Encounter-Driven City-Scale Human Mobility Generation with Large Language Models

Chao Li; Renhe Jiang; Chuan Xiao; Jiawei Wang

cs.AIOct 13, 2026

Encounter-Driven City-Scale Human Mobility Generation with Large Language Models

Authors: Chao Li, Renhe Jiang, Chuan Xiao, Jiawei Wang

Organizations: The Hong Kong University of Science and Technology · The University of Tokyo · The University of Osaka

Abstract

Large language models (LLMs) have shown strong potential for human mobility generation by modeling individual routines, activities, and mobility preferences. However, existing approaches remain largely resident-centric: generating realistic individual trajectories does not necessarily recover the encounters that emerge when multiple residents co-occur at the same place and time. To this end, we study encounter-aware human mobility generation, which jointly preserves individual mobility and population-level interaction structure. We propose EncounterMob, an encounter-driven LLM framework consisting of a coarse-to-fine Encounter Generator and an Iterative Coupling Framework. The Encounter Generator first captures city-level encounter rhythms and then instantiates participant-time-venue encounters, while the Iterative Coupling Framework coordinates these encounters with personal mobility behavior through successive refinement. We further introduce an encounter recovery protocol covering encounter marginal and joint distributions over residents, times, and venues. Experiments on the Japan and NYC datasets show that EncounterMob outperforms all external baselines on all seven encounter metrics and all six personal mobility metrics. Compared with ELLMob, it reduces the participant-time-venue joint error by 40.4%/41.7% and the mean personal mobility error by 58.8%/32.7% on Japan/NYC, respectively.

Figures & tables

Figure 1: Individual mobility generation versus encounter-driven mobility generation. Colored paths represent the trajectories of different residents, and highlighted locations indicate encounters across trajectories. (a) Individual mobility generation models each resident according to personal mobility patterns. (b) Encounter-driven mobility generation additionally coordinates encounter participants, times, and locations across the trajectories of multiple residents within the generated mobility data.

Dimension	Distribution	Characterization
Participant group	$G$ (GroupD)	Complete participant groups involved in encounters
Time	$T$ (TD)	Temporal distribution of encounters
Venue	$V$ (VD)	Venue distribution of encounters
Participant–Time	$G{\times}T$	Association between participant groups and times
Participant–Venue	$G{\times}V$	Association between participant groups and venues
Time–Venue	$T{\times}V$	Association between encounter times and venues

Table 1: Evaluation dimensions in the proposed encounter recovery protocol.

Figure 2: Overview of EncounterMob . (1) The coarse-to-fine Encounter Generator infers city-level encounter rhythms from historical mobility and resident profiles, then instantiates participant–time–venue encounter proposals. (2) The Iterative Coupling Framework compares these proposals with encounters realized by independently initialized personal trajectories. Guided by historical evidence, the joint critic provides feedback to revise personal trajectories and, when needed, city-level rhythms and encounter proposals. The personal trajectories form the final output. Colored dashed curves illustrate correspondences between proposed and realized encounters after coupling.

	Japan							NYC
Model	TD	VD	GroupD	$T{\times}V$	$G{\times}T$	$G{\times}V$	$G{\times}T{\times}V$	TD	VD	GroupD	$T{\times}V$	$G{\times}T$	$G{\times}V$	$G{\times}T{\times}V$
LSTM	0.09877	0.24476	0.68514	0.50963	0.69207	0.68664	0.69207	0.69315	0.66362	0.69315	0.69315	0.69315	0.69315	0.69315
GETNext	0.05961	0.23111	0.68071	0.49085	0.69315	0.68496	0.69315	0.65010	0.69315	0.69315	0.69315	0.69315	0.69315	0.69315
MHSA	0.06601	0.22904	0.68160	0.49340	0.69315	0.68468	0.69315	0.64424	0.66869	0.69315	0.69315	0.69315	0.69315	0.69315
TrajGAIL	0.08504	0.45497	0.68895	0.60298	0.69315	0.69167	0.69315	0.45976	0.66984	0.66637	0.69315	0.69315	0.69315	0.69315
DiffTraj	0.06280	0.22032	0.67645	0.46585	0.69286	0.68045	0.69286	0.66286	0.66728	0.69315	0.69315	0.69315	0.69315	0.69315

Table 2: Encounter fidelity on Japan and NYC. All metrics are JSDs computed with natural logarithms on normalized histograms; lower is better. Bold and underline indicate the best and second-best results among the compared methods; ablations are excluded from ranking.

	Japan						NYC
Model	SD	SI	DARD	STVD	CD	SGD	SD	SI	DARD	STVD	CD	SGD
LSTM	0.06371	0.10285	0.17937	0.58511	0.00493	0.11069	0.22653	0.11680	0.58435	0.68255	0.00870	0.20886
GETNext	0.05912	0.08853	0.15050	0.55073	0.01972	0.10460	0.25078	0.16924	0.57884	0.68183	0.00541	0.20827
MHSA	0.06143	0.05722	0.16247	0.55529	0.01929	0.11078	0.25595	0.15803	0.55372	0.67848	0.00734	0.19664
TrajGAIL	0.12680	0.08676	0.22410	0.60217	0.03489	0.22519	0.37290	0.20186	0.60749	0.68741	0.03541	0.35073
DiffTraj	0.08385	0.09709	0.14790	0.53861	0.01787	0.10721	0.23151	0.17015	0.55110	0.67917	0.01012	0.20594

Table 3: Personal mobility fidelity on Japan and NYC. All metrics are JSDs computed with natural logarithms; lower is better. Bold and underline indicate the best and second-best results among the compared methods; ablations are excluded from ranking.

