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

arXiv:2610.11442 · cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation".

Tom: The gist Human mobility generation can be characterized as a “why–where–when” decision process,

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So, wrapping up "MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation," the authors are proposing that modeling mobility as a sequence of motivation-conditioned check-in transitions is a solid approach.

Jane: They’ve shown that combining the step-level prediction of motivation proxies with motivation-conditioned state generation and inference-time rollout validation yields strong overall distributional fidelity across different user groups and temporal periods.

Lu: The main implication is that using these motivation proxies as intermediate representations can be really useful for generating mobility trajectories.

Meng: It moves us toward models that don't just see a list of places, but understand the underlying semantic reason driving the movement, which is something we need for building truly robust systems.

Lalam: The framework relies on three main components: the motivation predictor to infer motivation or termination, the state predictor to ground that in a location and time, and an implicit world model component learned through observed and alternative transitions.

Tom: The paper confirms that this combination results in strong fidelity under user-level and temporal distribution shifts, which is where many other methods struggle.

Jane: So for anyone listening who only cares about the practical use, it means the system is better at keeping a consistent activity pattern over time, even when things change.

Conclusion: Tom: So, this paper, "MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation," it boils down to using motivation to figure out how people move around and then making sure those moves actually make sense over time.

Jane: It's about breaking down movement into these little steps where the AI first guesses why someone is moving, then figures out where and when that move should happen.

Lu: The core idea is that you don't just track coordinates; you track the *reason* behind the movement, which gives it more meaning.

Meng: From an engineering standpoint, it means we can build a system that predicts not just "where" someone will go next, but "why" they're going there.

Lalam: The AI learns to predict this motivation and then uses that guess to ground the next location and time, creating a loop where prediction feeds back into the prediction.

Tom: So if we look at what the authors did, they put together three main pieces: a predictor for motivation, a predictor for state, and this implicit world model component that learns from both what actually happened and what might have been.

Jane: And those results showed it performs better than other methods when things change—like when users start moving in new ways or when the time of day shifts.

Lu: Specifically, they saw a reduction in distribution divergence, like ten point seven percent for seen users and eighteen point three percent for unseen ones, which is pretty solid compared to baselines.

Meng: That fidelity under shifts is key because real-world data isn't always perfectly consistent over long periods of time.

Tom: It’s about getting the AI to learn the underlying patterns that drive human behavior, rather than just memorizing routes or times.

Jane: And it suggests that these motivation proxies are actually useful intermediate steps for generating those complex mobility trajectories we need.

Lalam: If this works well, it means the AI can build a model of human culture and routine based on these semantic actions, which could change how we think about personal movement patterns in general.

Tom: So yeah, it’s a framework that uses motivation as the bridge between what someone *wants* to do and where they actually end up going.

Jane: And now we can start thinking about how this kind of understanding could apply to other complex decision-making processes outside of just walking or driving.

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

The University of Osaka

cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: The gist Human mobility generation can be characterized as a “why–where–when” decision process, and MotiveMob introduces a motivation-driven autoregressive framework that first forms a

Key concepts

Why–Where–When Decision Process: This is the core logic of MotiveMob, breaking down human movement into three sequential questions: why do I want to move (motivation), where should I go (location), and when should I arrive (time). The system uses this step-by-step approach to build a complete daily trajectory.
Motivation Predictor (PM): The PM predicts the latent motivation for the next movement or decides if the current activity sequence should end. It takes the current state, long-term behavioral reports, and recent history as input to infer what drives the user's next action. This helps determine if a trajectory should continue or terminate.
State Predictor (PS): The PS takes a predicted motivation from the PM and grounds it in a concrete candidate location and arrival time. Before accepting this candidate, the PS performs checks for speed feasibility (e.g., not exceeding 50 km/h) and repetition to ensure the proposed move is realistic.
Alternative Transition Augmentation: To improve training beyond just observed data, MotiveMob creates pseudo alternative transitions. It pairs known states with different motivation labels and uses an LLM judge to select relevant candidates, creating a larger, balanced dataset for training the State Predictor.

Summary

The gist Human mobility generation can be characterized as a “why–where–when” decision process, and MotiveMob introduces a motivation-driven autoregressive framework that first forms a hypothesis about why the next movement may occur and then jointly generates where and when it may occur.

MotiveMob Framework

MotiveMob is an LLM-based autoregressive framework that generates daily mobility trajectories through step-wise “why–where–when” decisions. At each step, a motivation predictor PM conditions on the current mobility state, a long-term behavioral report R, and a recent mobility history H to predict the motivation for the next movement or determine whether the trajectory should terminate. Given the hypothesized motivation, a state predictor PS grounds it in a candidate next location and arrival time. Before being fed back as the current mobility state for the next “why–where–when” decision, the candidate is subject to speed-feasibility and repetition checks.

