Pith. sign in

REVIEW 5 major objections 5 minor 64 references

This paper claims that a spatio-temporal foundation model pre-trained only on procedurally generated synthetic graph dynamical systems transfers zero-shot to real traffic networks, outperforming existing STFMs in ten of twelve benchmark set

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 09:44 UTC pith:B244PWIV

load-bearing objection NeoST has a genuinely new idea—synthetic-only pretraining for STFMs—and a concrete architecture, but the diversity claim rests on six traffic datasets and the missing artifacts make it hard to fully endorse the numbers. the 5 major comments →

arxiv 2607.16251 v1 pith:B244PWIV submitted 2026-06-27 cs.LG

Learning Spatio-Temporal Foundation Models from Pure Synthetic Data

Yutong Feng , Shiyuan Piao , Yutong Xia , Xu Liu , Wenqi Fan , Fugee Tsung , See-Kiong Ng , Yuxuan Liang This is my paper
classification cs.LG
keywords spatio-temporal foundation modelssynthetic datalatent-space reasoningzero-shot transfergraph dynamical systemstest-time adaptationtraffic forecastingLatent Alignment Loss
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets out to show that a spatio-temporal foundation model can be pre-trained on no real data at all—only procedurally generated graph dynamical systems—and still transfer to real traffic networks. Across six real-world benchmarks in zero-shot settings, it reports the lowest MAE/RMSE in ten of twelve configurations, with the largest margins at the 96-step horizon. The design turns three apparent liabilities into assets: a controllable synthetic corpus avoids real-world distribution bias, latent-space reasoning avoids autoregressive error accumulation and diffusion sampling cost, and a ridge-regression adapter corrects distribution shift at inference time. A sympathetic reader would care because it suggests the bottleneck for general spatio-temporal models is not data scale or domain coverage but whether pre-training exposes the right structural dynamics.

Core claim

The paper's central claim is that NeoST—pre-trained exclusively on synthetic graph-coupled stochastic processes—transfers zero-shot to real spatio-temporal systems and outperforms existing STFMs. The authors argue that real-world pre-training corpora entangle intrinsic dynamics with sensor noise and domain-specific topology, while a generator that samples random graphs, random latent dynamics, random observation projections, and random masking defines a broad manifold of dynamical regimes. Because the generator exposes the true latent state, the model is trained with a Latent Alignment Loss that supervises structural dynamics rather than point-wise observation reconstruction. The empirical r

What carries the argument

The load-bearing machinery is the synthetic data engine, which samples a distribution over stochastic graph dynamical systems: an Erdős–Rényi or Poisson graph, a normalized Laplacian filtered through Chebyshev polynomials, a second-order stochastic recurrence H_{t+1}=aH_t + tanh([T1Ht||T2H_{t-1}]W1)W2 + sigmoid(...)⊙ε_t + b with ε_t∼N(0,I), and a random linear observation map x_t=H_t w with random masking. This engine generates unlimited paired observation/latent-state sequences, enabling a latent-space reasoning transformer and a Latent Alignment Loss that projects predictions into the latent state via PCA. The hierarchical reasoner generates multiple candidate latent trajectories at differ

Load-bearing premise

The load-bearing premise is that the synthetic generator's random family of graph dynamical systems—Erdős–Rényi/Poisson graphs, Gaussian noise, tanh drift, and linear observations—is representative enough of real spatio-temporal processes that a model trained only on it will transfer to real traffic networks; if real scale-free topologies and heavy-tailed shocks lie outside that family, the zero-shot performance claim gives way.

What would settle it

Train NeoST on the described synthetic corpus and evaluate zero-shot on a network with scale-free topology and heavy-tailed exogenous shocks, such as a traffic dataset containing incident-induced outliers. If MAE at horizon 96 degrades to baseline level or worse than a real-data-trained STFM, the claim that synthetic-only pre-training suffices is falsified. A cheaper check is to compare t-SNE coverage against such a dataset: if the real manifold falls outside the synthetic manifold's envelope, the coverage argument fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If synthetic-only pre-training transfers, real-world data collection and curation cease to be the ceiling for STFM development; domain-specific benchmark training can be replaced by controlled generation.
  • Latent supervision from simulator ground truth produces representations of dynamics, not just observations; this suggests a path to world models that reason about latent state.
  • The 96-step results imply long-horizon forecasting can be stabilized without autoregressive decoding; error accumulation is not inevitable.
  • The test-time ridge-regression adapter offers a cheap way to correct distribution shift using incoming observations, applicable to any latent-embedding forecaster.
  • The generator can sample endlessly and configurations are controllable, so scaling laws can be studied without collecting more real data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the representativeness assumption is taken seriously, the same recipe should transfer beyond traffic to domains with similar spectral and dynamical structure, such as energy, weather, or biology; a testable prediction is that performance degrades gracefully with dynamical similarity rather than domain name.
  • The latent alignment loss could be used as a diagnostic: inverting a model's latent predictions could recover estimated latent states, making it possible to inspect what the model believes is driving a real system.
  • The heavy-tailed limitation suggests a concrete extension: injecting Lévy noise and scale-free graphs would test whether the zero-shot gains hold under extreme events; if they do not, the current claim is bounded to Gaussian and quasi-periodic regimes.
  • The cubic matrix inversion in test-time adaptation is likely the first bottleneck as graph size grows, so a low-rank or iterative variant would be needed to extend the result to very large networks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes NeoST, claimed to be the first spatio-temporal foundation model pre-trained exclusively on procedurally generated synthetic data. The synthetic engine (Section 3) samples random graphs and latent stochastic dynamics with Chebyshev filters, then produces observable signals via random linear projections and masking. The architecture (Section 4) uses patching, a spatio-temporal encoder, and a hierarchical latent reasoner that generates and refines multiple future trajectories without autoregressive decoding. Training (Section 5) combines a latent alignment loss with observation reconstruction, and test-time adaptation fits a ridge regression adapter on observed embeddings. Experiments on six traffic datasets (PEMS03/04/07/08, PEMS-BAY, METR-LA) report zero-shot MAE/RMSE against six baselines, with NeoST claimed to achieve the best result in 10 of 12 settings, plus ablations, efficiency analysis, and a t-SNE coverage plot.

