Pith. sign in

REVIEW 6 major objections 7 minor 49 references

Reprogramming Vision Foundation Models for Spatio-Temporal Forecasting

T0 review · 6 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read By feeding a frozen vision transformer both the raw spatio-temporal grid and a first-order temporal-difference flow, ST-VFM claims to outperform every compared baseline on all ten forecasting datasets.

desk verdict A genuinely interesting idea—frozen VFMs for spatio-temporal forecasting via two-stream input—but the empirical evidence is not yet solid enough to support the universal superiority claim. read the letter →

arxiv 2507.11558 v1 pith:N3LLRDTY submitted 2025-07-14 cs.CV cs.AI

classification cs.CVcs.AI
keywords spatio-temporalforecastingvisionfoundationmodelsmodelreprogrammingtemporaldifferenceflowcross-promptcoordinationfrozenbackboneurbanpredictiontransferlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a frozen vision foundation model—an image transformer pretrained only on natural pictures—can be turned into a general-purpose spatio-temporal forecaster by reprogramming its inputs rather than retraining it. The proposed framework, ST-VFM, feeds the model two streams: the raw spatio-temporal grid and a lightweight flow of first-order temporal differences that behave like motion cues in video. A Temporal-Aware Token Adapter embeds temporal context and aligns both streams to the vision model's token space, and a Bilateral Cross-Prompt Coordination module lets each stream condition the other through data-derived prompts. Across ten datasets spanning traffic, mobility, crowd flow, and cellular usage, the paper reports that ST-VFM outperforms every baseline on both MAE and RMSE, including models pretrained on spatio-temporal data. If correct, this matters because strong spatial priors from large-scale image pretraining could replace costly task-specific pretraining for urban forecasting.

What carries the argument

Two reprogramming stages carry the argument. Pre-VFM reprogramming is the Temporal-Aware Token Adapter: a Flow/ST Tokenizer patches the $H\times W$ grid at each time step, a Temporal Context Encoder runs a lightweight transformer over the time axis to inject temporal context, Token Adaptation linearly projects the tokens into the frozen VFM's embedding dimension, and a learnable positional embedding replaces interpolating the VFM's image-position embeddings. Post-VFM reprogramming is the Bilateral Cross-Prompt Coordination module, where each branch's VFM output supplies data-driven prompt tokens used as extra keys in the other branch's attention, so the ST branch is conditioned on flow-derived prompts and the flow branch on spatial prompts. Around these sit the two-branch input design (raw grid plus first-order temporal differences $\Delta_t X$), the frozen VFM backbone, and the joint loss $\mathcal{L} = \mathcal{L}_{ST} + \lambda\mathcal{L}_{Flow}$, which supervises both the future states and the future flow maps.

What would settle it

Take a benchmark whose dominant signal is a period longer than the six-step input window but shorter than the evaluation span (for example hourly data with daily periodicity and a one-day horizon), and run the paper's protocol: if ST-VFM's advantage over UniST shrinks or reverses compared with a variant that uses multi-step or optical-flow difference features, the first-order flow proxy is the part of the design that is not carrying the claimed generality.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that visual inductive biases transfer to spatio-temporal prediction once temporal information is encoded the way a vision model can read it. ST-VFM treats the historical grid $X \in \mathbb{R}^{C\times H\times W\times T}$ as a sequence of image-like frames and derives a second branch $\Delta_t X = X(t+1)-X(t)$ as a motion proxy, arguing by Taylor expansion of the brightness-constancy equation that the temporal derivative is a legitimate first-order stand-in for optical flow. The frozen VFM encodes both branches after pre-VFM reprogramming (patch tokenization, a lightweight temporal transformer shared across spatial locations, a linear token adaptation, and learnable positional embeddings), and a post-VFM module builds cross-prompt tokens from each branch's output so that the motion stream conditions the spatial stream and vice versa. The training objective couples the main forecast with an auxiliary forecast of the flow maps, weighted by $\lambda$. The paper reports this configuration outperforms all compared baselines on all ten datasets on both metrics, with an average improvement of more than 10% over the strongest baseline, and that the framework holds across DINO, CLIP, and DEIT backbones while beating a video foundation model adapted the same way.

