Relative Transitions, Not Absolute Destinations: A Transfer-and-Ground Framework for Target-Trajectory-Free Human Mobility Generation

Yidi Wang; Yunhe Zhang; Bangchao Deng; Dingqi Yang; Pengyang Wang

cs.LGOct 1, 2026

Relative Transitions, Not Absolute Destinations: A Transfer-and-Ground Framework for Target-Trajectory-Free Human Mobility Generation

Authors: Yidi Wang, Yunhe Zhang, Bangchao Deng, Dingqi Yang, Pengyang Wang

Organizations: SKL-IOTSC, Department of CIS, University of Macau Macau SAR, China

Abstract

Individual mobility trajectories support urban analysis and location-based services, yet most trajectory generators require observations from their deployment city. This assumption excludes precisely the cities where trajectories are unavailable even though points of interest (POIs) and their attributes can be obtained from public maps. We study target-trajectory-free generation: learning from POIs and trajectories in source cities while utilizing only POI coordinates and categories in a target city, with no target trajectory or trajectory-derived statistic available for training, model selection, or generation. Existing trajectory generators typically predict absolute destinations, entangling reusable movement behavior with city-specific POI identities and spatial layouts. Our core insight is to replace this city-bound output with context-conditioned relative transitions. We propose Nomad, a transfer-and-ground framework that separates learning how people move from determining where those movements are realized. Specifically, a history-conditioned flow-matching model learns from source trajectories a transition prior over semantic displacement between POI contexts, geographic displacement, and elapsed time; at inference, a behavior graph and an exploration--return walk ground sampled transitions onto the target POI map. This factorization enables a direct test of representation level transferability without assuming invariance of the full mobility distribution. Extensive experiments across ten cities and 14 transfers show that Nomad outperforms adaptation baselines in trajectory fidelity and downstream utility, lowering the average error over the best baseline of each metric by about 15% in distributional fidelity and about 3% in downstream utility.

Figures & tables

Figure 1 . NYC vs. TKY: same transitions, different POIs.

Figure 2 . Architecture overview of Nomad.

	Methods	Downstream Utility				Distributional Fidelity
	Methods	Pred-MAPE $\downarrow$	Pred-MSPE $\downarrow$	Rec-MAPE $\downarrow$	Rec-MSPE $\downarrow$	OD-Sim $\downarrow$	Distance $\downarrow$	Speed $\downarrow$	Time-Cat $\downarrow$
NYC $\to$ TKY	Markov ( Gambs et al., 2012 )	0.390 $\pm$ 0.040	0.165 $\pm$ 0.036	0.180 $\pm$ 0.049	0.054 $\pm$ 0.037	0.616 $\pm$ 0.005	0.227 $\pm$ 0.001	–	–	+24.76%
	DITRAS ( Pappalardo and Simini, 2018 )	0.396 $\pm$ 0.030	0.181 $\pm$ 0.030	0.276 $\pm$ 0.068	0.111 $\pm$ 0.043	0.516 $\pm$ 0.007	0.181 $\pm$ 0.001	0.077 $\pm$ 0.006	0.295 $\pm$ 0.002
	Gravity ( Barbosa et al., 2018 )	0.419 $\pm$ 0.049	0.193 $\pm$ 0.045	0.251 $\pm$ 0.132	0.119 $\pm$ 0.112	0.460 $\pm$ 0.003	0.126 $\pm$ 0.001	0.117 $\pm$ 0.004	0.265 $\pm$ 0.001
	MIRAGE ( Deng et al., 2025b )	0.390 $\pm$ 0.015	0.188 $\pm$ 0.010	0.197 $\pm$ 0.004	0.110 $\pm$ 0.001	0.925 $\pm$ 0.068	0.137 $\pm$ 0.001	0.055 $\pm$ 0.001	0.273 $\pm$ 0.002
	GTG ( Wang et al., 2025 )	0.392 $\pm$ 0.015	0.191 $\pm$ 0.010	0.168 $\pm$ 0.025	0.020 $\pm$ 0.010	0.593 $\pm$ 0.018	0.236 $\pm$ 0.001	0.186 $\pm$ 0.001	0.341 $\pm$ 0.001

Table 1. Main results on four zero-shot transfer settings. Each value is the mean

\pm

std over five generation seeds. Bold and underline mean the best and the second-best results. Markov ’s outputs do not include timestamps, so ‘–’ is used as a placeholder. The last column reports Nomad’s average relative gain over the second-best method across the eight metrics.