Significance. If the central claims hold, the paper would make a useful contribution: it demonstrates a plausible route to avoiding real-world pre-training bias by generating an unlimited, controllable synthetic corpus, and it proposes a non-autoregressive latent-space reasoning architecture with test-time adaptation. The theoretical appendix (B) provides some internal consistency guarantees (spectral bounds, moment control, conditional Gaussian transitions) and is a strength, as is the explicit description of the synthetic data engine. The main weaknesses are empirical scope and validation rigor: the 'diverse real-world' claim rests entirely on traffic data, the t-SNE-based coverage argument is not statistically valid, the key hyperparameter α is tuned on test benchmarks, and the '10 of 12' ranking depends on excluding several baselines. These issues are load-bearing for the paper's headline claim but are addressable with additional experiments and sharper reporting.

major comments (5)
  1. [Abstract; §6.1; Table 2; Appendix C.1] The abstract claims 'diverse real-world benchmarks' and 'diverse real-world spatio-temporal systems', but all six evaluation datasets are California traffic flow/speed at 5-minute resolution. Appendix C.1 even states the benchmarks span 'traffic flow, speed, electricity consumption, and meteorology', yet Table 2 lists no electricity or meteorological datasets. This internal inconsistency directly weakens the central zero-shot transfer claim. Please either add non-traffic benchmarks (e.g., electricity load, weather, air quality) or explicitly scope the paper's claims to traffic-like spatio-temporal systems.
  2. [§6.5, Figure 7] The claim that synthetic data 'comprehensively envelopes' real clusters is inferred solely from a t-SNE plot. t-SNE preserves local neighborhoods but not densities, distances, or global containment; it cannot support the 'enveloping' conclusion. Please provide quantitative distributional coverage evidence, such as density ratios, nearest-neighbor distances, or manifold precision/recall, and ideally report the same analysis on non-traffic domains. Without this, the representativeness of the synthetic corpus remains unestablished.
  3. [§6.3, Figure 4; Table 1] The hyperparameter α is selected on PEMS03 and PEMS07 (Figure 4), both of which appear as test datasets in Table 1. For these datasets the reported zero-shot results are therefore not fully zero-shot with respect to the objective trade-off. Please fix α using validation data or a separate tuning dataset, then report Table 1 with that fixed value. Also state how the ridge regularization λ in the test-time adapter (Section 5) is chosen, since the same circularity could apply.
  4. [Table 1, footnote] The '10 of 12' claim and the statement that NeoST 'consistently surpasses' baselines rely on excluding asterisked baselines from the ranking. That exclusion is methodologically defensible for zero-shot comparison, but the text should state the comparison class explicitly. When all reported numbers are considered, NeoST is not best in several settings: PEMS-Bay 12 (Moirai 1.97, FactoST 2.02 vs NeoST 2.67), PEMS-08 96 (TimesFM 92.72 vs NeoST 94.03), PEMS-03 96 (Moirai 74.85 vs NeoST 84.49), and METR-LA 12 (OpenCity 4.30 vs NeoST 4.90). Please avoid the unqualified 'consistently outperforms existing STFMs' claim.
  5. [Appendix D.1] The paper itself acknowledges a 'sim-to-real' gap: the generator uses Erdős–Rényi/Poisson graphs and Gaussian noise, while real networks are often scale-free/small-world and exogenous shocks are heavy-tailed. Given that all downstream evaluations are traffic data, this concession directly affects the representativeness assumption on which the pure-synthetic pre-training claim rests. The theoretical properties in Appendix B establish internal consistency of the generator, not coverage of real phenomena. Please provide evidence of transfer under the acknowledged gap, or extend the generator to non-Gaussian/heavy-tailed regimes and scale-free topologies and re-evaluate.
minor comments (5)
  1. [Figure 1] The caption contains a stray '??' and the diagram labels are hard to read. Please clean up the figure and caption.
  2. [Eq. (4), Section 5] The training objective applies 'PCA' to project model outputs into the latent space. PCA is not a standard differentiable layer; please specify how the projection is computed during training (e.g., fixed basis, online/batch-wise, separate of gradients) and whether the same projection is used for all samples.
  3. [Table 1] No error bars or multiple-seed results are reported. For a foundation-model comparison, at least the main zero-shot results should include mean ± std over a few seeds to assess significance.
  4. [Appendix C.3.2, UniST] The text says UniST is 'retrained on 13 datasets from its original release' to handle 12- and 96-step horizons. Please clarify whether any of the six evaluation datasets are among those 13; if so, apply the same asterisk convention and exclusion rule to UniST.
  5. [Availability] The paper does not release code, the synthetic generator, or checkpoints. Given that the central contribution is a data engine, releasing the generator is important for reproducibility and for the community to verify the claims.

