability, and practical applicability.

  1. Methodological design and interpretability. We clarify the complementary roles of MRFM, DSERM, ESGLM, and PGPM, explaining how their integration addresses distinct forecasting challenges rather than merely increasing model complexity. We also distinguish the implemented Deng-entropy-inspired surrogate from strict focal-set-wise Deng entropy and clarify the role of Hovmöller images in representing spatiotemporal patterns without treating sensor ordering as physical spatial coordinates.

  2. Physical modeling and theoretical foundations. We explain how topology-guided corrections and graph-diffusion regularization jointly incorporate physical road connectivity into recursive forecasting. We further clarify that the non-expansiveness and residual-based deviation bounds provide conditional theoretical guarantees, rather than unconditional stability or exact traffic conservation. Additional ablation experiments with λpinn=0\lambda_{\mathrm{pinn}}=0 isolate the contribution of physics regularization across six datasets.

  3. Experimental reliability and reproducibility. We clarify dataset splits, forecasting protocols, and the relationship between single-run and multi-run results. We also provide ten-seed experimental records, corrected paired t-tests, and reproducible statistical procedures to substantiate the consistency and significance of the reported improvements.

  4. Component effectiveness and noise robustness. We introduce controlled comparisons between ViLT and a lightweight statistical MLP, together with physics-loss ablations, to quantify the contributions of multimodal representation learning and physical regularization. Additional Gaussian-noise experiments compare EviPhys with HimNet and a variant without evidence modeling, demonstrating improved robustness under the evaluated perturbation levels.

  5. Computational efficiency and large-scale evaluation. We report trainable and frozen parameter counts, inference times, and GPU memory usage to clarify the trade-off between predictive accuracy and computational cost. We further evaluate EviPhys on SD and GBA, containing 716 and 2,352 sensors, respectively, and add comparisons with HimNet and PDG2Seq across multiple forecasting horizons. These experiments provide additional evidence of predictive effectiveness on large-scale traffic networks while acknowledging the computational demands of dynamic graph construction.

Together, these clarifications and additional experiments strengthen the methodological transparency, theoretical rigor, reproducibility, and empirical validation of EviPhys. We have also addressed reporting inconsistencies and refined the manuscript to better communicate the scope, contributions, and limitations of the proposed framework.

AuthorsOpenReview ↗

Request for Further Discussion

Dear Area Chair and Reviewers,

We sincerely appreciate your time and careful consideration of our submission. We have provided detailed responses and additional experimental results to address the reviewers' concerns regarding methodological design, theoretical foundations, experimental reliability, and computational efficiency. These clarifications and supplementary experiments further strengthen the validity and reproducibility of EviPhys.

As the discussion deadline is approaching, we would greatly appreciate any further feedback or suggestions. We remain available to discuss any outstanding concerns and provide additional clarifications if needed.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.