Alternative Transition Augmentation

To train PS beyond observed transitions, MotiveMob constructs pseudo alternative transitions by pairing observed states with alternative motivation labels. This involves a two-stage pipeline where a lightweight Transformer-based candidate generator produces candidates, and an LLM-based judge evaluates them using the user’s historical information Hi, the specified motivation m j i, and their spatial and temporal distances from the current state. The selected states form a pool of pseudo-labeled alternative transitions denoted by Dalt. The augmented training set is then defined as Daug = Dobs ∪ D(ρ)alt, balancing observed and alternative supervision during training.

Predictor Training Objectives

The state predictor PS is trained to model the resulting next-state distribution as pθS(s' ci, m), minimizing the average negative log-likelihood over the augmented transition set Daug. The motivation predictor PM is trained to infer a step-level potential motivation or predict trajectory termination from context ci = (si, Ru, Hu d), where it minimizes the average negative log-likelihood of the observed and terminal labels DM.

Inference-Time Rollout Validation

At inference time, PM and PS operate alternately to generate a daily trajectory through step-level “why–where–when” decisions. The process involves:

PM greedily predicting mˆ i = arg max y∈M∪END pθM (y ci) and terminating if mˆ i = END.
PS sampling a provisional next state s˜i+1 ∼ pθS(· ci, mˆ i).
Validation checks include speed-feasibility (vmax = 50 km/h) and repetition checks.

If a proposal exceeds vmax, it is rejected and regenerated, or re-grounded to a feasible location of the same activity category using the user’s observed mobility information.

Evaluation Results

MotiveMob consistently achieves better distributional fidelity than competitive pretraining-based and prompting-based methods under user-level and temporal distribution shifts. Specifically, MotiveMob reduces the corresponding divergence by 10.7% for seen users and 18.3% for unseen users relative to the best-performing baseline in each setting. Spatial analysis showed that MotiveMob better preserves core activity regions for unseen users under temporal shifts.

Ablation and Sensitivity Analysis

Removing the behavioral report Ru worsens SI and STVD in all six settings, confirming its role as a long-term prior over when and where a user typically performs an activity. Removing Hu d worsens SI in all settings and substantially increases DARD and SD for both user groups. The motivation predictor PM contributes to improved temporal and activity-semantic coherence, as removing it increases both SI and DARD across all six evaluation settings. Furthermore, one epoch of IWM pretraining on a balanced mixture of observed and alternative transitions helps the motivation predictor learn SI and DARD more quickly after 8 epochs of motivation predictor training.

Conclusion

By combining step-level prediction of motivation proxies, motivation-conditioned state generation, and inference-time rollout validation, MotiveMob achieved the strongest overall distributional fidelity across the evaluated user groups and temporal periods. The findings support motivation proxies as useful intermediate representations for mobility trajectory generation.

How it works

The framework decomposes mobility generation into a sequence of motivation-conditioned check-in transitions. At each step, PM predicts the motivation and PS grounds it in a feasible next location and arrival time. The training objective for PS is to model the resulting next-state distribution as pθS(s' ci, m), combining factual grounding with broader coverage of motivation-conditioned mobility dynamics.

Key Components

The framework relies on three main components:

Motivation Predictor (PM): Infers the latent motivation for the next move or determines trajectory termination.
State Predictor (PS): Grounds a given motivation in a candidate next location and arrival time.
Implicit World Model (IWM): PS is trained using observed and alternative transitions to model the dynamics component of an implicit world model.

Inference Process

During inference, PM and PS operate sequentially to generate a daily trajectory through step-level “why–where–when” decisions. The process begins with a START token, followed by PM predicting the first motivation and PS generating the initial physical state s0. Subsequent steps alternate between motivation prediction and state prediction until termination conditions are met.

Limitations

MotiveMob relies on a predefined motivation taxonomy and pseudo-labeled alternative transitions whose correspondence to unobserved user intentions cannot be directly verified. The IWM augmentation exhibits a metric-specific trade-off, where improvements in the joint time–location distribution do not consistently coincide with improvements in the step-distance distribution. Future work could explore adaptive motivation representations and hierarchical sequential objectives.

Improvements for AI systems

Bold header: Motivation-driven autoregressive framework for human mobility generation

Human mobility generation can be characterized as a “why–where–when” decision process, and MotiveMob explicitly models this by first forming a hypothesis about why the next movement may occur and then jointly generating where and when it may occur.

Bold header: Offline alternative transition augmentation procedure for state predictor training

We develop an offline alternative transition augmentation procedure for training the state predictor as an IWM, where a lightweight candidate generator proposes outcomes for hypothetical motivations, while an LLM-based judge selects the most plausible candidate to train PS beyond observed transitions.

Bold header: Motivation-conditioned trajectory generation via sequential decision making

At inference time, PM and PS operate sequentially to generate a daily trajectory through step-level “why–where–when” decisions, where PM determines why the next move occurs, while PS grounds the predicted motivation in where and when it occurs.

Bold header: Implicit world model training using motivation-conditioned dynamics

We train the state predictor PS as an IWM of motivation-conditioned mobility dynamics by modeling si+1 ∼ T (· si, mi, Ru, Hu,d, d), which characterizes the distribution of the next location and arrival time resulting from a motivation under the user’s behavioral context.

Bold header: Robust generalization to out-of-distribution mobility patterns

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."

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.