Circularity Check

2 steps flagged

Core synthetic-to-real transfer is externally benchmarked, but two local steps are self-referential: alpha is tuned on the very test sets reported in Table 1, and the synthetic-coverage guarantee reduces to the generator's own sampling distribution.

specific steps
  1. fitted input called prediction [Section 6.3 (Ablation and Hyperparameter Analysis, Figure 4); results reported in Table 1; objective in Eq. (4)]
    "we evaluate the zero-shot forecasting Mean Absolute Error (MAE) under both short-term (12 steps) and long-term (96 steps) horizons on the PEMS03 and PEMS07 datasets. This analysis systematically ablates critical architectural modules, examines the sensitivity of the latent alignment objective by varying the trade-off hyperparameterα∈[0,1] (Figure 4)... The optimal balance consistently resides within α∈[0.2,0.6]"

    The scalar α in LAL (Eq. 4) is selected by minimizing the zero-shot MAE on PEMS03 and PEMS07, and the same two datasets, at the same 12/96 horizons, appear in Table 1 as evidence that 'NeoST consistently surpasses both spatio-temporal and time series foundation models... in ten out of the twelve evaluated settings.' The reported wins on PEMS03/PEMS07 are therefore partly a function of fitting the hyperparameter to the evaluation data rather than independent zero-shot predictions. This is a local, two-dataset instance of fitting a parameter to a subset of data and then reporting a closely related quantity as a prediction; the other four benchmarks are not used in the α sweep.

  2. self definitional [Appendix B (Corpus-level coverage, Eq. (22)); Section 6.5 (Figure 7); concession in Appendix D.1]
    "Thus, any regime with nonzero probability under the sampling distribution appears in sufficiently large synthetic corpora with high probability. ... the synthetic data forms an expansive, continuous manifold that comprehensively envelopes the localized clusters of all real-world test sets. This empirical evidence validates our core hypothesis."

    The coverage guarantee in Eq. (22) is a restatement of the definition of the generator's sampling distribution ν: it only ensures that regimes already assigned nonzero probability by ν will eventually be sampled. It says nothing about whether real traffic/weather/power regimes lie in the support of ν, which is exactly the transfer premise under test. Section 6.5 then treats the assertion that the synthetic t-SNE manifold 'comprehensively envelopes' real clusters as validation of that premise; t-SNE does not preserve the containment/density relations claimed. The coverage conclusion therefore reduces to the generator's own definition, and Appendix D.1 concedes the gap: real networks are scale-free/small-world with heavy-tailed shocks, while ν draws Erdős–Rényi/Poisson graphs and Gaussian ε.

full rationale

The main derivation chain is not circular by construction: Section 3 defines a synthetic generator (Erdős–Rényi/Poisson graphs, tanh drift, Gaussian noise, x_t=H_t w); Section 5 trains LAL on the generator's own latent states and observations; Table 1 evaluates zero-shot on external traffic datasets (PEMS/METR). The evaluation data never enters the generator parameters or pre-training weights, and the ridge-regression TTA fits only the observed input window before predicting out-of-window future values, so the headline forecasts are not statistically forced by the fit. Self-citations in the paper ([47] synthetic-data survey, settings per Zhong et al. [44]) are numerous but not load-bearing for the central result, which is judged against external baselines and external benchmarks. Two partial self-referential steps are flagged above: α is tuned on PEMS03/PEMS07 and then reported in Table 1, and the coverage argument (Eq. 22 + Figure 7) guarantees only coverage of the synthetic distribution's own support while claiming real-manifold containment. Appendix D.1 explicitly concedes the sim-to-real gap (scale-free/small-world topology, heavy-tailed shocks, 'sim-to-real' distribution gap), so the representativeness premise is an acknowledged assumption — a correctness risk rather than a hidden circularity. Additional reporting caveats weigh on the verdict without changing its nature: Table 1 excludes starred baselines from the ranking despite some beating NeoST (e.g., PEMS-Bay 12: FactoST 2.02 vs NeoST 2.67; PEMS-08 96: TimesFM 92.72 vs NeoST 94.03), and Appendix C.1's text claims electricity/meteorology benchmarks while Table 2 lists only six traffic datasets. Because the central synthetic-to-real claim rests on external data and does not reduce to a fit or a self-citation chain, the score is moderate rather than high.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests primarily on a domain-coverage assumption: the synthetic generator's random graphs, Gaussian noise, tanh drift, and linear observation maps are representative of real spatio-temporal systems. The LAL pretraining also assumes the generator's latent H_t is a meaningful universal supervisory target. Standard spectral-graph facts (Appendix B) are used for stability but do not establish transfer.