Load-bearing premise

The load-bearing premise is that the difference between consecutive frames, $\Delta_t X = X(t+1)-X(t)$, carries enough of the temporal story for the frozen vision model to learn evolution; if a dataset's key dynamics are periodic cycles longer than the input window, slow trends, or movements that shift more than one cell per step, the flow branch may be nearly blind to them.

Editorial extensions

If this is right

  • Spatio-temporal forecasting no longer requires domain-specific pretraining: an image-pretrained frozen VFM with lightweight adapters beats models that were pretrained on urban data.
  • The same frozen backbone serves traffic, mobility, crowd-flow, and cellular-usage tasks, with the reprogramming modules as the only trainable components.
  • The flow branch plus cross-prompts gives the image model an explicit handle on change over time, so the framework's gains depend on temporal-difference signals being present in the data.
  • Across self-supervised (DINO), contrastive (CLIP), and supervised (DEIT) visual backbones, the method reports consistent gains over the strongest prior baseline.
  • A video foundation model adapted the same way trails the reprogrammed image VFMs, suggesting image-pretrained spatial priors plus explicit light-weight motion cues is the better division of labor.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The flow branch is only a first-order proxy; periodic or long-range dynamics invisible to adjacent-frame differences are a plausible failure mode, and replacing $\Delta_t X$ with multi-step differences or estimated optical flow would test how much of the gain is really motion-aware.
  • Because the VFM is frozen and the adapters are small, the same pipeline could be dropped onto other grid-structured forecasting problems (weather, energy, epidemiology) without retraining the backbone.
  • The cross-prompt attention weights are instance-specific and could be inspected as an interpretability tool for which spatial regions and which change signals drive a forecast.
  • If the reported margin over UniST generalizes beyond the ten benchmarks, reprogrammed image models would offer a cheaper alternative to LLM-based spatio-temporal forecasting, which the paper argues is limited by one-dimensional sequential modeling.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 7 minor

Summary. The paper proposes ST-VFM, a framework that adapts frozen vision foundation models (VFMs) to spatio-temporal forecasting. The method uses a dual-branch input: raw spatio-temporal tensors and first-order temporal-difference "flow" maps. A Temporal-Aware Token Adapter performs pre-VFM reprogramming by tokenizing and temporally contextualizing both branches, and a Bilateral Cross-Prompt Coordination module provides post-VFM interaction between branch outputs. The model is trained with a main ST forecasting loss plus an auxiliary flow-forecasting loss. Experiments on ten urban datasets report that ST-VFM outperforms all baselines on every dataset and both metrics (MAE and RMSE), with additional cross-backbone experiments (DINO, CLIP, DEIT), a comparison to a video foundation model, and ablations on five datasets.

Significance. If the empirical claims hold, the paper makes a useful contribution: it demonstrates a systematic recipe for transferring image-pretrained vision transformers to spatio-temporal forecasting without modifying the backbone, and it evaluates the approach on a broader set of urban forecasting tasks than is typical in this literature. The inclusion of strong recent baselines such as UniST and ST-LLM, and the cross-backbone analysis, are valuable. The main result is, however, entirely empirical and currently rests on a single unverified evaluation protocol: no code, no data splits, no seeds, no error bars, and no significance tests are provided, and some reported numbers and equations contain inconsistencies. The central design idea is interpretable and the direction of the reported gains is consistent, but the evidence as presented does not yet support the universal claim of superiority.

