Plan-to-Synthesis: Cross-City Human Mobility Generation via Semantic Latent Flow Matching

Zhoufu Wang; Baoshen Guo; Zhiqing Hong; Junyi Li; Kailai Sun; Heye Huang; Alok Prakash; Shenhao Wang

cs.SISep 26, 2026

Plan-to-Synthesis: Cross-City Human Mobility Generation via Semantic Latent Flow Matching

Authors: Zhoufu Wang, Baoshen Guo, Zhiqing Hong, Junyi Li, Kailai Sun, Heye Huang, Alok Prakash, Shenhao Wang,

Organizations: Singapore-MIT Alliance for Research and Technology Centre, MIT · Nanyang Technological University · Hong Kong University of Science and Technology (Guangzhou) · Korea Advanced Institute of Science & Technology ·

Abstract

Human mobility generation aims to synthesize realistic point-of-interest (POI) visitation trajectories and has become an important tool for travel behavior modeling, transportation management, and urban planning. Existing diffusion-based methods achieve high fidelity but require per-city generation, given the inherent heterogeneity of geospatial locations and POI categories, while large language model-based methods generalize across cities but remain too costly at scale, especially for long-horizon trajectory generation. To address this, we propose SeMoFlow, a Semantic human Mobility generation framework based on latent Flow matching. We first encode heterogeneous POIs from different cities into a shared cross-city representation space via hierarchical Semantic IDs, where shared prefixes capture transferable semantics, and successive codes progressively refine the representation toward individual POIs. Building on the semantic IDs, SeMoFlow follows a plan-to-synthesis hierarchical generation paradigm, in which an autoregressive planner generates coarse-grained semantic and recurrence patterns, and a flow matching realizer synthesizes fine-grained suffix latents. The generated latents are subsequently decoded and grounded to concrete POIs. Extensive experiments on large-scale multi-city datasets show that SeMoFlow achieves higher trajectory fidelity than existing baselines, preserves city-specific mobility motifs, and supports both joint multi-city generation and effective cross-city transfer. The code is available at https://github.com/urban-mobility-generation/SeMoFlow.

Figures & tables

Figure 1: (a) A trajectory admits a coarse function or category sequence and a fine POI sequence. Residual quantization aligns these two levels with Semantic IDs, where a shared prefix encodes the coarse-grained semantics (e.g., categories) and a local suffix resolves the concrete POI (e.g., two food visits share the prefix but differ in their suffix). (b) An individual trajectory over space and time, where a set of POIs is visited repeatedly. (c) Mean cumulative repeat visits per individual, and recurrence is consistently stronger at the function and category levels than at the POI level.

Figure 2: Overview of SeMoFlow . (a) A semantic tokenizer encodes POI features into hierarchical Semantic IDs. (b) An autoregressive planner generates time gaps and SID prefixes to form a coarse-grained trajectory plan. (c) A conditional latent flow realizer learns to denoise corrupted suffix latents and generates suffix representations. (d) The generated suffix latents are decoded into full SIDs under the planned prefixes and subsequently grounded to concrete POIs.

Figure 3: Hierarchical SID tokenization of POIs. Shared prefix codes capture transferable activity semantics, while successive suffix codes refine the representation toward individual POI identities.

City	Model	$JSD_{\mathrm{POI}}\downarrow$	$JSD_{\mathrm{geo}}\downarrow$	$JSD_{\mathrm{travel}}\downarrow$	$JSD_{\mathrm{gyr}}\downarrow$	$JSD_{\mathrm{motif}}\downarrow$	$JSD_{\mathrm{recur}}\downarrow$
Chicago	MoveSim	0.1115	0.0087	0.0479	0.1346	0.4227	0.3399
	TimeGeo	0.3286	0.0220	0.3240	0.2905	0.0810	0.0640
	MIRAGE	0.1136	0.0091	0.0711	0.0709	0.0071	0.0257
	MobiDiff	0.6187	0.4815	0.0167	0.0265	0.0187	0.0275
	Traveller	0.4096	0.1329	0.1909	0.2339	0.0037	0.0073
	SeMoFlow (Per-City)	0.0943	0.0052	0.0172	0.0018	0.0025	0.0020

Table 1: Comparison of trajectory generation performance across four cities. Bold and underline indicate the best and second-best values within each city block, respectively.