free parameters (3)
  • α (latent alignment weight) = optimal range [0.2, 0.6] (not a single reported value)
    Controls balance between latent H alignment and observation reconstruction in Eq. (4); tuned on PEMS03/PEMS07 validation, Section 6.3/Figure 4.
  • λ (TTA ridge regularization) = not reported
    In closed-form ridge solution of Section 5 (Test-Time Adaptation); no numerical value given, required to reproduce TTA.
  • Generator sampling ranges (N, d_h, a, b, masking ratio) = N∈[64,512], d_h∈[4,16]; a∈(0,1), b and mask ratio unspecified
    Chosen by hand in Section 6.1; these define the synthetic pre-training distribution and are not fitted to downstream data.
axioms (4)
  • domain assumption Real spatio-temporal systems are sufficiently covered by random Erdős–Rényi/Poisson graphs plus Chebyshev-filtered stochastic latent dynamics.
    Section 3 and Figure 2 construct the corpus; Appendix D.1 admits scale-free/small-world topologies and heavy-tailed noise are absent, so coverage is assumed, not demonstrated.
  • domain assumption Observable signals are low-dimensional random linear projections x_t=H_t w of an underlying latent state.
    Section 3, Observation Generation; transfer to real traffic depends on this structural postulate.
  • domain assumption The synthetic latent state H_t is a valid privileged supervision target; LAL/PCA alignment teaches transferable structural dynamics.
    Eq. (4) and Section 5; no independent evidence connects H_t to real-world latent processes.
  • standard math Normalized Laplacian spectrum lies in [0,2] for sampled undirected graphs with nonnegative weights, so Chebyshev filters are contractive.
    Appendix B, Theorem B.3; standard spectral graph theory, used to justify filter stability.
invented entities (1)
  • Synthetic latent state H_t no independent evidence
    purpose: Privileged supervision signal for Latent Alignment Loss; the model is trained to predict it on synthetic data.
    H_t is internal to the generator (Eq. 1) and is not observable in real deployments; no falsifiable handle connects it to real traffic dynamics.

pith-pipeline@v1.3.0-alltime-deepseek · 19219 in / 14674 out tokens · 122247 ms · 2026-08-02T09:44:30.086439+00:00 · methodology

0 comments
read the original abstract

Spatio-Temporal Foundation Models (STFMs) aim to learn generalizable representations of complex dynamical systems across space and time. However, existing approaches suffer from distributional bias in real-world pre-training data, structural bottlenecks of autoregressive or diffusion-based paradigms, and objectives that overemphasize point-wise reconstruction in noisy observation space.We propose \textbf{NeoST}, the first spatio-temporal foundation model pre-trained solely on procedurally generated synthetic systems. NeoST introduces a scalable synthetic pre-training corpus to mitigate real-world bias, a latent-space reasoning architecture that generates and iteratively refines multiple future trajectories without sequential error accumulation, and latent-space objectives that emphasize structural dynamics and enable inference-time correction under distribution shifts.Extensive experiments across diverse real-world benchmarks show that NeoST consistently outperforms existing STFMs in diverse real-world spatio-temporal systems, achieves superior long-horizon stability and inference efficiency.

Figures

Figures reproduced from arXiv: 2607.16251 by Fugee Tsung, See-kiong Ng, Shiyuan Piao, Wenqi Fan, Xu Liu, Yutong Feng, Yutong Xia, Yuxuan Liang.

Figure 1
Figure 1. Figure 1: Comparison of Paradigms: NeoST vs. Recent STFM. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic data pipeline. Graph Sampling & Operator Construction. To promote structural diversity, the graph topology is sampled on-the-fly from random graph families, including Erdos–Rényi and Poisson degree models. ˝ Self-loops are added to stabilize the spectrum. Given A, we compute the normalized Laplacian L = I − D−1/2AD−1/2 , where D is the degree matrix. To model spatial interactions in a principled … view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of NeoST. 4.2 Spatio-Temporal Encoder The spatio-temporal encoder transforms the patch-level representation Z1:PT ∈ R PT ×N×d , where PT denotes the number of patches obtained from patchifying the input sequence, into a contextual latent memory. This is accomplished through alternating between two complementary mechanisms: intra￾node temporal attention, which captures temporal evolutio… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation and hyperparameter analysis. • Architectural Contributions: The full NeoST framework attains the lowest MAE and exhibits superior predictive stability across most evaluated scenarios. Specifically, disabling multi-scale prediction significantly impairs forecasting robustness over long-term horizons, whereas omitting the trajectory refinement module results in a substantial increase in cumulative s… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of Test-Time Adaptation (TTA). 6.4 Inference Efficiency (RQ3) 2 5 10 25 100 500 1000 Inference Time (s) [Log Scale] 15 20 25 30 40 60 100 150 Mean Absolute Error [Log Scale] Efficiency Comparison: MAE vs. Inference Time vs. Params UniST 2.42M | 2.1s FactoST 1.3M | 8.1s MOIRAI 91.4M | 22.2s OpenCity 1.67M | 25.3s TimesFM 200.0M | 470.1s NeoST (Ours) 89.26M | 24.0s ROSE MAE: 27.5 | Params: 6.4M Time: … view at source ↗
Figure 6
Figure 6. Figure 6: Efficiency comparison. To answer RQ3, we evaluate the computational efficiency of NeoST by comparing its inference time, parameter scale, and predictive perfor￾mance (MAE) against representative baselines on PEMS03 dataset, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: t-SNE visualization. To address RQ4, we visualize the representational manifold of our synthetic corpus against six real￾world benchmarks using t-SNE. After applying in￾stance normalization to fixed-length temporal se￾quences (12 steps) from all datasets to isolate intrinsic temporal dynamics, we projected these features into a 2D space. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 1 canonical work pages

  1. [1]

    Foundation models for spatio-temporal data science: A tutorial and survey,

    Y . Liang, H. Wen, Y . Xia, M. Jin, B. Yang, F. Salim, Q. Wen, S. Pan, and G. Cong, “Foundation models for spatio-temporal data science: A tutorial and survey,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .2, 2025, pp. 6063–6073

    2025

  2. [2]

    Unraveling spatio-temporal foundation models via the pipeline lens: A comprehensive review,

    Y . Fang, H. Miao, Y . Liang, L. Deng, Y . Cui, X. Zeng, Y . Xia, Y . Zhao, T. B. Pedersen, C. S. Jensen et al., “Unraveling spatio-temporal foundation models via the pipeline lens: A comprehensive review,”IEEE Transactions on Knowledge and Data Engineering, 2026

    2026

  3. [3]

    Spatio-temporal foundation models: Vision, challenges, and opportunities,

    A. Goodge, W. S. Ng, B. Hooi, and S. K. Ng, “Spatio-temporal foundation models: Vision, challenges, and opportunities,” arXiv preprint arXiv:2501.09045, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Spatio-temporal graph neural networks for predictive learning in urban computing: A survey,

    G. Jin, Y . Liang, Y . Fang, Z. Shao, J. Huang, J. Zhang, and Y . Zheng, “Spatio-temporal graph neural networks for predictive learning in urban computing: A survey,”IEEE transactions on knowledge and data engineering, vol. 36, no. 10, pp. 5388–5408, 2023

    2023

  5. [5]

    Flownet: Modeling dynamic spatio-temporal systems via flow propagation,

    Y . Feng, X. Liu, Y . Xia, and Y . Liang, “Flownet: Modeling dynamic spatio-temporal systems via flow propagation,” arXiv preprint arXiv:2511.05595, 2025

    arXiv 2025

  6. [6]

    Representation learning for spatiotemporal physical systems,

    H. Qu, R. Morel, M. McCabe, A. Bietti, F. Lanusse, S. Ho, and Y . LeCun, “Representation learning for spatiotemporal physical systems,” arXiv preprint arXiv:2603.13227, 2026

    arXiv 2026

  7. [7]

    Hierarchical planning with latent world models,

    W. Zhang, B. Terver, A. Zholus, S. Chitnis, H. Sutaria, M. Assran, R. Balestriero, A. Bar, A. Bardes, Y . LeCunet al., “Hierarchical planning with latent world models,” arXiv preprint arXiv:2604.03208, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    An observed value consistent diffusion model for imputing missing values in multivariate time series,

    X. Wang, H. Zhang, P. Wang, Y . Zhang, B. Wang, Z. Zhou, and Y . Wang, “An observed value consistent diffusion model for imputing missing values in multivariate time series,” in Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, 2023, pp. 2409–2418

    2023

  9. [9]

    Foundation models for time series analysis: A tutorial and survey,

    Y . Liang, H. Wen, Y . Nie, Y . Jiang, M. Jin, D. Song, S. Pan, and Q. Wen, “Foundation models for time series analysis: A tutorial and survey,” in Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, 2024, pp. 6555–6565

    2024

  10. [10]

    Deep learning for time series forecasting: Tutorial and literature survey,

    K. Benidis, S. S. Rangapuram, V . Flunkert, Y . Wang, D. Maddix, C. Turkmen, J. Gasthaus, M. Bohlke-Schneider, D. Salinas, L. Stella et al., “Deep learning for time series forecasting: Tutorial and literature survey,”ACM Computing Surveys, vol. 55, no. 6, pp. 1–36, 2022

    2022

  11. [11]

    Time-series forecasting with deep learning: a survey,

    B. Lim and S. Zohren, “Time-series forecasting with deep learning: a survey,” Philosophical transactions of the royal society a: mathematical, physical and engineering sciences, vol. 379, no. 2194, 2021

    2021

  12. [12]

    A survey on diffusion models for time series and spatio-temporal data,

    Y . Yang, M. Jin, H. Wen, C. Zhang, Y . Liang, L. Ma, Y . Wang, C. Liu, B. Yang, Z. Xu, S. Pan, and Q. Wen, “A survey on diffusion models for time series and spatio-temporal data,” ACM Comput. Surv., vol. 58, no. 8, Feb. 2026. [Online]. Available: https://doi.org/10.1145/3783986

    doi:10.1145/3783986 2026

  13. [13]

    A survey on diffusion models for time series and spatio-temporal data,

    Y . Yang, M. Jin, H. Wen, C. Zhang, Y . Liang, L. Ma, Y . Wang, C. Liu, B. Yang, Z. Xu, J. Bian, S. Pan, and Q. Wen, “A survey on diffusion models for time series and spatio-temporal data,” 2024

    2024

  14. [14]

    Diffstg: Probabilistic spatio-temporal graph forecasting with denoising diffusion models,