major comments (6)
  1. [Section 4.1 (Experimental Setup)] The paper states that each dataset is divided into training, validation, and test splits using a "standard 70%/15%/15% ratio" but never states that the split is chronological. All ten datasets are time-ordered urban sensing data with strong temporal autocorrelation, so a random split can place near-duplicate consecutive timesteps in both training and test sets, directly leaking the short-term dynamics that the model is asked to forecast. The reported margins over UniST are small in several cases (e.g., BikeNYC2 RMSE 5.42 vs. 5.50; TrafficSH RMSE 0.645 vs. 0.665), so the universal claim depends critically on split integrity. Please specify the exact chronological split boundaries or provide the split indices, and if the protocol follows UniST, cite and describe it precisely.
  2. [Section 3.4, Eq. (10)] The attention formula in Eq. (10) is dimensionally inconsistent as written. QST_i has dimension D, while the concatenated vector [KST_j, PFlow_j] has dimension 2D because both KST_j and PFlow_j are D-dimensional prompt/key vectors. The inner product QST_i · [KST_j, PFlow_j]^T is therefore undefined unless QST_i is also 2D or a projection is applied to the concatenated key. Additionally, the softmax numerator and denominator treat the concatenated term and the regular key term in a way that is not consistent with a standard attention weighting. Please correct the equation or define the projection that maps the concatenated keys into the query dimension; as written, this module cannot be implemented from the text.
  3. [Table 2, MAU row] The MAU row appears garbled: after "Cellular 39.09 18.73," the next entries "5.22 2.06 1.28 22.1 6.12 2.95 28.70 11.23 ..." do not align with the TDrive, BikeNYC, and BikeNYC2 columns. If this row is incorrect or shifted, it affects the reader's ability to verify the baseline comparison. The table is the sole evidence for the headline claim, so all baseline rows must be checked against the original papers or released code and corrected.
  4. [Section 4.2 and Table 2 (reproducibility and statistics)] No code, random seeds, data-split indices, or per-seed results are provided, and no statistical significance tests are reported. The paper claims ST-VFM "outperforms all baselines on all datasets and metrics," but several improvements over the strongest baseline are within 2–5% on RMSE, and without multiple runs or paired tests it is impossible to establish that these differences are not noise. Please provide code or a detailed implementation specification, run at least three to five seeds per configuration, and report mean ± std, together with a significance test on the main comparisons.
  5. [Section 4.3, Table 3 (ablation completeness)] The ablation study is reported on only five of the ten datasets, and the final ablation step adds the ST Flow input and the Bilateral Cross-Prompt Coordination module together, so the individual contribution of the cross-prompt module—one of the paper's two named contributions—is never isolated. The components should be ablated separately, and the ablation either extended to all datasets or justified with a statistical analysis showing that the five datasets are representative. As written, the claim that each component contributes to the final result is not fully supported by the evidence.
  6. [Section 3.3, Eq. (4) (motion proxy assumption)] The method relies on first-order temporal differences, ΔtX = X(t+1) − X(t), as a lightweight proxy for optical flow and as the sole vehicle for supplying temporal information to the VFM. The authors state that this "effectively captures motion magnitude and dynamic change patterns," but this assumption is not tested. It is plausible that periodic patterns, long-range trends, or sub-grid movements are poorly represented by adjacent-frame differences. Please compare the proposed proxy with alternatives (e.g., two-frame optical flow, multi-step differences, or normalized differences) on at least a subset of datasets, or provide an analysis of cases where the proxy is informative and where it is not.
minor comments (7)
  1. [Section 3.5, Eq. (11)] The loss weight λ is introduced but its value is not given in the main text; it appears only in supplementary sensitivity plots. Please state the value used for the main results.
  2. [Section 3.3 and References] The paper refers to "DINO (Oquab et al. 2023)" but the cited reference is DINOv2; please use consistent naming and citations throughout.
  3. [Table 4 and Supplementary Table 3] The VideoMAE baseline appears only in a separate comparison table and not in the main results table, and the supplementary CLIP/DEIT table has N/A entries for TDrive. Please clarify why these entries are N/A and report the complete comparison in the main text if VideoMAE is claimed as a serious alternative.
  4. [Section 4.1 (Implementation Details)] The description of the Adapter is ambiguous: the backbone is said to remain frozen, but "a lightweight Adapter ... is inserted into the VFM backbone to allow optional parameter-efficient fine-tuning, though it is not activated in all configurations." Since Table 3 includes a VFM+Adapter configuration, please state clearly whether the final model uses the Adapter and, if so, how it is integrated.
  5. [Section 3.3 (Learnable Positional Embedding)] The learnable positional embedding Epos is said to be "initialized independently," but the initialization scheme is not specified, and it is not stated whether the VFM's pretrained positional embeddings are discarded. Please report the initialization and its effect.
  6. [Figure 2] The cross-backbone comparison bars in Figure 2 have no error bars and are shown for only four datasets, despite the text claiming consistency; please include error bars or clarify that these are single runs.
  7. [Supplementary Material, Section 4] Several baseline descriptions in the supplementary contain unresolved citation placeholders (e.g., STResNet (?), ACFM (?), MC-STL (?)). These should be completed before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ST-VFM is an empirical architecture paper whose reported gains rest on held-out test evaluation against external baselines, not on self-referential definitions or fitted inputs renamed as predictions.