Figure 4: Cross-city adaptation to Dallas using different percentages of target-city data.

Variant	$JSD_{\mathrm{POI}}\downarrow$	$JSD_{\mathrm{geo}}\downarrow$	$JSD_{\mathrm{travel}}\downarrow$	$JSD_{\mathrm{gyr}}\downarrow$	$JSD_{\mathrm{motif}}\downarrow$	$JSD_{\mathrm{recur}}\downarrow$
w/o Graph	0.0978	0.0049	0.2133	0.2798	0.0026	0.0056
w/o Planner	0.4093	0.0409	0.0745	0.0304	0.0055	0.0052
w/o Router	0.0943	0.0054	0.0184	0.0019	0.0048	0.0060
SeMoFlow	0.0943	0.0052	0.0172	0.0018	0.0025	0.0020

Table 2: Ablation studies of SeMoFlow on Chicago dataset. Bold indicates the best result.

Figure 5: Hierarchical SIDs in Chicago. (a) Coarse-to-fine spatial distributions across six SID levels. (b) Mean POIs per prefix and category purity at each depth. (c) An illustrative example in which the prefix captures a dominant POI category and subsequent codes progressively distinguish individual POIs.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	City	# POIs	# Weekly Segments	Length Range ( $L$ )	Experimental Settings
Veraset	Boston	122,731	390,148	$5\leq L\leq 128$	Per-City / Multi-City training
	Chicago	206,852	679,417	$5\leq L\leq 128$	Per-City / Multi-City training
	New York City	394,469	1,609,108	$5\leq L\leq 128$	Per-City / Multi-City training
	Pittsburgh	68,869	219,842	$5\leq L\leq 128$	Per-City / Multi-City training
	Dallas	180,268	939,831	$5\leq L\leq 128$	Only for Cross-City Transfer
Foursquare	Los Angeles	25,889	10,197	$5\leq L\leq 128$	Per-City / Multi-City training

Appendix

Table 3: Two multi-city human mobility datasets, i.e., Veraset and Foursquare check-in datasets.

SID Tokenizer		Prefix Planner		Latent-Flow Realizer
Parameter	Value	Parameter	Value	Parameter	Value
SID levels	6	Depth	6	Transformer width	256
Codebook size	32 (Boston / Pittsburgh)	Attention heads	8	Latent width	256
Codebook size	48 (Chicago / NYC)	Gap bins	16	Attention heads	8
Prefix levels	1–3	Time bins	42	Flow trunk	6 layers
Suffix levels	4–6	Max. occurrences	128	Max. occurrences	128

Appendix

Table 4: Architecture hyperparameters of SeMoFlow.

Module	Per-City Training		Joint Training
	Parameter	Value	Parameter	Value
Planner	Learning rate	$1.4\times 10^{-3}$ (Boston/NYC/Pittsburgh) $2\times 10^{-3}$ (Chicago)	Learning rate	$1.4\times 10^{-3}$
	Batch size	1024 (Boston); 4096 (Chicago) 2048 (NYC/Pittsburgh)	Batch size	2304
	Epochs	30 (Boston/Pittsburgh) 50 (Chicago); 60 (NYC)	Epochs	100
	Weight decay	$10^{-4}$	Weight decay	$10^{-4}$
	Routing temperature	1.7	Routing temperature	2.0

Appendix

Table 5: Training hyperparameters of SeMoFlow.

Figure 6: Mobility flow networks of real and SeMoFlow-generated trajectories across four cities.

Figure 7: Spatial distributions of real and SeMoFlow-generated POI visits across four cities. Visit frequencies are aggregated into hexagonal spatial bins and visualized using a shared logarithmic color scale.

Figure 8: POI category distributions of real and SeMoFlow-generated trajectories across four cities. Bars compare the share of visits to the twelve most frequently visited POI categories in each city.

