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

Mengkun Gao; Zengqing Wu; Renhe Jiang; Jiawei Wang; Yusong Wang; Chuang Yang; Shuyuan Zheng; Makoto Onizuka

cs.LGOct 8, 2026

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

Authors: Mengkun Gao, Zengqing Wu, Renhe Jiang, Jiawei Wang, Yusong Wang, Chuang Yang, Shuyuan Zheng, Makoto Onizuka, +1 more

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

Abstract

Human mobility generation, an important task in urban research, synthesizes trajectory data for urban planning and transportation management. Human mobility can be characterized as a "why-where-when" decision process: people form an intention to move and then determine where and when the corresponding activity will take place. Trajectory generation under user-level and temporal distribution shifts may benefit from explicitly modeling this decision structure. However, many existing human mobility generation methods either represent behavioral intent at a coarse granularity, such as a daily plan or a trajectory-level description, or directly predict future locations without explicitly reasoning about a possible motivation for each movement step. We introduce MotiveMob, a motivation-driven autoregressive framework for human mobility generation that first forms a hypothesis about why the next movement may occur and then jointly generates where and when it may occur. At each step, a motivation predictor conditions on the current mobility state, a long-term behavioral report, and the mobility history to infer a plausible motivation or determine whether the trajectory should terminate. Given the hypothesized motivation, a state predictor grounds it in a candidate next location and arrival time. The candidate then undergoes speed-feasibility and repetition checks before being fed back for the next decision. We evaluate MotiveMob under distribution shifts involving unseen users and unseen temporal periods, including seasonal changes and the substantial behavioral disruption caused by the COVID-19 pandemic. Experiments show that MotiveMob consistently achieves better distributional fidelity than competitive pretraining-based and prompting-based methods under user-level and temporal distribution shifts, demonstrating robust generalization to out-of-distribution mobility patterns.

Figures & tables

Figure 1: Overview of the MotiveMob framework. Given the user context, MotiveMob generates a daily trajectory step by step, first predicting why (

P_{M}

) and then where and when (

P_{S}

). The implicit world model

P_{S}

is trained using observed and alternative transitions. The marker denotes rollout validation , which performs speed-feasibility checking and repetition-based termination, as described in Section 4.3 .

Table 1: Evaluation results for seen and unseen users across the three temporal settings. Panels (a) and (b) report results for seen and unseen users, respectively. In each panel, the upper and lower baseline sections correspond to pretraining-based and prompting-based methods, respectively. All values are JSD; lower is better. The best result in each column of each panel is shown in bold, and the runner-up is underlined. Speed and Rep. denote speed validation and repetition-aware termination, respectively. Best results are in bold, and runner-up results are underlined.

Seen users
State predictor $P_{S}$	Seen period				Regular unseen period				COVID-19 unseen period
State predictor $P_{S}$	SD	SI	DARD	STVD	SD	SI	DARD	STVD	SD	SI	DARD	STVD
MotiveMob (IWM $P_{S}$ )	0.0241	0.0159	0.0857	0.3841	0.0307	0.0212	0.1054	0.4558	0.0880	0.0309	0.1632	0.5745
w/ positive-only $P_{S}$	0.0328	0.0115	0.0707	0.3870	0.0376	0.0185	0.0943	0.4469	0.0914	0.0390	0.1586	0.5602
w/ rule-based $P_{S}$	0.0374	0.0101	0.0992	0.4286	0.0405	0.0165	0.1111	0.4932	0.0687	0.0198	0.1576	0.6230
Unseen users

Table 2: State predictor comparison without rollout validation for seen and unseen users across the three temporal settings. All values are JSD; lower is better. The best result in each column of each panel is shown in bold, and the runner-up is underlined.