	Variants	Downstream Utility				Distributional Fidelity
		Pred-MAPE $\downarrow$	Pred-MSPE $\downarrow$	Rec-MAPE $\downarrow$	Rec-MSPE $\downarrow$	OD-Sim $\downarrow$	Distance $\downarrow$	Speed $\downarrow$	Time-Cat $\downarrow$
NYC $\to$ TKY	w/o graph encoder	0.419 $\pm$ 0.004	0.247 $\pm$ 0.003	0.224 $\pm$ 0.010	0.117 $\pm$ 0.001	0.501 $\pm$ 0.002	0.060 $\pm$ 0.001	0.025 $\pm$ 0.000	0.254 $\pm$ 0.001
	w/o flow-matching	0.484 $\pm$ 0.016	0.224 $\pm$ 0.014	0.141 $\pm$ 0.015	0.053 $\pm$ 0.004	0.513 $\pm$ 0.004	0.148 $\pm$ 0.000	0.037 $\pm$ 0.000	0.249 $\pm$ 0.000
	w/o copy-gate	0.431 $\pm$ 0.005	0.178 $\pm$ 0.001	0.160 $\pm$ 0.000	0.042 $\pm$ 0.003	0.400 $\pm$ 0.002	0.049 $\pm$ 0.000	0.025 $\pm$ 0.000	0.235 $\pm$ 0.000
	w/o AR-mask	0.427 $\pm$ 0.016	0.148 $\pm$ 0.015	0.156 $\pm$ 0.011	0.076 $\pm$ 0.002	0.437 $\pm$ 0.002	0.070 $\pm$ 0.001	0.026 $\pm$ 0.001	0.265 $\pm$ 0.001
	w/o condition	0.409 $\pm$ 0.007	0.161 $\pm$ 0.005	0.150 $\pm$ 0.002	0.084 $\pm$ 0.001	0.465 $\pm$ 0.005	0.133 $\pm$ 0.000	0.026 $\pm$ 0.000	0.255 $\pm$ 0.001

Table 2. Ablation study on the four transfers. Each value is the mean

\pm

std over five generation seeds. Bold mean the best results.

Panel of Figure 3 — Figure 3 . Impact of the target trajectory ratio for few-shot adapting on four transfers.

Panel of Figure 4 — Figure 4 . The embedding visualization of destinations and transitions on NYC and TKY.

Figure 5 . Impact of

K

and

N

for grounding on NYC

\to

TKY.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Region	City	#POI	#Cat.	#Traj.	#Check-in	Avg. len.
Metropolis	NYC	38,333	250	27,662	135,717	4.9
Metropolis	TKY	61,858	247	73,999	397,416	5.4
Japan	FUK	3,673	256	2,577	22,294	8.7
	KYO	5,124	265	3,546	30,967	8.7
	NGO	5,796	286	4,321	36,147	8.4
	OSA	11,974	344	10,038	87,257	8.7

Appendix

Table 3. Per-city statistics of the ten Foursquare cities.

Transfer	MIRAGE (s)	GTG (s)	COLA (s)	Nomad (s)
TKY $\to$ NYC	27.3	446.0	68.0	9.6
NYC $\to$ TKY	8.9	277.0	25.8	5.4
OSA $\to$ NGO	3.2	86.0	9.1	3.9
OSA $\to$ KYO	3.2	86.0	9.1	3.9
OSA $\to$ FUK	3.2	86.0	9.1	3.9
NGO $\to$ OSA	1.6	41.8	4.5	3.4

Appendix

Table 4 . Training time per epoch (summed over all training stages) and generation time for 10k trajectories.