City	Model	$JSD_{\mathrm{POI}}\downarrow$	$JSD_{\mathrm{geo}}\downarrow$	$JSD_{\mathrm{travel}}\downarrow$	$JSD_{\mathrm{gyr}}\downarrow$	$JSD_{\mathrm{motif}}\downarrow$	$JSD_{\mathrm{recur}}\downarrow$
Los Angeles	MoveSim	0.5282	0.0467	0.4850	0.4333	0.5963	0.4274
	TimeGeo	0.5558	0.0424	0.0597	0.0683	0.0311	0.0720
	MIRAGE	0.4729	0.0462	0.1367	0.1841	0.0020	0.0182
	MobiDiff	0.6144	0.3996	0.0945	0.1795	0.0782	0.2363
	Traveller	0.5596	0.0635	0.1496	0.1538	0.0037	0.0064
	SeMoFlow (Per-City)	0.4883	0.0564	0.3968	0.1474	0.0629	0.1072

Appendix

Table 6: Trajectory generation comparison across three Foursquare cities. Bold and underline indicate the best and second-best values within each city block, respectively.

Figure 9: Comparison of daily mobility motifs between real and SeMoFlow-generated trajectories across four cities. The top row illustrates representative motif structures, while the bottom panels compare their frequencies in real and generated trajectories.

Trajectory Generation

Trajectory generation creates plausible, goal-directed paths and motion sequences for people, vehicles, and robots.

See all 182 papers

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation
Jun 10, 2026Siyu Li, Toan Tran, Lingyi Zhao +2Trajectory Generation Human Behavior Simulation
Dept. of Computer Science, Emory University, Atlanta, GA, USA · Novateur Research Solutions, Ashburn, VA, USA
ActivityEditor: Learning to Synthesize Physically Valid Human Mobility
Apr 7, 2026Chenjie Yang, Yutian Jiang, Anqi Liang +3Trajectory Generation LLM Agent Trajectory Synthesis
Southwest Jiaotong University · HKUST (Guangzhou) · Shanghai Jiao Tong University +2
MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation
Oct 8, 2026Mengkun Gao, Zengqing Wu, Renhe Jiang +6Human Behavior Modeling Trajectory Generation
The University of Osaka · The University of Tokyo · The Hong Kong University of Science and Technology +1

Jun 10, 2026J/K move · Enter open · S save

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

Siyu Li, Toan Tran, Lingyi Zhao, Khurram Shafique+1

Dept. of Computer Science, Emory University, Atlanta, GA, USA · Novateur Research Solutions, Ashburn, VA, USA

Trajectory Generation Human Behavior Simulation Synthetic Data Generation LLM Agent Trajectory Synthesis LLM Agent Planning

Open paper AskPDF arXiv

Human mobility data is important for transportation, urban planning, and epidemic control, but large-scale trajectory collection is often costly and privacy-constrained, motivating realistic synthetic trajectory generation. Existing LLM-based generators typically rely on either prompt engineering, which preserves zero-shot reasoning but lacks fine-grained spatiotemporal grounding, or trajectory-level fine-tuning, which improves statistical precision but incurs substantial computational cost and may weaken general reasoning. We propose TrajGenAgent, a semantic-aware hierarchical LLM-agent framework for human mobility trajectory generation without model fine-tuning. TrajGenAgent uses a two-stage orchestrator-worker design: an LLM first synthesizes an individual- and weekday-conditioned activity chain from historical evidence via in-context learning, and a deterministic workflow then grounds each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. To evaluate realism beyond aggregate spatiotemporal statistics, we introduce an anomaly-detection-based evaluation framework using two complementary detectors to assess behavioral and semantic plausibility. Experiments on benchmark and large-scale simulation datasets show that TrajGenAgent improves spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over representative neural and LLM-based baselines, while avoiding parameter updates.

Abstract

Trajectory Generation

Explore similar work

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

ActivityEditor: Learning to Synthesize Physically Valid Human Mobility

MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation