UrbanFM: Scaling Urban Spatio-Temporal Foundation Models

Wei Chen; Yuqian Wu; Junle Chen; Xiaofang Zhou; Yuxuan Liang

cs.LGFeb 24, 2026

UrbanFM: Scaling Urban Spatio-Temporal Foundation Models

Authors: Wei Chen, Yuqian Wu, Junle Chen, Xiaofang Zhou, Yuxuan Liang

Organizations: HKUST(GZ) · HKUST

Abstract

Urban systems, as dynamic complex systems, continuously generate spatio-temporal data streams that encode the fundamental laws of human mobility and city evolution. While AI for Science has witnessed the transformative power of foundation models in disciplines like genomics and meteorology, urban computing remains fragmented due to "scenario-specific" models, which are overfitted to specific regions or tasks, hindering their generalizability. To bridge this gap and advance spatio-temporal foundation models for urban systems, we adopt scaling as the central perspective and systematically investigate two key questions: what to scale and how to scale. Grounded in first-principles analysis, we identify three critical dimensions: heterogeneity, correlation, and dynamics, aligning these principles with the fundamental scientific properties of urban spatio-temporal data. Specifically, to address heterogeneity through data scaling, we construct WorldST. This billion-scale corpus standardizes diverse physical signals, such as traffic flow and speed, from over 100 global cities into a unified data format. To enable computation scaling for modeling correlations, we introduce the MiniST unit, a novel split mechanism that discretizes continuous spatio-temporal fields into learnable computational units to unify representations of grid-based and sensor-based observations. Finally, addressing dynamics via architecture scaling, we propose UrbanFM, a minimalist self-attention architecture designed with limited inductive biases to autonomously learn dynamic spatio-temporal dependencies from massive data. Furthermore, we establish EvalST, the largest-scale urban spatio-temporal benchmark to date. Extensive experiments demonstrate that UrbanFM achieves remarkable zero-shot generalization across unseen cities and tasks, marking a pivotal first step toward large-scale urban spatio-temporal foundation models.

Figures & tables

Figure 1: Top: The fundamental nature of spatio-temporal data and challenges of urban ST foundational models. Bottom: Three core perspectives of scaling the urban ST foundation model: data, computation, and architecture.

Figure 2: Data Overview: Breakthroughs in multi-domain coverage and spatio-temporal scale. Notably, we far surpass UniST ( Yuan et al., 2024a ) , OpenCity ( Li et al., 2025 ) , and BigCity ( Yu et al., 2025 ) in terms of spatial regions and temporal spans, by as much as 33 to 145 times.

Figure 3: A schematic diagram of the scaling mechanism: MiniST tokenization and UrbanFM model.

Figure 4: Zero-shot forecasting effectiveness on various spatial-temporal benchmarks (Full results in Table D.1 , D.2 , D.3 , and D.4 ).

Figure 5: Evaluation of UrbanFM ’s performance gain through few-shot tuning and comparison with full-shot expert models.

Figure 6: Performance evaluation of UrbanFM and classic methods for spatio-temporal point and block data imputation tasks.

Figure 7: Scaling law analysis across model depth, data volume. (Full results in Table D.6 and D.6 ).

Figure 8: Scaling law analysis of spatio-temporal resolutions (Full results in Table D.8 and D.8 ).

Figure 9: A visualization study of the robustness of predictions in noisy environments using two datasets.

Figure 10: Efficiency study of our method.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Data Format	Domain Type	Dataset Name	Spatial Region	# Spatial Location	Temporal Range	# Temporal Step	Overlap
Sensor -based	Flow	PEMS-03	Los Angles, US	358	2018/09/01 – 2018/11/30	26208	$\times$
		PEMS-04	Los Angles, US	307	2018/01/01 – 2018/02/28	16992	$\times$
		PEMS-07	Los Angles, US	883	2017/05/01 – 2017/08/06	28224	$\times$
		PEMS-08	Los Angles, US	170	2016/07/01 – 2016/08/31	17856	$\times$
	Occupancy	OCC-Pairs	Pairs, FR	106	2016/01/01 – 2016/12/01	96410	$\times$
	Occupancy	OCC-Hamburg	Hamburg, DE	240	2016/08/27 – 2016/12/09	30085	$\times$

Appendix

Table A.1 : Evaluation Benchmark Summary

Panel of Figure 12 — Figure A.1 : Spatial Visualization of EvalST Benchmark.