	Method	Downstream Utility				Distributional Fidelity
	Method	Pred-MAPE $\downarrow$	Pred-MSPE $\downarrow$	Rec-MAPE $\downarrow$	Rec-MSPE $\downarrow$	OD-Sim $\downarrow$	Distance $\downarrow$	Speed $\downarrow$	Time-Cat $\downarrow$
TKY $\to$ NYC	Markov ( Gambs et al., 2012 )	0.619 $\pm$ 0.071	0.400 $\pm$ 0.094	0.449 $\pm$ 0.076	0.229 $\pm$ 0.071	0.585 $\pm$ 0.005	0.181 $\pm$ 0.001	–	–	+24.43%
	DITRAS ( Pappalardo and Simini, 2018 )	0.776 $\pm$ 0.083	0.641 $\pm$ 0.122	0.696 $\pm$ 0.092	0.547 $\pm$ 0.121	0.473 $\pm$ 0.008	0.143 $\pm$ 0.000	0.040 $\pm$ 0.000	0.284 $\pm$ 0.002
	Gravity ( Barbosa et al., 2018 )	0.647 $\pm$ 0.068	0.462 $\pm$ 0.103	0.391 $\pm$ 0.062	0.183 $\pm$ 0.067	0.312 $\pm$ 0.002	0.128 $\pm$ 0.002	0.029 $\pm$ 0.000	0.236 $\pm$ 0.001
	MIRAGE ( Deng et al., 2025b )	0.608 $\pm$ 0.033	0.418 $\pm$ 0.042	0.445 $\pm$ 0.054	0.290 $\pm$ 0.058	0.946 $\pm$ 0.038	0.083 $\pm$ 0.001	0.063 $\pm$ 0.001	0.270 $\pm$ 0.002
	GTG ( Wang et al., 2025 )	0.575 $\pm$ 0.004	0.328 $\pm$ 0.004	0.285 $\pm$ 0.037	0.088 $\pm$ 0.024	0.710 $\pm$ 0.025	0.235 $\pm$ 0.002	0.121 $\pm$ 0.001	0.223 $\pm$ 0.002

Appendix

Table 5. Full results on Metropolis transfer pair. Conventions as in Table 1 .

	Method	Downstream Utility				Distributional Fidelity
	Method	Pred-MAPE $\downarrow$	Pred-MSPE $\downarrow$	Rec-MAPE $\downarrow$	Rec-MSPE $\downarrow$	OD-Sim $\downarrow$	Distance $\downarrow$	Speed $\downarrow$	Time-Cat $\downarrow$
OSA $\to$ NGO	Markov ( Gambs et al., 2012 )	0.365 $\pm$ 0.028	0.172 $\pm$ 0.020	0.233 $\pm$ 0.091	0.103 $\pm$ 0.069	0.399 $\pm$ 0.024	0.199 $\pm$ 0.003	–	–	+6.24%
	DITRAS ( Pappalardo and Simini, 2018 )	0.382 $\pm$ 0.039	0.180 $\pm$ 0.032	0.287 $\pm$ 0.108	0.155 $\pm$ 0.114	0.312 $\pm$ 0.013	0.159 $\pm$ 0.003	0.058 $\pm$ 0.003	0.344 $\pm$ 0.002
	Gravity ( Barbosa et al., 2018 )	0.463 $\pm$ 0.113	0.253 $\pm$ 0.113	0.449 $\pm$ 0.135	0.398 $\pm$ 0.160	0.350 $\pm$ 0.013	0.182 $\pm$ 0.002	0.024 $\pm$ 0.001	0.367 $\pm$ 0.001
	MIRAGE ( Deng et al., 2025b )	0.455 $\pm$ 0.030	0.233 $\pm$ 0.021	0.335 $\pm$ 0.094	0.224 $\pm$ 0.153	0.837 $\pm$ 0.145	0.056 $\pm$ 0.003	0.034 $\pm$ 0.003	0.405 $\pm$ 0.008
	GTG ( Wang et al., 2025 )	0.398 $\pm$ 0.039	0.171 $\pm$ 0.029	0.365 $\pm$ 0.020	0.395 $\pm$ 0.010	0.736 $\pm$ 0.030	0.153 $\pm$ 0.003	0.165 $\pm$ 0.002	0.317 $\pm$ 0.007

Appendix

Table 6. Full results on Japanese-cluster transfer pairs. Conventions as in Table 1 .

	Method	Downstream Utility				Distributional Fidelity
	Method	Pred-MAPE $\downarrow$	Pred-MSPE $\downarrow$	Rec-MAPE $\downarrow$	Rec-MSPE $\downarrow$	OD-Sim $\downarrow$	Distance $\downarrow$	Speed $\downarrow$	Time-Cat $\downarrow$
IST $\to$ IZM	Markov ( Gambs et al., 2012 )	0.242 $\pm$ 0.055	0.081 $\pm$ 0.041	0.316 $\pm$ 0.177	0.216 $\pm$ 0.147	0.502 $\pm$ 0.010	0.162 $\pm$ 0.002	–	–	-10.84%
	DITRAS ( Pappalardo and Simini, 2018 )	0.319 $\pm$ 0.066	0.138 $\pm$ 0.032	0.375 $\pm$ 0.046	0.238 $\pm$ 0.043	0.408 $\pm$ 0.012	0.127 $\pm$ 0.002	0.062 $\pm$ 0.003	0.257 $\pm$ 0.002
	Gravity ( Barbosa et al., 2018 )	0.291 $\pm$ 0.088	0.123 $\pm$ 0.067	0.242 $\pm$ 0.166	0.123 $\pm$ 0.129	0.334 $\pm$ 0.003	0.114 $\pm$ 0.003	0.004 $\pm$ 0.000	0.274 $\pm$ 0.002
	MIRAGE ( Deng et al., 2025b )	0.434 $\pm$ 0.029	0.229 $\pm$ 0.028	0.322 $\pm$ 0.133	0.224 $\pm$ 0.122	0.976 $\pm$ 0.033	0.146 $\pm$ 0.003	0.063 $\pm$ 0.003	0.380 $\pm$ 0.009
	GTG ( Wang et al., 2025 )	0.329 $\pm$ 0.030	0.151 $\pm$ 0.033	0.178 $\pm$ 0.030	0.112 $\pm$ 0.009	0.633 $\pm$ 0.031	0.311 $\pm$ 0.003	0.149 $\pm$ 0.003	0.282 $\pm$ 0.003