Figure 2: Ground-truth and predicted spatial mobility distributions for unseen users across two unseen periods in Tokyo. Darker red indicates higher visit density.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Motivation	Behavioral interpretation	Destination venue categories
Return	Returning to a residence or temporary accommodation	Residential venues, hotels, hostels, inns, resorts, and other lodging venues
Personal Service	Obtaining a personal, household, financial, repair, or travel-related service	Salons, banks, repair shops, rental services, post offices, travel agencies, and related service venues
Entertainment	Participating in leisure, cultural, or recreational activities	Arts and entertainment, landmarks and outdoors, sports and recreation, cruises, auditoriums, and related venues
Dining	Eating or drinking	Dining and drinking venues, food courts, and cafeterias
Transit	Traveling or transferring between places	Airports, stations, public transport facilities, roads, and other transportation venues
Religious	Participating in religious or spiritual activities	Shrines, temples, churches, mosques, monasteries, and other religious venues

Appendix

Table 3: Mobility motivation taxonomy and its mapping from destination venue categories. The listed venues summarize the category groups associated with each motivation.

Figure 3: Distributional diagnostics for the log-transformed transition speeds used to calibrate $v_{\max}$ . (a) Empirical distribution and a Gaussian distribution with the same mean and standard deviation. The red line marks the two-standard-deviation upper bound. (b) Normal Q–Q plot showing the remaining deviations from Gaussianity, particularly in the tails.

Parameter	Setting
Construction data	25 seen users from January–October 2019, containing 3,559 daily trajectories and 31,063 observed transition instances
Alternative motivations	Three distinct motivations per observed transition, sampled without replacement while excluding the observed motivation
Prepared alternative cases	93,189 cases before candidate generation and judging; 31,063 judged alternative transitions are used for balanced state-predictor training
Candidate-generator architecture	Three-layer pre-layer-normalized Transformer encoder with hidden dimension 128, 4 attention heads, feed-forward dimension 256, and dropout 0.1
Prediction heads	Separate classification heads for location subcategory, spatial grid, and arrival-time bin
Candidate-generator objective	Equal-weighted cross-entropy losses for the subcategory, spatial-grid, and time-bin prediction heads

Appendix

Table 4: Configuration for alternative-transition construction.

Parameter	State predictor $P_{S}$	Motivation predictor $P_{M}$
Initialization	Llama-3.1-8B-Instruct	Independently initialized from Llama-3.1-8B-Instruct
Training data	31,063 observed transitions and 31,063 judged alternative transitions	34,622 observed instances, comprising 31,063 step-level motivations and 3,559 END labels
Observed-to-alternative ratio	$1{:}1$	Not applicable
Training objective	Assistant-token-only causal language-modeling loss	Assistant-token-only causal language-modeling loss
Quantization	4-bit NF4 with double quantization and bfloat16 computation	4-bit NF4 with double quantization and bfloat16 computation
QLoRA configuration	Rank $r=16$ , scaling factor $\alpha=16$ , dropout 0, and no bias	Rank $r=16$ , scaling factor $\alpha=16$ , dropout 0, and no bias

Appendix

Table 5: Training configurations for the state predictor

P_{S}

and motivation predictor

P_{M}

Parameter	Setting
Initial state	A START state; the first motivation and check-in are generated without applying a travel-speed constraint because no previous spatial state is available
Mobility-history context	Up to 30 check-ins from the three active days preceding the target day
Motivation decoding	Greedy decoding with at most 8 newly generated tokens
State decoding	Sampling with temperature 0.2, top- $p$ 0.9, and at most 32 newly generated tokens
State-generation attempts	At most 8 proposals per step, consisting of one initial proposal and up to seven retries
Retry decoding	Temperature increases from 0.3 to 0.9 in increments of 0.1; top- $p$ remains 0.9 and the maximum output length remains 32 tokens

Appendix

Table 6: Configuration for autoregressive trajectory inference.

Protocol component	Data usage and restriction
Predictor training	Trajectories from 25 seen users between January and October 2019. For the seen-user/seen-period evaluation, 20% of user–date trajectories are held out and excluded from predictor training.
User-level shift	The 50 unseen users are disjoint from the 25 users used to train $P_{M}$ and $P_{S}$ .
Temporal shift	The regular unseen period (November 2019–February 2020) and the COVID-19 period (April–October 2020) do not overlap with the predictor-training period.
Recent history $\mathcal{H}_{u,d}$	Contains at most 30 check-ins from the three active days strictly preceding target date $d$ . Target-day and future observations are never included.
Behavioral report $R_{u}$	In the temporal-shift settings, the report is constructed only from observations preceding the evaluation period. In the seen-period settings, it is a static aggregate profile constructed from the query-disjoint support subset and may include support dates later than the target date.
COVID-19 update	For targets in April–October 2020, the report is updated using observations available through March 2020, without using any check-ins from the target period.