Data Format	Domain Type	Spatial Region	# Spatial Location	Temporal Range	# Temporal Step
Sensor	Flow	Augsburg, DE	40	2017/05/06 – 2017/05/25	5757
Sensor	Flow	Basel, CH	66	2016/10/24 – 2016/10/31	2016
Sensor	Flow	Bern, CH	708	2016/10/24 – 2016/10/31	2016
Sensor	Flow	Birmingham, GB	34	2017/10/23 – 2017/11/18	7540
Sensor	Flow	Bolton, GB	56	2017/11/13 – 2017/11/18	1491
Sensor	Flow	Bordeaux, FR	409	2016/11/21 – 2016/11/27	2016

Appendix

Table A.2 : Pre-training Datasets Summary

Figure A.2 : Temporal Visualization of EvalST Benchmark.

Figure A.3 : Spatio-Temporal Visualization of EvalST Benchmark.

Figure B.1 : Visualization of greedy capacity-constrained clustering with different $S_{p}$ on PEMS-08 and Taxi-NYC.

Zero-Shot of Foundation Model

Few-Shot (10% Training Set) of Expert Model

Method Type

ST-Graph Method

ST-Grid Method

Time Series Method

Method

UrbanFM

FactoST

OpenCity

TimesFM++

Moirai

TimeMoE

Chronos

STAEformer

STID

D2STGNN

GWNet

TAU

PredRNN++

DSAN

ST-ResNet

PatchTST

Dlinear

LSTM

Traffic-SH

Short

MAE

0.44

0.40

0.45

0.49

0.46

0.43

0.45

1.26{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.06}}

0.47{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.39{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.37{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.66{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

1.11{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.50{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.31{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.46{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.36{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.53{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

RMSE

0.61

0.70

0.82

1.14

0.91

0.74

0.80

1.87{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.08}}

0.77{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.69{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.69{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.98{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.05}}

1.02{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

1.84{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

2.11{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.61{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.79{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.71{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.92{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

MAPE(%)

5.39

5.84

7.21

7.31

8.76

6.65

6.77

18.45{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.12}}

7.96{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.14}}

6.33{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.10}}

6.21{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.06}}

9.41{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.30}}

11.03{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.12}}

10.74{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

17.16{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.17}}

5.47{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

8.18{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

6.44{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

7.76{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Long

MAE

0.49

0.58

0.48

0.65

0.66

0.60

0.67

2.94{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.52}}

0.60{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.68{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.70{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

0.80{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

1.21{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.54{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.48{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.59{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.51{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.53{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Appendix

Table D.1 : Comparison of the zero-shot results of the foundation models with the few-shot results of the expert models on the short-term (

12\rightarrow 12_{average}

) and the long-term (

24\rightarrow 24_{average}

) forecasting based on the grid dataset benchmark. indicates best, indicates second best.

Zero-Shot of Foundation Model

Few-Shot (10% Training Set) of Expert Model

Method Type

ST-Graph Method

Time Series Method

Model

UrbanFM

FactoST

OpenCity

TimesFM++

Moirai

TimeMoE

Chronos

STAEformer

STID

D2STGNN

GWNet

PatchTST

Dlinear

LSTM

PEMS-03

Short

MAE

15.62

25.84

25.57

27.13

29.02

29.12

26.48

20.07{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.39}}

17.88{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.12}}

17.78{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.17}}

17.96{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.68}}

21.04{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.07}}

21.17{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

19.08{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.38}}

32.01{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

RMSE

23.24

40.26

49.25

40.94

44.11

43.36

41.10

32.35{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.38}}

28.50{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.31}}

29.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.59}}

28.82{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.89}}

32.92{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.09}}

33.26{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

31.04{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.71}}

55.52{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

MAPE(%)

15.24

24.91

23.86

25.19

30.76

38.87

24.96

20.26{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.95}}

19.35{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.21}}

20.51{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.89}}

21.09{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 3.11}}

19.28{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

20.88{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.26}}

19.53{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.56}}

31.76{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Long

MAE

17.03

40.31

27.85

31.37

47.19

33.32

38.66

22.72{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 1.00}}

22.23{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.10}}

21.43{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.44}}

25.70{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.53}}

28.63{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

29.05{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

23.20{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.08}}

31.89{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Appendix

Table D.2 : Comparison of the zero-shot results of the foundation models with the few-shot results of the expert models on the short-term (

12\rightarrow 12_{average}

) and the long-term (

24\rightarrow 24_{average}

) forecasting based on the sensor dataset benchmark. indicates best, indicates second best. OOM indicates Out-Of-Memory (NVIDIA A100-80G).

Few-Shot (10% Training Set) of Foundation Model

Full-Shot of Expert Model

Method Type

ST-Graph Method

Time Series Method

Model

UrbanFM

FactoST

OpenCity

TimesFM++

Moirai

TimeMoE

Chronos

STAEformer

STID

D2STGNN

GWNet

PatchTST

Dlinear

LSTM

PEMS-03

Short

MAE

13.00

21.08

16.91

21.56

20.54

20.59

21.31

15.13{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.10}}

15.21{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

14.88{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.11}}

14.91{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.12}}

20.84{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

21.04{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

16.69{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.06}}

26.17{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

RMSE

19.87

32.88

27.18

33.72

31.68

31.98

33.72

26.84{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.56}}

25.52{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.28}}

25.76{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.23}}

25.82{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.30}}

32.57{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

33.08{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

28.21{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.20}}

47.52{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

MAPE(%)

12.43

19.28

18.57

19.61

19.63

22.11

19.21

15.33{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.12}}

16.36{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.13}}

15.02{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.28}}

15.19{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.42}}

19.13{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

20.78{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.09}}

16.05{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.27}}

26.89{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Long

MAE

13.82

28.62

19.29

30.08

28.74

26.59

29.81

17.30{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.34}}

17.07{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.06}}

16.45{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.09}}

20.37{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.39}}

27.84{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.14}}

28.66{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

19.57{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.03}}

26.19{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Appendix

Table D.3 : Comparison of the few-shot results of the foundation models with the full-shot results of the expert models on the short-term (

12\rightarrow 12_{average}

) and the long-term (

24\rightarrow 24_{average}

) forecasting based on the sensor dataset benchmark. indicates best, indicates second best.

Few-Shot (10% Training Set) of Foundation Model

Full-Shot of Expert Model

Method Type

ST-Graph Method

ST-Grid Method

Time Series Method

Method

UrbanFM

FactoST

OpenCity

TimesFM++

Moirai

TimeMoE

Chronos

STAEformer

STID

D2STGNN

GWNet

TAU

PredRNN++

DSAN

ST-ResNet

PatchTST

Dlinear

LSTM

Traffic-SH

Short

MAE

0.10

0.34

0.44

0.39

0.34

OOM

0.34

0.24{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.25{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.24{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.27{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.40{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.32{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.04{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.09{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.30{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.43{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.31{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.92{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

RMSE

0.26

0.62

0.77

0.69

0.62

OOM

0.64

0.57{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.50{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

0.74{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.62{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.80{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.80{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.61{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.76{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.62{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

1.22{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

MAPE(%)

1.47

5.25

7.33

6.14

5.31

OOM

5.29

4.41{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

4.23{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.08}}

4.38{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.08}}

4.78{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.18}}

6.86{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.05}}

5.65{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.09}}

9.58{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

10.70{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.04}}

5.34{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.05}}

7.72{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.02}}

5.57{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.06}}

12.76{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Long

MAE

0.20

0.54

0.48

0.57

OOM

0.55

0.33{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.35{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.32{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.48{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.44{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.41{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

1.16{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

1.13{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.01}}

0.47{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.56{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.41{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

0.93{\color[rgb]{0.5,0.5,0.5}\text{\scriptsize\textpm 0.00}}

Appendix

Table D.4 : Comparison of the few-shot results of the foundation models with the full-shot results of the expert models on the short-term (

12\rightarrow 12_{average}

) and the long-term (

24\rightarrow 24_{average}

) forecasting based on the grid dataset benchmark. indicates best, indicates second best.

Table 21

Table 22

Geospatial Foundation Models

Geospatial foundation models turn satellite, map, and Earth-system data into reusable representations for spatial analysis and prediction.

See all 123 papers

Jul 7, 2026

WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence

Humans can navigate an unfamiliar city and gradually form a coherent spatial mental map spanning tens of square kilometers. Can AI build spatial representations at a comparable scale? Although recent foundation models have advanced scene reconstruction and embodied intelligence, scaling to entire cities remains an open challenge, primarily due to the lack of city-scale data. To bridge the gap, we introduce WildCity, a real-world multimodal dataset collected by autonomous fleets traversing complex urban environments. Our dataset includes 18 trajectories, each averaging 83.7 kilometers in length, and preserves the core challenges of in-the-wild perception, e.g., dynamic objects, lighting variations, and imperfect camera poses. We further establish an urban-tailored reconstruction baseline and convert the reconstructed environments into a closed-loop simulator. Beyond the dataset and baseline, we systematically analyze the key challenges on the path to simulation-ready urban digital twins: scalability, extrapolation, and uncertainty. Ultimately, WildCity aims to catalyze progress not only in city-scale rendering, but more broadly in the pursuit of AI that can perceive, remember, and reason across space at a scale comparable to human cognition. Project page: https://han-xiangyu.github.io/Wild-City/