Figure 3: City-level encounter patterns on Japan over 30 days. (a–e) Hourly encounter distributions across the ten most frequent ground-truth venue categories for ground truth, ELLMob, w/o Rhythm, w/o Instantiation, and final output. Each heatmap is normalized within these categories; colors are scaled by the maximum for each category across all five panels and hours. (f) Hourly encounter counts across all venue categories for the same five sources, pooled over the 30 days. Generated encounters are extracted from final trajectories.

Dataset	Method	Count	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$
Japan	ELLMob	1,627	13.83	15.53	14.63
	w/o Rhythm	940	44.55	38.75	41.45
	w/o Instantiation	1,104	27.86	28.73	28.29
	EncounterMob	1,111	49.66	49.86	49.76
NYC	ELLMob	26	50.00	30.77	38.10
	w/o Rhythm	123	69.23	69.23	69.23

Table 4: Participant-group recovery in Japan and NYC. Count is the number of generated encounter events, precision, recall, and F1 are percentages; higher is better.

Figure 4: Effect of Iterative Coupling on Japan over R0–R3. Encounter JSD decreases across rounds, while group recovery F1 of the generated trajectories increases and personal mean JSD remains relatively stable. JSD curves use the left axis; group recovery F1 uses the right axis. Lower JSD and higher F1 are better.

Figure 5: Case studies of Iterative Coupling: (a) revising personal activity to realize a supported encounter; (b) revising an encounter proposal to preserve supported personal activity. Center panels paraphrase Critic feedback and subsequent revisions. Japan shows a morning segment; NYC shows full daily stops. Lines connect stops, not road routes. The cases use different saved configurations, with ground truth reserved for evaluation.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	Japan	NYC
Historical period (days)	60	123
Test period (days)	30	61
Residents	1,049	226
Activity stops	43,147	3,432
Active resident-days	12,887	2,017
Encounters	921	77

Appendix

Table 5: Dataset statistics. Activity stops, active resident-days, and encounters refer to the test period.

Figure 6: Spatial distributions of actual encounters in Greater Tokyo over the same 30-day evaluation period. Rows represent venue categories, and columns compare ground truth with encounters extracted from the final trajectories generated by each method. Panels within each row share the ground-truth count denominator and color scale.

Method	Active RD	Total tokens (M)	Tokens/RD (k)
ELLMob	14,279	130.058	9.11
EncounterMob	12,887	587.586	45.60
ELLMob	2,017	14.645	7.26
EncounterMob	2,017	49.823	24.70

Appendix

Table 6: Inference token usage of the retained runs. M denotes million tokens; k denotes thousand tokens. Per-day usage is normalized by active resident-days (RD).

\dagger

marks a calibrated estimate; — indicates unavailable token accounting.

Human Behavior Simulation

Human behavior simulation models how individuals and groups decide, interact, move, and adapt in social, economic, educational, and physical settings.

See all 239 papers

ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework
Mar 9, 2026Yusong Wang, Chuang Yang, Jiawei Wang +5
Institute of Science Tokyo · The University of Tokyo · The University of Osaka
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
MobEvolve: An Agentic Self-Evolving Heuristic System for Interpretable Human Mobility Generation
Jun 1, 2026Junlin He, Yihong Tang, Tong Nie +6Human Behavior Simulation Self-Evolving Agents
1The Hong Kong Polytechnic University · 2McGill University · 8Mila – Quebec AI Institute +4

Mar 9, 2026J/K move · Enter open · S save

ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework

Yusong Wang, Chuang Yang, Jiawei Wang, Xiaohang Xu+4

Institute of Science Tokyo · The University of Tokyo · The University of Osaka

Open paper AskPDF arXiv

Human mobility generation aims to synthesize plausible trajectory data, which is widely used in urban system research. While Large Language Model-based methods excel at generating routine trajectories, they struggle to capture deviated mobility during large-scale societal events. This limitation stems from two critical gaps: (1) the absence of event-annotated mobility datasets for design and evaluation, and (2) the inability of current frameworks to reconcile competitions between users' habitual patterns and event-imposed constraints when making trajectory decisions. This work addresses these gaps with a twofold contribution. First, we construct the first event-annotated mobility dataset covering three major events: Typhoon Hagibis, COVID-19, and the Tokyo 2021 Olympics. Second, we propose ELLMob, a self-aligned LLM framework that first extracts competing rationales between habitual patterns and event constraints, based on Fuzzy-Trace Theory, and then iteratively aligns them to generate trajectories that are both habitually grounded and event-responsive. Extensive experiments show that ELLMob wins state-of-the-art baselines across all events, demonstrating its effectiveness. Our codes and datasets are available at https://github.com/deepkashiwa20/ELLMob.

Abstract

Human Behavior Simulation

Explore similar work

ELLMob: Event-Driven Human Mobility Generation with Self-Aligned LLM Framework

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

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