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

Yusong Wang; Chuang Yang; Jiawei Wang; Xiaohang Xu; Jiayi Xu; Dongyuan Li; Chuan Xiao; Renhe Jiang

cs.LGMar 9, 2026

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

Authors: Yusong Wang, Chuang Yang, Jiawei Wang, Xiaohang Xu, Jiayi Xu, Dongyuan Li, Chuan Xiao, Renhe Jiang

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

Abstract

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.

Figures & tables

Figure 1: Event-driven mobility generation by LLMs, which incorporates event context to capture real-world human mobility on three different events: (1) Typhoon: evacuation from seaside, (2) COVID-19 Pandemic: self-restraint, and (3) Olympics: restricted zones and traffic jam.

Event	Event Period	Pre-Event Period	Description
Typhoon Hagibis	2019-10-12 $\sim$ 10-13	2019-08-13 $\sim$ 10-11	Natural disaster.
COVID-19 Pandemic	2020-04-07 $\sim$ 04-13	2020-02-07 $\sim$ 04-06	Public health emergency.
Tokyo 2021 Olympics	2021-07-23 $\sim$ 07-29	2021-05-24 $\sim$ 07-22	Pandemic-era large event.
Normal Period	2019-09-01 $\sim$ 09-30	2019-07-03 $\sim$ 08-31	Regular urban mobility.

Table 1: Specifications for experimental evaluation scenarios.

Table 3

Figure 2: Normalized radar charts of four mobility metrics on pre-event and during-event patterns.

Figure 3: The ELLMob framework architecture comprising three interconnected modules: Trajectory Generation, Gist Extraction, and Reflection-based alignment.

Gist Type	Attribute	Description	Example
Pattern Gist	Core Behavior	The dominant pattern of action.	Daily commute to a office.
	Points of Inertia	Deeply embedded, non-negotiable components.	Returning home to a specific neighborhood at night.
	Points of Fracture	Critical dependencies and single points of failure.	Reliance on a single train line that might be suspended.
Event Gist	Primary Intent	Core implication of the event for mobility decisions.	High risk outdoors, strong incentive to stay home.
	Behavioral Implications	Survival, social dynamics, and compliance.	Evacuation from coastal areas, seeking indoor shelter.
	Risk-Reward Calculus	A cost-benefit analysis of the response to event risks.	Risk of injury outweighs reward of a non-essential outing.

Table 4: A set of defined core attributes that guide the gist extraction from source information.

Models	Typhoon Hagibis				COVID-19 Pandemic				Tokyo 2021 Olympics
Models	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓
LSTM	0.1336	0.1039	0.0555	0.1111	0.1928	0.1047	0.1300	0.2571	0.1147	0.0651	0.0598	0.0634
DeepMove	0.1697	0.0826	0.0266	0.0759	0.1838	0.0834	0.0423	0.1688	0.1667	0.0492	0.0587	0.0555
GETNext	0.3031	0.2007	0.0274	0.1037	0.2891	0.2241	0.0142	0.1354	0.2701	0.1473	0.0176	0.1204
MHSA	0.1430	0.1815	0.0118	0.0711	0.2180	0.3083	0.0254	0.0437	0.1815	0.2013	0.0120	0.0525
TrajGAIL	0.1034	0.3591	0.0155	0.0275	0.1600	0.3557	0.0195	0.0444	0.0863	0.2913	0.0121	0.0104

Table 5: Comparison of different methods under three events. Performance is evaluated by JSD across four dimensions with the best performance highlighted in bold .

Figure 4: (a) Comparison of generated activity proportions (relative to the total number of activities) during the COVID-19 Pandemic. Each chart contrasts the ground truth (G.T.), distribution with: Without internal alignment (w/o I.A.), without external alignment (w/o E.A.), and the complete model (Comp.). (b) Comparison of three key activity categories distributions generated by ELLMob with various LLM-based baselines. (c) Performance comparison on the active user prediction task.

Models	Typhoon Hagibis				COVID-19 Pandemic				Tokyo 2021 Olympics
Models	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓
Reflexion	0.1106	0.1282	0.0841	0.0855	0.1685	0.0588	0.0146	0.0269	0.1741	0.0308	0.0378	0.0704
SELF-REFINE	0.1979	0.0637	0.0135	0.0193	0.2122	0.1053	0.0344	0.0294	0.0826	0.0320	0.0073	0.0058
Air	0.0764	0.0710	0.0198	0.0204	0.1858	0.0454	0.0256	0.0291	0.0774	0.0441	0.0053	0.0035
ELLMob	0.0642	0.0200	0.0041	0.0173	0.1003	0.0444	0.0080	0.0268	0.0617	0.0061	0.0022	0.0035

Table 7: Comparison with iterative-reflection baselines with the best performance in bold .