full rationale

The paper's central claim is an experimental performance comparison (Section 4.2, Table 2), not an analytic derivation. The dual-branch inputs—the raw ST tensor and the ST flow defined in Eq. 4 as Δ_t X = X(t+1) − X(t)—are computed from the historical window X_P only, so the target X_Q is not fed into the model as an input. The auxiliary flow loss L_Flow (supplementary Eq. 2) supervises a separate branch with future difference maps and is explicitly described as a regularizer, not as a mechanism that encodes the target into the input. No parameter is fitted to the test portion and then reported as a prediction; the validation split is used only for checkpoint selection, and hyperparameters such as λ and the number of prompt layers are fixed. The evaluation protocol is inherited from an external prior work (UniST, Yuan et al. 2024), not from the authors' own prior claims, and no uniqueness theorem or self-citation is invoked to justify the architecture. Concerns about the unspecified temporal ordering of the 70/15/15 split and the absence of seed-level variance bear on experimental validity and reproducibility, but they are not circularity: they do not make the reported output equivalent to the input by construction. The paper therefore receives a circularity score of 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical performance claim, so the main free parameters are training hyperparameters and the learned token/position embeddings. The key domain assumptions are that grid frames behave like images and that frame differencing carries enough temporal information. No new physical or conceptual entities are postulated; ST Flow is a temporal difference map, not a new force or particle.

free parameters (5)
  • lambda (flow loss weight) = not stated (sensitivity tested over 0.001 to 0.1)
    Controls the contribution of the auxiliary flow forecasting loss in Eq. 11; the paper reports stable performance across a range but never gives the final value used for the main results.
  • Patch size Ps = 2
    Chosen for all datasets; affects tokenization granularity and is not ablated.
  • Number of cross-prompt layers = 2 (1 for BikeNYC and TaxiNYC)
    Set smaller on small datasets; not ablated.
  • Number of Temporal Transformer layers = 3
    Used in both pre-VFM and post-VFM stages.
  • Learnable positional embedding E_pos = learned
    Learned from scratch for each dataset; not inherited from the VFM.
assumptions (4)
  • domain assumption Spatio-temporal grid frames can be treated as images for a vision transformer.
    The entire method rests on the structural analogy X in R^(C x H x W x T) being image-like; stated in Section 3.2 and Figure 1.
  • domain assumption First-order temporal differences (ST Flow) are a sufficient proxy for temporal dynamics.
    Eq. (4) defines delta_t X in place of full optical flow; the paper admits it omits displacement direction but relies on it as the temporal signal.
  • domain assumption The frozen VFM's spatial priors transfer to non-image grid data after a linear token projection.
    Token Adaptation (Section 3.3) projects ST data into the VFM token space without a modality-alignment loss; the success of the framework depends on this transfer.
  • standard math Brightness constancy and first-order Taylor expansion for optical flow are used only as motivation.
    The paper derives the flow constraint in Eqs. (2)-(3) but does not actually solve for u and v; it only uses the temporal derivative term.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reprogramming Vision Foundation Models for Spatio-Temporal Forecasting." pith.science (2026). https://pith.science/paper/N3LLRDTY