    H. Wen, Y . Lin, Y . Xia, H. Wan, Q. Wen, R. Zimmermann, and Y . Liang, “Diffstg: Probabilistic spatio-temporal graph forecasting with denoising diffusion models,” in Proceedings of the 31st ACM international conference on advances in geographic information systems, 2023, pp. 1–12

    2023

  15. [15]

    Towards unifying diffusion models for probabilistic spatio-temporal graph learning,

    J. Hu, X. Liu, Z. Fan, Y . Liang, and R. Zimmermann, “Towards unifying diffusion models for probabilistic spatio-temporal graph learning,” in Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems, 2024, pp. 135–146

    2024

  16. [16]

    A comprehensive survey of regression-based loss functions for time series forecasting,

    A. Jadon, A. Patil, and S. Jadon, “A comprehensive survey of regression-based loss functions for time series forecasting,” in International Conference on Data Management, Analytics & Innovation. Springer, 2024, pp. 117–147

    2024

  17. [17]

    Soft-dtw: a differentiable loss function for time-series,

    M. Cuturi and M. Blondel, “Soft-dtw: a differentiable loss function for time-series,” in International conference on machine learning. PMLR, 2017, pp. 894–903

    2017

  18. [18]

    Shape and time distortion loss for training deep time series forecasting models,

    V . Le Guen and N. Thome, “Shape and time distortion loss for training deep time series forecasting models,” Advances in neural information processing systems, vol. 32, 2019. 10

    2019

  19. [19]

    Large models for time series and spatio-temporal data: A survey and outlook,

    M. Jin, Q. Wen, Y . Liang, C. Zhang, S. Xue, X. Wang, J. Zhang, Y . Wang, H. Chen, X. Liet al., “Large models for time series and spatio-temporal data: A survey and outlook,” arXiv preprint arXiv:2310.10196, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Streasoner: Empowering llms for spatio- temporal reasoning in time series via spatial-aware reinforcement learning,

    J. Ni, S. Wang, M. Jin, Q. He, and W. Jin, “Streasoner: Empowering llms for spatio- temporal reasoning in time series via spatial-aware reinforcement learning,” arXiv preprint arXiv:2601.03248, 2026

    Pith/arXiv arXiv 2026

  21. [21]

    A decoder-only foundation model for time-series forecasting,

    A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder-only foundation model for time-series forecasting,” ICML, 2024, pretrained TimesFM on 100B time points from Google Trends, Wikipedia, synthetic and real-world data

    2024

  22. [22]

    Chronos: Learning the language of time series,

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Ranga- puram, S. P. Arango, S. Kapoor et al., “Chronos: Learning the language of time series,” arXiv preprint arXiv:2403.07815, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Unified training of universal time series forecasting transformers,

    G. Woo, C. Liu, A. Kumar, C. Xiong, S. Savarese, and D. Sahoo, “Unified training of universal time series forecasting transformers,” 2024

    2024

  24. [24]

    Moirai-moe: Empowering time series foundation models with sparse mixture of experts,

    X. Liu, J. Liu, G. Woo, T. Aksu, Y . Liang, R. Zimmermann, C. Liu, S. Savarese, C. Xiong, and D. Sahoo, “Moirai-moe: Empowering time series foundation models with sparse mixture of experts,” arXiv preprint arXiv:2410.10469, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Moirai 2.0: When less is more for time series forecasting,

    C. Liu, T. Aksu, J. Liu, X. Liu, H. Yan, Q. Pham, S. Savarese, D. Sahoo, C. Xiong, and J. Li, “Moirai 2.0: When less is more for time series forecasting,” arXiv preprint arXiv:2511.11698, 2025

    arXiv 2025

  26. [26]

    Unitime: A language- empowered unified model for cross-domain time series forecasting,

    X. Liu, J. Hu, Y . Li, S. Diao, Y . Liang, B. Hooi, and R. Zimmermann, “Unitime: A language- empowered unified model for cross-domain time series forecasting,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 4095–4106

    2024

  27. [27]

    Moment: A family of open time-series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foundation models,” arXiv preprint arXiv:2402.03885, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Time-moe: Billion-scale time series foundation models with mixture of experts,

    X. Shi, S. Wang, Y . Nie, D. Li, Z. Ye, Q. Wen, and M. Jin, “Time-moe: Billion-scale time series foundation models with mixture of experts,” arXiv preprint arXiv:2409.16040, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Opencity: Open spatio-temporal foundation models for traffic prediction,

    Z. Li, L. Xia, L. Shi, Y . Xu, D. Yin, and C. Huang, “Opencity: Open spatio-temporal foundation models for traffic prediction,” arXiv preprint arXiv:2408.10269, 2024

    arXiv 2024

  30. [30]

    Unist: A prompt-empowered universal model for urban spatio-temporal prediction,

    Y . Yuan, J. Ding, J. Feng, D. Jin, and Y . Li, “Unist: A prompt-empowered universal model for urban spatio-temporal prediction,” Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 4095–4106, 2024

    2024

  31. [31]

    Urbangpt: Spatio-temporal large language models,

    Z. Li, L. Xia, J. Tang, Y . Xu, L. Shi, L. Xia, D. Yin, and C. Huang, “Urbangpt: Spatio-temporal large language models,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 5351–5362