Model	Token Count	Inference Time (s)	Cost (USD)
LLMOB	1,271	10.12	0.00030
LLM-MOB	3,954	3.72	0.00064
LLM-MOVE	4,954	4.05	0.00078
LLM-ZS	5,184	3.34	0.00080
Reflexion	26,057	27.12	0.00417
SELF-REFINE	15,382	21.16	0.00258

Table 8: Computational Efficiency Analysis. The reported total token count includes both input and output tokens, and the overall cost is computed by accounting for their respective pricing rates.

Figure 6: Spatial mobility patterns. Darker red in the heatmaps indicates higher visit frequency.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A2: Log-log plots of the Complementary Cumulative Distribution Function (CCDF) of check-in counts. Left part of the figure shows the distribution of check-ins per POI. Right part of the figure shows the distribution of check-ins per user. Both distributions exhibit a linear trend, characteristic of a power-law.

User	Lat.	Long.	POI ID	Subcategory	Category	Timestamp	Comments (Translated and rewritten)
0118	35.652	139.543	1003	Home Appliance Store	Retail	2020-04-07 18:33	Chofu-chan: Oh no, an emergency declaration has been announced!!
0118	35.633	139.577	14932	Clothing Store	Retail	2020-04-10 19:16	In response to coronavirus precautions, credit card transactions have transitioned to self-scanning! (This is what we’ve been aiming for.).
0118	35.632	139.577	4859	Rail Station	Travel & Transport	2020-04-10 19:21	Wait a minute? Isn’t this considered close contact?

Appendix

Table A2: A user’s mobility sample, showcasing activity trajectories with spatio-temporal features and comments.

Figure A3: A visualization of category distributions across four event scenarios. Each data point represents the percentage share of the category out of the total activities in that scenario.

Figure A4: This figure shows three event schema instances (Typhoon Hagibis, COVID-19 pandemic, and Tokyo 2021 Olympics), serving as structured contexts for event-driven mobility generation.

Models	Typhoon Hagibis				COVID-19 Pandemic				Tokyo 2021 Olympics
Models	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓
w/ verbatim	0.1511	0.0724	0.0230	0.0595	0.2071	0.0713	0.0324	0.0422	0.1030	0.0452	0.0054	0.0189
w/o bottom	0.0978	0.0443	0.0101	0.0327	0.1687	0.0690	0.0305	0.0333	0.0892	0.0364	0.0036	0.0082
ELLMob	0.0642	0.0200	0.0041	0.0173	0.1003	0.0444	0.0080	0.0268	0.0617	0.0061	0.0022	0.0035

Appendix

Table A3: Ablation study on the effectiveness of different FTT components with the best performance highlighted in bold .

Models	Typhoon Hagibis				COVID-19 Pandemic				Tokyo 2021 Olympics
Models	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓
LLMOB	0.1069	0.0743	0.0126	0.0196	0.0846	0.0693	0.0105	0.0292	0.0572	0.0281	0.0031	0.0087
LLM-MOB	0.3180	0.1726	0.2004	0.0968	0.1428	0.0660	0.0280	0.0459	0.0379	0.0154	0.0053	0.0065
LLM-Move	0.2383	0.0721	0.0887	0.0465	0.3683	0.0644	0.0378	0.0353	0.1979	0.0287	0.0097	0.0063
LLM-ZS	0.3466	0.1537	0.2556	0.1084	0.2788	0.1344	0.2489	0.1479	0.0967	0.0313	0.0137	0.0057
ELLMob	0.0850	0.0267	0.0087	0.0160	0.0546	0.0586	0.0069	0.0210	0.0113	0.0074	0.0016	0.0048

Appendix

Table A4: Performance comparison of different methods using Gemini Flash 2.0. Performance is evaluated by JSD across four dimensions with the best performance highlighted in bold .

Models	Typhoon Hagibis				COVID-19 Pandemic				Tokyo 2021 Olympics
Models	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓	SI↓	SD↓	CD↓	SGD↓
LLaMA3-8B	0.0663	0.0512	0.0004	0.0132	0.0669	0.0624	0.0087	0.0263	0.0407	0.0322	0.0023	0.0030
Qwen-2.5-14B	0.1594	0.0607	0.0091	0.0115	0.0836	0.0646	0.0016	0.0225	0.0530	0.0238	0.0035	0.0021
R1-Q7B	0.0881	0.0516	0.0033	0.0203	0.1104	0.0993	0.0163	0.0392	0.0570	0.0251	0.0015	0.0071

Appendix

Table A5: Performance comparison of ELLMob across different open-source LLM backbones. Performance is evaluated by JSD across four dimensions

Model	SI	SD	CD	SGD
LLMOB	0.1934	0.1589	0.0344	0.1201
LLM-MOB	0.1732	0.1617	0.0304	0.0988
LLM-MOVE	0.2134	0.1555	0.0698	0.1227
LLM-ZS	0.1531	0.1788	0.0299	0.1161
ELLMob	0.1131	0.1001	0.0120	0.0556

Appendix

Table A6: Regional generalizability analysis on the Osaka data during the COVID-19 Pandemic. Performance is evaluated by JSD across four dimensions with the best performance in bold .