Appendix

Table 7. Full results on Turkish-cluster transfer pairs. Conventions as in Table 1 .

Panel of Figure 6 — Figure 6 . Hyper-parameter sensitivity on {ANK+BUR+IZM $\to$ IST}.

Panel of Figure 7 — Figure 7 . Hyper-parameter sensitivity on {KYO+NGO $\to$ OSA}.

Explore similar work

Jun 10, 2026cs.AI

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

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.

Siyu Li, Toan Tran, Lingyi Zhao +2

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

Mobility Traccia Human Motion Generation Large Language Model Agents

Ask PDF

May 11, 2026cs.AI

TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding

Urban mobility is naturally expressed both as trajectories in space and as natural-language descriptions of travel intent, constraints, and preferences. However, prior work rarely evaluates these two modalities together on the same real-world trajectories: trajectory modeling often stays geometry-centric, while language-centric mobility benchmarks frequently target route planning and tool use rather than fine-grained, verifiable alignment between text and the underlying route. We introduce TrajPrism, a multi-task benchmark for language-trajectory alignment that unifies (i) instruction-conditioned trajectory generation, (ii) language-driven semantic trajectory retrieval, and (iii) trajectory captioning, together with an evaluation protocol that measures trajectory fidelity, retrieval quality, and language groundedness. We construct TrajPrism by pairing real urban trajectories with judge-filtered language annotations generated under a four-dimensional travel-intent taxonomy. The benchmark contains 300K selected trajectories across Porto, San Francisco, and Beijing, yielding 2.1M task instances from three instruction variants, three retrieval queries, and one caption per trajectory. We further develop proof-of-concept models for each task: TrajAnchor for instruction-conditioned trajectory generation, TrajFuse for semantic trajectory retrieval, and TrajRap for trajectory captioning. These models instantiate the proposed tasks and show that geometry-only trajectory baselines leave a large gap on our protocol, especially where language is part of the input-output interface. We release TrajPrism with code and a reproducible annotation pipeline that is designed to be portable across cities, given compatible trajectory inputs and map resources.

Lihuan Li, Wilson Wongso, Baiyu Chen +6

UNSW Sydney · HKUST (GZ)

Urban Mobility Journey Prism

Ask PDF

Jun 1, 2026cs.AI

MobEvolve: An Agentic Self-Evolving Heuristic System for Interpretable Human Mobility Generation

Human mobility generation aims to synthesize realistic trip chains for target populations based on individual features. Existing paradigms, including deep generative models, LLM-based methods, and traditional heuristics, struggle to satisfy the complex demands of this task while simultaneously maintaining interpretability, behavioral plausibility, population-level distributional alignment, and inference efficiency. To bridge this gap, we introduce MobEvolve, the first agentic self-evolving heuristic framework for human mobility generation. MobEvolve initializes a behavior-inspired heuristic system and employs an LLM agent to iteratively evolve its internal logic. By diagnosing empirical misalignments and failure cases on a validation set, the agent proposes targeted updates and accumulates evolution memory for cumulative self-improvement. Extensive evaluations on the Singapore and Montreal benchmarks demonstrate that MobEvolve significantly outperforms state-of-the-art deep generative and LLM-based methods in individual trajectory fidelity, population-level distribution alignment, and behavioral plausibility, while preserving interpretability and high inference efficiency.

Junlin He, Yihong Tang, Tong Nie +6

1The Hong Kong Polytechnic University · 2McGill University · 8Mila – Quebec AI Institute +4

Mobility Human Motion Generation Student-Generated Trajectories Self-Evolving Agents

Ask PDF