    2024

  32. [32]

    Airformer: Predicting nationwide air quality in china with transformers,

    Y . Liang, Y . Xia, S. Ke, Y . Wang, Q. Wen, J. Zhang, Y . Zheng, and R. Zimmermann, “Airformer: Predicting nationwide air quality in china with transformers,” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, no. 12, 2023, pp. 14 329–14 337

    2023

  33. [33]

    Spatio-temporal field neural networks for air quality inference,

    Y . Feng, Q. Wang, Y . Xia, J. Huang, S. Zhong, and Y . Liang, “Spatio-temporal field neural networks for air quality inference,” in Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, K. Larson, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2024, pp. 7260–7268, aI for Good....

    work page doi:10.24963/ijcai.2024/803 2024

  34. [34]

    Accurate medium-range global weather forecasting with 3d neural networks,

    K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian, “Accurate medium-range global weather forecasting with 3d neural networks,” Nature, vol. 619, no. 7970, pp. 533–538, 2023

    2023

  35. [35]

    Climax: A foundation model for weather and climate,

    T. Nguyen, J. Brandstetter, A. Kapoor, J. K. Gupta, and A. Grover, “Climax: A foundation model for weather and climate,” arXiv preprint arXiv:2301.10343, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    Hyper- connections,

    D. Zhu, H. Huang, Z. Huang, Y . Zeng, Y . Mao, B. Wu, Q. Min, and X. Zhou, “Hyper- connections,” arXiv preprint arXiv:2409.19606, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Reversible instance normalization for accurate time-series forecasting against distribution shift,

    T. Kim, J. Kim, Y . Tae, C. Park, J.-H. Choi, and J. Choo, “Reversible instance normalization for accurate time-series forecasting against distribution shift,” in International conference on learning representations, 2021. 11

    2021

  38. [38]

    Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free,

    Z. Qiu, Z. Wang, B. Zheng, Z. Huang, K. Wen, S. Yang, R. Men, L. Yu, F. Huang, S. Huanget al., “Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free,” arXiv preprint arXiv:2505.06708, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: Enhanced transformer with rotary position embedding,” Neurocomputing, vol. 568, p. 127063, 2024

    2024

  40. [40]

    Structured sequence modeling with graph convolutional recurrent networks,

    Y . Seo, M. Defferrard, P. Vandergheynst, and X. Bresson, “Structured sequence modeling with graph convolutional recurrent networks,” 2016. [Online]. Available: https://arxiv.org/abs/1612.07659

    Pith/arXiv arXiv 2016

  41. [41]

    Principal component analysis,

    S. Wold, K. Esbensen, and P. Geladi, “Principal component analysis,” Chemometrics and intelligent laboratory systems, vol. 2, no. 1-3, pp. 37–52, 1987

    1987

  42. [42]

    Learning dynamics and heterogeneity of spatial-temporal graph data for traffic forecasting,

    S. Guo, Y . Lin, H. Wan, X. Li, and G. Cong, “Learning dynamics and heterogeneity of spatial-temporal graph data for traffic forecasting,” IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 11, pp. 5415–5428, 2021

    2021

  43. [43]

    Diffusion convolutional recurrent neural network: Data- driven traffic forecasting,

    Y . Li, R. Yu, C. Shahabi, and Y . Liu, “Diffusion convolutional recurrent neural network: Data- driven traffic forecasting,” in International Conference on Learning Representations (ICLR ’18), 2018

    2018

  44. [44]

    Learning to factorize spatio-temporal foundation models,

    S. Zhong, J. Qiu, Y . Wu, X. Zou, Z. Rao, B. Yang, C. Guo, H. Xu, and Y . Liang, “Learning to factorize spatio-temporal foundation models,” in The Thirty-ninth Annual Conference on Neural Information Processing Systems

  45. [45]

    Rose: Register-assisted general time series forecasting with decomposed frequency learning,

    Y . Wang, Y . Qiu, P. Chen, K. Zhao, Y . Shu, Z. Rao, L. Pan, B. Yang, and C. Guo, “Rose: Register-assisted general time series forecasting with decomposed frequency learning,” 2024

    2024

  46. [46]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  47. [47]

    Empowering time series analysis with synthetic data: A survey and outlook in the era of foundation models,

    X. Liu, T. Aksu, J. Liu, Q. Wen, Y . Liang, C. Xiong, S. Savarese, D. Sahoo, J. Li, and C. Liu, “Empowering time series analysis with synthetic data: A survey and outlook in the era of foundation models,” arXiv preprint arXiv:2503.11411, 2025

    Pith/arXiv arXiv 2025

  48. [48]

    Largest: A benchmark dataset for large-scale traffic forecasting,

    X. Liu, Y . Xia, Y . Liang, J. Hu, Y . Wang, L. Bai, C. Huang, Z. Liu, B. Hooi, and R. Zimmer- mann, “Largest: A benchmark dataset for large-scale traffic forecasting,” Advances in Neural Information Processing Systems, vol. 36, pp. 75 354–75 371, 2023

    2023

  49. [49]

    Predicting parking availability in singapore with cross-domain data: A new dataset and a data-driven approach,