@misc{pith2026250711558,
  author       = {Pith},
  title        = {Pith review of: Reprogramming Vision Foundation Models for Spatio-Temporal Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N3LLRDTY}},
  note         = {Machine review of arXiv:2507.11558}
}
read the original abstract

Foundation models have achieved remarkable success in natural language processing and computer vision, demonstrating strong capabilities in modeling complex patterns. While recent efforts have explored adapting large language models (LLMs) for time-series forecasting, LLMs primarily capture one-dimensional sequential dependencies and struggle to model the richer spatio-temporal (ST) correlations essential for accurate ST forecasting. In this paper, we present \textbf{ST-VFM}, a novel framework that systematically reprograms Vision Foundation Models (VFMs) for general-purpose spatio-temporal forecasting. While VFMs offer powerful spatial priors, two key challenges arise when applying them to ST tasks: (1) the lack of inherent temporal modeling capacity and (2) the modality gap between visual and ST data. To address these, ST-VFM adopts a \emph{dual-branch architecture} that integrates raw ST inputs with auxiliary ST flow inputs, where the flow encodes lightweight temporal difference signals interpretable as dynamic spatial cues. To effectively process these dual-branch inputs, ST-VFM introduces two dedicated reprogramming stages. The \emph{pre-VFM reprogramming} stage applies a Temporal-Aware Token Adapter to embed temporal context and align both branches into VFM-compatible feature spaces. The \emph{post-VFM reprogramming} stage introduces a Bilateral Cross-Prompt Coordination module, enabling dynamic interaction between branches through prompt-based conditioning, thus enriching joint representation learning without modifying the frozen VFM backbone. Extensive experiments on ten spatio-temporal datasets show that ST-VFM outperforms state-of-the-art baselines, demonstrating effectiveness and robustness across VFM backbones (e.g., DINO, CLIP, DEIT) and ablation studies, establishing it as a strong general framework for spatio-temporal forecasting.

Figures

Figures reproduced from arXiv: 2507.11558 by the authors.

Figure 1
Figure 1. The overall framework of ST-VFM. ST-VFM takes two input sources: spatio-temporal input and ST flow. In the Pre-VFM Reprogramming stage, a Temporal-Aware Token Adapter projects both inputs into a VFM-compatible token space, embedding temporal context while aligning modalities. The frozen VFM then encodes these adapted ST and flow tokens. In the Post-VFM Reprogramming stage, a Bilateral Cross-Prompt Coordination modul… view at source ↗
Figure 2
Figure 2. Comparison of different pretrained VFMs (DINO, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 1
Figure 1. Sensitivity w.r.t. λ on Cellular and BikeNYC. Training (min) Inference (min) RMSE MAE UniST 1.06 0.0423 3.00 1.38 # 4 1.08 0.0432 2.65 1.17 Ours (full) 2.06 0.0745 2.51 1.06 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 31 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, L.; Yao, L.; Li, C.; Wang, X.; and Wang, C. 2020. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in NeurIPS, 33: 17804--17815

  4. [4]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in NeurIPS, 33: 1877--1901

  5. [5]

    O.; Pfister, T.; Zheng, Y.; Ye, W.; and Liu, Y

    Cao, D.; Jia, F.; Arik, S. O.; Pfister, T.; Zheng, Y.; Ye, W.; and Liu, Y. 2023. Tempo: Prompt-based generative pre-trained transformer for time series forecasting. arXiv preprint arXiv:2310.04948

  6. [6]

    Caron, M.; Touvron, H.; Misra, I.; J \'e gou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the international conference on computer vision, 9650--9660

  7. [7]

    Chang, Z.; Zhang, X.; Wang, S.; Ma, S.; Ye, Y.; Xinguang, X.; and Gao, W. 2021. Mau: A motion-aware unit for video prediction and beyond. Advances in NeurIPS, 34: 26950--26962

  8. [8]

    Chung, J.; Gulcehre, C.; Cho, K.; and Bengio, Y. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555

Show all 49 references
  1. [9]

    Deng, J.; Chen, X.; Jiang, R.; Song, X.; and Tsang, I. W. 2021. St-norm: Spatial and temporal normalization for multi-variate time series forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 269--278

  2. [10]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

  3. [11]

    Gao, Z.; Tan, C.; Wu, L.; and Li, S. Z. 2022. Simvp: Simpler yet better video prediction. In Proceedings of the conference on CVPR, 3170--3180

  4. [12]

    Houlsby, N.; Giurgiu, A.; Jastrzebski, S.; Morrone, B.; De Laroussilhe, Q.; Gesmundo, A.; Attariyan, M.; and Gelly, S. 2019. Parameter-efficient transfer learning for NLP. In International conference on machine learning, 2790--2799. PMLR

  5. [13]

    Huang, Y.; Mao, X.; Guo, S.; Chen, Y.; Shen, J.; Li, T.; Lin, Y.; and Wan, H. 2024. STD-PLM: Understanding Both Spatial and Temporal Properties of Spatial-Temporal Data with PLM. arXiv preprint arXiv:2407.09096

  6. [14]

    Ji, J.; Wang, J.; Huang, C.; Wu, J.; Xu, B.; Wu, Z.; Zhang, J.; and Zheng, Y. 2023. Spatio-temporal self-supervised learning for traffic flow prediction. In Proceedings of the AAAI, volume 37, 4356--4364

  7. [15]

    X.; and Wang, J

    Jiang, J.; Han, C.; Zhao, W. X.; and Wang, J. 2023. Pdformer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction. In Proceedings of the AAAI, volume 37, 4365--4373

  8. [16]

    Y.; Shi, X.; Chen, P.-Y.; Liang, Y.; Li, Y.-F.; Pan, S.; et al

    Jin, M.; Wang, S.; Ma, L.; Chu, Z.; Zhang, J. Y.; Shi, X.; Chen, P.-Y.; Liang, Y.; Li, Y.-F.; Pan, S.; et al. 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728

  9. [17]

    Jin, Y.; Chen, K.; and Yang, Q. 2022. Selective cross-city transfer learning for traffic prediction via source city region re-weighting. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 731--741

  10. [18]

    Li, Y.; Yu, R.; Shahabi, C.; and Liu, Y. 2017. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926

  11. [19]

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

  12. [20]

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

  13. [21]

    Liu, C.; Yang, S.; Xu, Q.; Li, Z.; Long, C.; Li, Z.; and Zhao, R. 2024 a . Spatial-temporal large language model for traffic prediction. In 2024 25th IEEE International Conference on Mobile Data Management (MDM), 31--40. IEEE

  14. [22]

    Liu, L.; Yu, S.; Wang, R.; Ma, Z.; and Shen, Y. 2024 b . How can large language models understand spatial-temporal data? arXiv preprint arXiv:2401.14192

  15. [23]

    Liu, L.; Zhang, R.; Peng, J.; Li, G.; Du, B.; and Lin, L. 2018. Attentive crowd flow machines. In Proceedings of the 26th ACM international conference on Multimedia, 1553--1561

  16. [24]

    Liu, Y.; Hu, T.; Zhang, H.; Wu, H.; Wang, S.; Ma, L.; and Long, M. 2023. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625

  17. [25]

    Lu, B.; Gan, X.; Zhang, W.; Yao, H.; Fu, L.; and Wang, X. 2022. Spatio-temporal graph few-shot learning with cross-city knowledge transfer. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 1162--1172

  18. [26]

    D.; and Kanade, T

    Lucas, B. D.; and Kanade, T. 1981. An iterative image registration technique with an application to stereo vision. In IJCAI'81: 7th international joint conference on Artificial intelligence, volume 2, 674--679

  19. [27]

    H.; Sinthong, P.; and Kalagnanam, J

    Nie, Y.; Nguyen, N. H.; Sinthong, P.; and Kalagnanam, J. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730

  20. [28]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

  21. [29]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

  22. [30]