Appendix

Table 7: Data separation and information available under the evaluation protocol.

Method	Seen period (1186 trajectories)				Regular unseen period (466 trajectories)				COVID-19 unseen period (689 trajectories)
Method	SD	SI	DARD	STVD	SD	SI	DARD	STVD	SD	SI	DARD	STVD
LSTM	0.0914	0.1375	0.2673	0.4161	0.1109	0.1698	0.3230	0.5779	0.1259	0.2073	0.3556	0.5725
MHSA	0.0908	0.1357	0.2468	0.4286	0.1210	0.1817	0.2706	0.5803	0.1262	0.2152	0.3044	0.5552
DeepMove	0.1777	0.1024	0.2348	0.4933	0.1645	0.1267	0.3180	0.6065	0.2308	0.1595	0.3340	0.6029
GetNext	0.0395	0.2820	0.2149	0.4635	0.0433	0.2834	0.2057	0.5601	0.0455	0.2952	0.1953	0.5294
TrajGAIL	0.2251	0.1256	0.0799	0.5836	0.2458	0.1227	0.1250	0.6576	0.2494	0.1118	0.1137	0.6424

Appendix

Table 8: Evaluation results on seen users in Osaka across seen and unseen time periods. The upper and lower baseline sections correspond to pretraining-based and prompting-based methods, respectively. Lower values are better; the best and second-best results are highlighted in bold and underlined, respectively.

Method	Seen period (2804 trajectories)				Regular unseen period (1655 trajectories)				COVID-19 unseen period (1301 trajectories)
Method	SD	SI	DARD	STVD	SD	SI	DARD	STVD	SD	SI	DARD	STVD
LSTM	0.1489	0.2437	0.3872	0.6207	0.1740	0.2420	0.3810	0.6569	0.1493	0.2645	0.4358	0.6551
MHSA	0.1897	0.1965	0.3578	0.5961	0.1715	0.2352	0.3762	0.6466	0.1582	0.2500	0.4182	0.6589
DeepMove	0.2754	0.2267	0.3548	0.6510	0.2746	0.2257	0.3876	0.6714	0.2709	0.2338	0.3918	0.6704
GetNext	0.0832	0.3220	0.2568	0.5752	0.0752	0.3215	0.2376	0.6408	0.0994	0.3351	0.2511	0.6438
TrajGAIL	0.1737	0.0825	0.1287	0.5750	0.1954	0.0912	0.1578	0.6513	0.1915	0.0943	0.1427	0.6532

Appendix

Table 9: Evaluation results on unseen users in Osaka across seen and unseen time periods. The upper and lower baseline sections correspond to pretraining-based and prompting-based methods, respectively. Lower values are better; the best and second-best results are highlighted in bold and underlined, respectively.

Seen users
Backbone	Seen period				Regular unseen period				COVID-19 unseen period
Backbone	SD	SI	DARD	STVD	SD	SI	DARD	STVD	SD	SI	DARD	STVD
Qwen3.5-9B	0.0080	0.0088	0.0671	0.3552	0.0096	0.0116	0.0902	0.4464	0.0452	0.0173	0.1396	0.5728
Gemma-4-E4B	0.0081	0.0191	0.0828	0.4041	0.0114	0.0192	0.0953	0.4731	0.0425	0.0200	0.1300	0.5431
Gemma-2-9B	0.0099	0.0064	0.0641	0.3760	0.0097	0.0118	0.0889	0.4296	0.0482	0.0138	0.1319	0.5696
Unseen users

Appendix

Table 10: MotiveMob results with different LLM backbones for seen and unseen users across the three temporal settings. All values are JSD; lower is better.

State generation	END predictor	Mean length	Max-step rate	SI $\downarrow$	DARD $\downarrow$
Direct, w/o $P_{M}$ (IWM)	Direct predictor	7.52	1.48%	0.0282	0.1401
$P_{S}$ (IWM)	$P_{M}$	8.08	2.27%	0.0205	0.1087
$P_{S}$ (positive-only)	$P_{M}$	7.42	2.06%	0.0176	0.1132
Rule-based state predictor	$P_{M}$	9.06	1.76%	0.0215	0.1163
Ground truth	–	6.82	–	–	–

Appendix

Table 11: Trajectory-length and termination statistics for the motivation-mediation ablation. Mean length and termination rates are aggregated over all 4,671 test days from the six evaluation splits. SI and DARD are macro averages over the six splits. The maximum rollout length is 25 states.

Figure 4: Trajectory comparisons with and without the trained motivation predictor. From left to right, the columns show the ground-truth trajectory, MotiveMob with the motivation predictor, and the corresponding ablation without it. Numbered markers indicate visit order, and labels show the visit time and POI.

Figure 5: Motivation-predictor sensitivity to IWM initialization and training duration. Each evaluation split contains four $3\times 3$ matrices corresponding to SD, SI, DARD, and STVD. Rows show results after 4, 8, and 12 epochs of motivation-predictor training, while columns show direct IL, balanced IWM initialization ( $p_{+}=0.5$ ), and observed-only IWM initialization ( $p_{+}=1.0$ ). Cell annotations report raw JSD values; lower values and lighter shading indicate better distributional alignment.

Initialization	Epoch	SD	SI	DARD	STVD	Mean rank
IL	4	0.0200	0.0372	0.1550	0.4771	5.42
IL	8	0.0204	0.0208	0.1200	0.4898	5.04
IL	12	0.0203	0.0198	0.1043	0.4857	3.81
IWM, $p_{+}=0.5$	4	0.0209	0.0378	0.1616	0.4742	5.42
IWM, $p_{+}=0.5$	8	0.0213	0.0192	0.1134	0.4949	4.92
IWM, $p_{+}=0.5$	12	0.0204	0.0207	0.1109	0.4969	5.02

Appendix

Table 12: Mean JSD over the six evaluation splits. All metrics are lower-is-better. Mean rank is calculated by ranking the nine configurations separately for every test split and metric and then averaging the resulting 24 ranks.

User	Period	Days	States	Resampling	Resampling	Correction	Feasibility	No feasible	Repetition
status				required	resolved	invoked	corrected	substitute	terminated
Seen	Seen	908	7,647	932 (12.2%)	528 (6.9%)	404 (5.3%)	231 (3.0%)	173 (2.3%)	121 (13.3%)
Seen	Regular unseen	690	5,446	686 (12.6%)	415 (7.6%)	271 (5.0%)	134 (2.5%)	137 (2.5%)	88 (12.8%)
Seen	COVID-19 unseen	392	3,218	830 (25.8%)	459 (14.3%)	371 (11.5%)	237 (7.4%)	134 (4.2%)	77 (19.6%)
Unseen	Seen	1,345	10,066	2,582 (25.7%)	1,534 (15.2%)	1,048 (10.4%)	629 (6.2%)	419 (4.2%)	207 (15.4%)
Unseen	Regular unseen	795	6,016	1,610 (26.8%)	929 (15.4%)	681 (11.3%)	432 (7.2%)	249 (4.1%)	150 (18.9%)

Appendix

Table 13: Inference-time validation and termination statistics. All percentages are normalized by generated states, except repetition termination, which is normalized by generated trajectories.

Figure 6: Ground-truth and predicted spatial mobility distributions for four additional evaluation settings. From top to bottom, the rows show seen users in the seen period, seen users in the regular unseen period, seen users in the COVID-19 unseen period, and unseen users in the seen period. Darker red indicates higher visit density.

Explore similar work

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
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
Relative Transitions, Not Absolute Destinations: A Transfer-and-Ground Framework for Target-Trajectory-Free Human Mobility Generation
Oct 1, 2026Yidi Wang, Yunhe Zhang, Bangchao Deng +2Trajectory Generation
SKL-IOTSC, Department of CIS, University of Macau Macau SAR, China

Jun 1, 2026, cs.AIJ/K move · Enter open · S save

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

Junlin He, Yihong Tang, Tong Nie, Ao Qu+5

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

Human Behavior Simulation Self-Evolving Agents

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

TrajGenAgent: A Hierarchical LLM Agent for Human Mobility Trajectory Generation

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