    H. Zhang, Y . Xia, S. Zhong, K. Wang, Z. Tong, Q. Wen, R. Zimmermann, and Y . Liang, “Predicting parking availability in singapore with cross-domain data: A new dataset and a data-driven approach,” arXiv preprint arXiv:2405.18910, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Lade: The first comprehensive last-mile express dataset from industry,

    L. Wu, H. Wen, H. Hu, X. Mao, Y . Xia, E. Shan, J. Zheng, J. Lou, Y . Liang, L. Yanget al., “Lade: The first comprehensive last-mile express dataset from industry,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 5991–6002

    2024

  51. [51]

    Generative adversarial networks for spatio-temporal data: A survey,

    N. Gao, H. Xue, W. Shao, S. Zhao, K. K. Qin, A. Prabowo, M. S. Rahaman, and F. D. Salim, “Generative adversarial networks for spatio-temporal data: A survey,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 13, no. 2, pp. 1–25, 2022

    2022

  52. [52]

    Synthetic datasets for machine learning on spatio-temporal graphs using pdes,

    J. Arndt, U. Isil, M. Detzel, W. Samek, and J. Ma, “Synthetic datasets for machine learning on spatio-temporal graphs using pdes,” arXiv preprint arXiv:2502.04140, 2025

    Pith/arXiv arXiv 2025

  53. [53]

    Difftraj: Generating gps trajectory with diffusion probabilistic model,

    Y . Zhu, Y . Ye, S. Zhang, X. Zhao, and J. Yu, “Difftraj: Generating gps trajectory with diffusion probabilistic model,” Advances in Neural Information Processing Systems, vol. 36, pp. 65 168– 65 188, 2023

    2023

  54. [54]

    Controltraj: Controllable trajectory generation with topology-constrained diffusion model,

    Y . Zhu, J. J. Yu, X. Zhao, Q. Liu, Y . Ye, W. Chen, Z. Zhang, X. Wei, and Y . Liang, “Controltraj: Controllable trajectory generation with topology-constrained diffusion model,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 4676–4687

    2024

  55. [55]

    Causal time series generation via diffusion models,

    Y . Xia, C. Xu, Y . Liang, Q. Wen, R. Zimmermann, and J. Bian, “Causal time series generation via diffusion models,” arXiv preprint arXiv:2509.20846, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Self-supervised learning for time series analysis: Taxonomy, progress, and prospects,

    K. Zhang, Q. Wen, C. Zhang, R. Cai, M. Jin, Y . Liu, J. Y . Zhang, Y . Liang, G. Pang, D. Song et al., “Self-supervised learning for time series analysis: Taxonomy, progress, and prospects,” IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 10, pp. 6775–6794, 2024. 12

    2024

  57. [57]

    Self-supervised learning from images with a joint-embedding predictive architecture,

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Ballas, “Self-supervised learning from images with a joint-embedding predictive architecture,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 15 619–15 629

    2023

  58. [58]

    Lejepa: Provable and scalable self-supervised learning without the heuristics,

    R. Balestriero and Y . LeCun, “Lejepa: Provable and scalable self-supervised learning without the heuristics,” arXiv preprint arXiv:2511.08544, 2025

    Pith/arXiv arXiv 2025

  59. [59]

    T-jepa: A joint-embedding predictive architecture for trajectory similarity computation,

    L. Li, H. Xue, Y . Song, and F. Salim, “T-jepa: A joint-embedding predictive architecture for trajectory similarity computation,” in Proceedings of the 32nd ACM international conference on advances in geographic information systems, 2024, pp. 569–572

    2024

  60. [60]

    Hit-jepa: A hierarchical self-supervised trajectory embedding framework for similarity computation,

    L. Li, H. Xue, S. Ao, Y . Song, and F. Salim, “Hit-jepa: A hierarchical self-supervised trajectory embedding framework for similarity computation,” arXiv preprint arXiv:2507.00028, 2025

    Pith/arXiv arXiv 2025

  61. [61]

    Towards neural scaling laws for time series foundation models,

    Q. Yao, C.-H. H. Yang, R. Jiang, Y . Liang, M. Jin, and S. Pan, “Towards neural scaling laws for time series foundation models,” arXiv preprint arXiv:2410.12360, 2024

    Pith/arXiv arXiv 2024

  62. [62]

    Scaling law for time series forecasting,

    J. Shi, Q. Ma, H. Ma, and L. Li, “Scaling law for time series forecasting,” Advances in Neural Information Processing Systems, vol. 37, pp. 83 314–83 344, 2024

    2024

  63. [63]

    Scaling-laws for large time-series models,

    T. D. Edwards, J. Alvey, J. Alsing, N. H. Nguyen, and B. D. Wandelt, “Scaling-laws for large time-series models,” arXiv preprint arXiv:2405.13867, 2024. 13 A Related Works Synthetic Spatio-Temporal Data.The use of synthetic data for time series and spatio-temporal modeling has gained increasing attention, particularly in the era of foundation models. A co...

    Pith/arXiv arXiv 2024

  64. [64]

    sim-to-real

    construct synthetic spatio-temporal graphs governed by physical dynamics, offering improved controllability. Diffusion models for trajectory [ 53, 54] and time series [55] generation have also been explored. Despite these efforts, synthetic data is rarely used as theprimarysource for STFM pre-training. Most existing STFMs still rely heavily on aggregated ...

    2018