    Shao, Z.; Zhang, Z.; Wang, F.; Wei, W.; and Xu, Y. 2022. Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecasting. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, 4454--4458

  23. [31]

    C.; See, S.; Qin, H.; Dai, J.; and Li, H

    Shi, X.; Huang, Z.; Bian, W.; Li, D.; Zhang, M.; Cheung, K. C.; See, S.; Qin, H.; Dai, J.; and Li, H. 2023. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In Proceedings of the International Conference on Computer Vision, 12469--12480

  24. [32]

    Simonyan, K.; and Zisserman, A. 2014. Two-stream convolutional networks for action recognition in videos. Advances in NeurIPS, 27

  25. [33]

    Tan, C.; Gao, Z.; Wu, L.; Xu, Y.; Xia, J.; Li, S.; and Li, S. Z. 2023. Temporal attention unit: Towards efficient spatiotemporal predictive learning. In Proceedings of the Conference on CVPR, 18770--18782

  26. [34]

    Tong, Z.; Song, Y.; Wang, J.; and Wang, L. 2022. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35: 10078--10093

  27. [35]

    Touvron, H.; Cord, M.; Douze, M.; Massa, F.; Sablayrolles, A.; and J \'e gou, H. 2021. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, 10347--10357. PMLR

  28. [36]

    Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.; Babaei, Y.; Bashlykov, N.; Batra, S.; Bhargava, P.; Bhosale, S.; et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

  29. [37]

    Wang, B.; Lin, Y.; Guo, S.; and Wan, H. 2021. GSNet: Learning Spatial-Temporal Correlations from Geographical and Semantic Aspects for Traffic Accident Risk Forecasting. In AAAI

  30. [38]

    Wang, S.; Cao, J.; and Philip, S. Y. 2020. Deep learning for spatio-temporal data mining: A survey. IEEE transactions on knowledge and data engineering, 34(8): 3681--3700

  31. [39]

    Wang, Y.; Long, M.; Wang, J.; Gao, Z.; and Yu, P. S. 2017. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in NeurIPS, 30

  32. [40]

    Wang, Y.; Zhang, J.; Zhu, H.; Long, M.; Wang, J.; and Yu, P. S. 2019. Memory in memory: A predictive neural network for learning higher-order non-stationarity from spatiotemporal dynamics. In Proceedings of the conference on CVPR, 9154--9162

  33. [41]

    Yan, Y.; Wen, H.; Zhong, S.; Chen, W.; Chen, H.; Wen, Q.; Zimmermann, R.; and Liang, Y. 2024. Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web. In Proceedings of the ACM on Web Conference 2024, 4006--4017

  34. [42]

    Yu, B.; Yin, H.; and Zhu, Z. 2017. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875

  35. [43]

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

  36. [44]

    Zhang, J.; Zheng, Y.; and Qi, D. 2017. Deep spatio-temporal residual networks for citywide crowd flows prediction. In Proceedings of the AAAI, volume 31

  37. [45]

    Zhang, X.; Gong, Y.; Zhang, X.; Wu, X.; Zhang, C.; and Dong, X. 2023. Mask-and contrast-enhanced spatio-temporal learning for urban flow prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 3298--3307

  38. [46]

    Zhao, L.; Gao, M.; and Wang, Z. 2022. St-gsp: Spatial-temporal global semantic representation learning for urban flow prediction. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, 1443--1451

  39. [47]

    Zheng, C.; Fan, X.; Wang, C.; and Qi, J. 2020. Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI, volume 34, 1234--1241

  40. [48]

    Zhou, T.; Niu, P.; Sun, L.; Jin, R.; et al. 2023 a . One fits all: Power general time series analysis by pretrained lm. Advances in NeurIPS, 36: 43322--43355

  41. [49]

    Zhou, Z.; Yang, K.; Liang, Y.; Wang, B.; Chen, H.; and Wang, Y. 2023 b . Predicting collective human mobility via countering spatiotemporal heterogeneity. IEEE Transactions on Mobile Computing

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.