Xiangyu Han, Mengyu Yang, Jiaqi Li +9

May Mobility, Ann Arbor, MI, USA · New York University, New York, NY, USA · Stanford University, Stanford, CA, USA +1

Autonomous Driving Datasets 3D Reconstruction Dynamic Scene Reconstruction Digital Twins Autonomous Driving Simulation Urban Scene Reconstruction

Ask PDF

Oct 17, 2026

$M^2Net$ : Unifying Scale-Specific Structures and Prototype Decomposition for Spatial-Temporal Forecasting

Spatial-temporal forecasting is pivotal for intelligent urban management yet remains challenging due to the intricate coupling of spatial dynamics and temporal non-stationarity. Existing approaches, while effective, often overlook two fundamental characteristics of real-world data: (1) Granularity-dependence, where dominant spatial correlations shift significantly across different time scales, and (2) Signal entanglement, where global common patterns are inextricably mixed with node-specific disturbances. In this paper, we propose a novel framework named

M^2Net

(Multi-Scale Masked attention and Prototype network) to unify scale-specific structural learning and semantic decomposition. First, to capture the evolution of spatial dependencies, we introduce a Multi-Scale Masked attention mechanism. By downsampling temporal features into multiple resolutions, we dynamically generate scale-adaptive spatial masks, allowing the model to distinguish between micro-scale physical neighbors and macro-scale functional correlations. Second, to disentangle unique variations from shared trends, we propose a Prototype Residual Learning module. It retrieves global common patterns from a momentum-updated prototype bank and subtracts them from node representations, explicitly forcing the network to model the residual unique deviations. By integrating these two orthogonal views, i.e., structural multiscaling and semantic decomposition, our approach effectively isolates valid spatial-temporal signals from noise. Extensive experiments on real-world datasets demonstrate that

M^2Net

outperforms state-of-the-art baselines on extensive empirical verifications.

Chengyang Zhou, Zijian Zhang, Hao Miao +5

College of Computer Science and Technology, Jilin University, Changchun, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China +2

Spatiotemporal Forecasting Multiscale Time-Series Modeling Spatial Context Modeling Spatio-Temporal Attention Prototype Learning

Ask PDF

May 25, 2026

CITYREP: A Unified Benchmark for Urban Representations Across Cities, Tasks, and Modalities

Urban representation learning encodes complex urban environments into general-purpose embeddings for diverse downstream tasks and emerging urban foundation models. However, current evaluations are limited, typically focusing on one or two cities and tasks and relying on random splits that introduce spatial leakage, leading to inflated performance and weak support for cross-location generalization and fair comparison. To address this, we propose CityRep, a unified benchmark that evaluates urban representations across data modalities, cities, and tasks using spatially structured splits. CityRep consists of three key components: (1) a spatial unit-agnostic evaluation framework that supports heterogeneous urban representations through a standardized alignment module; (2) a unified evaluation protocol using block-based spatial splits to mitigate spatial leakage and enable rigorous model comparison; and (3) an extensible multi-city, multi-task benchmark suite spanning 8 cities and 8 tasks across regression, classification, and distribution prediction. We evaluate 11 representative urban representation models. Results show that performance is highly sensitive to the split protocol, with random splits inflating scores and altering model rankings. We also observe substantial variability across cities and tasks, underscoring the need for generalization-aware evaluation. CityRep is released as a reproducible benchmark with datasets, evaluation pipelines, and diagnostic tools to facilitate fair comparison and support future research in urban representation learning towards urban foundation models.

Junyuan Liu, Xinglei Wang, Zichao Zeng +5

SpaceTimeLab, University College London, UK · 3DIMPact, University College London, UK · School of Resource and Environmental Sciences, Wuhan University, China +1

Representation Learning Geospatial Representation Learning Data Contamination Urban Representation Learning

Ask PDF

Abstract

Geospatial Foundation Models

Explore similar work

WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence

M2NetM^2NetM2Net: Unifying Scale-Specific Structures and Prototype Decomposition for Spatial-Temporal Forecasting

CITYREP: A Unified Benchmark for Urban Representations Across Cities, Tasks, and Modalities

$M^2Net$ : Unifying Scale-Specific Structures and Prototype Decomposition for Spatial-Temporal Forecasting