Model	SI	SD	CD	SGD
LLMOB	0.0776	0.0391	0.0230	0.0422
LLM-MOB	0.1061	0.0413	0.0318	0.0379
LLM-MOVE	0.2248	0.0493	0.0272	0.0550
LLM-ZS	0.0996	0.0497	0.0230	0.0481
ELLMob	0.0598	0.0250	0.0200	0.0317

Appendix

Table A7: Event generalizability analysis on the New Year scenario in Tokyo. Performance is evaluated by JSD across four dimensions with the best performance in bold .

Figure A5: Performance comparison across iteration numbers $K$ = 1 to 5 for three major events: Typhoon Hagibis, COVID-19 Pandemic, and Tokyo 2021 Olympics. Performance is measured using Jensen-Shannon Divergence (JSD), where lower values indicate better performance.

Model	SI	SD	CD	SGD
LLMOB	0.1057 $\pm$ 0.0277	0.1509 $\pm$ 0.0386	0.0497 $\pm$ 0.0243	0.0319 $\pm$ 0.0170
LLM-MOB	0.1501 $\pm$ 0.0170	0.0558 $\pm$ 0.0208	0.0352 $\pm$ 0.0096	0.0532 $\pm$ 0.0430
LLM-MOVE	0.1242 $\pm$ 0.0035	0.0407 $\pm$ 0.0021	0.0131 $\pm$ 0.0007	0.0290 $\pm$ 0.0014
LLM-ZS	0.1601 $\pm$ 0.0071	0.1335 $\pm$ 0.0055	0.0155 $\pm$ 0.0008	0.0750 $\pm$ 0.0025
ELLMob	0.0649 $\pm$ 0.0030	0.0220 $\pm$ 0.0044	0.0046 $\pm$ 0.0012	0.0162 $\pm$ 0.0018

Appendix

Table A8: Stability analysis on the Typhoon Hagibis dataset. Results are reported as Mean

\pm

Standard Deviation across 5 independent runs.

Model	SI	SD	CD	SGD
LLMOB	0.1128 $\pm$ 0.0173	0.1221 $\pm$ 0.0298	0.0158 $\pm$ 0.0038	0.0295 $\pm$ 0.0047
LLM-MOB	0.1099 $\pm$ 0.0107	0.0613 $\pm$ 0.0263	0.0191 $\pm$ 0.0072	0.0359 $\pm$ 0.0066
LLM-MOVE	0.1789 $\pm$ 0.0241	0.0511 $\pm$ 0.0044	0.0457 $\pm$ 0.0185	0.0544 $\pm$ 0.0067
LLM-ZS	0.1108 $\pm$ 0.0103	0.0551 $\pm$ 0.0025	0.0442 $\pm$ 0.0074	0.0605 $\pm$ 0.0056
ELLMob	0.1002 $\pm$ 0.0094	0.0443 $\pm$ 0.0037	0.0079 $\pm$ 0.0015	0.0279 $\pm$ 0.0027

Appendix

Table A9: Stability analysis on the COVID-19 Pandemic dataset. Results are reported as Mean

\pm

Standard Deviation across 5 independent runs.

Model	SI	SD	CD	SGD
LLMOB	0.1042 $\pm$ 0.0131	0.0291 $\pm$ 0.0036	0.0115 $\pm$ 0.0017	0.0054 $\pm$ 0.0006
LLM-MOB	0.1123 $\pm$ 0.0067	0.0338 $\pm$ 0.0083	0.0157 $\pm$ 0.0084	0.0172 $\pm$ 0.0127
LLM-MOVE	0.2027 $\pm$ 0.0050	0.0304 $\pm$ 0.0014	0.0111 $\pm$ 0.0009	0.0053 $\pm$ 0.0004
LLM-ZS	0.0953 $\pm$ 0.0021	0.0320 $\pm$ 0.0014	0.0147 $\pm$ 0.0021	0.0060 $\pm$ 0.0011
ELLMob	0.0606 $\pm$ 0.0016	0.0063 $\pm$ 0.0002	0.0021 $\pm$ 0.0002	0.0037 $\pm$ 0.0002

Appendix

Table A10: Stability analysis on the Tokyo 2021 Olympics dataset. Results are reported as Mean

\pm

Standard Deviation across 5 independent runs.

Figure A6: The details of event schema under three different events.

Figure A7: The prompt of initial trajectory generation.

Figure A8: The prompt of event gist generation.

Figure A9: The prompt of action gist generation.

Figure A10: The prompt of pattern gist generation.

Figure A11: The prompt of conflict judgment.

Figure A12: The prompt of trajectory regeneration.

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
Encounter-Driven City-Scale Human Mobility Generation with Large Language Models
Oct 13, 2026Chao Li, Renhe Jiang, Chuan Xiao +1Human Behavior Simulation LLM Agent Trajectory Synthesis
The Hong Kong University of Science and Technology · The University of Tokyo · The University of Osaka
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

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

Explore similar work

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

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

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