Abstract
Mobility trajectory data provide essential support for smart city applications. However, such data are often difficult to obtain. Meanwhile, most existing trajectory generation methods implicitly assume that at least a subset of real mobility data from target city is available, which limits their applicability in data-inaccessible scenarios. In this work, we propose a new problem setting, called bus-conditioned zero-shot trajectory generation, where no mobility trajectories from a target city are accessible. The generation process relies solely on source city mobility data and publicly available bus timetables from both cities. Under this setting, we propose MobTA, the first approach to introduce task arithmetic into trajectory generation. MobTA models the parameter shift from bus-timetable-based trajectory generation to mobility trajectory generation in source city, and applies this shift to target city through arithmetic operations on task vectors. This enables trajectory generation that reflects target-city mobility patterns without requiring any real mobility data from it. Furthermore, we theoretically analyze MobTA's stability across base and instruction-tuned LLMs. Extensive experiments show that MobTA significantly outperforms existing methods, and achieves performance close to models finetuned using target city mobility trajectories.
Introduction
High-quality human mobility data is the lifeblood of smart city applications like urban planning, traffic management, and epidemic control. However, this data is usually locked up by telecom operators or ride-hailing companies due to strict commercial and privacy constraints. To circumvent this, researchers build generative models to synthesize statistically realistic trajectories. However, almost all existing models implicitly assume the availability of at least some real mobility data from the target city for training. If a city has no accessible data, these models fail.
This paper introduces a highly practical new setting: bus-conditioned zero-shot trajectory generation. Bus timetables are public, non-sensitive, and globally available. While bus routes do not perfectly capture how people move—missing first/last-mile walks and flexible trips—they provide a structural "backbone" of a city's layout and temporal rhythm. The challenge lies in bridging the gap between rigid bus timetables and actual human mobility without access to the target city's real data.
Methodology: Mobility Task Arithmetic (MobTA)
MobTA addresses this challenge using Task Arithmetic, a parameter-efficient model editing technique. The core intuition is that the behavioral gap between rigid bus routes and actual human movement is a transferable mathematical vector.
1. Unified Spatial Grid and Serialization
To make different cities mathematically comparable, the authors divide cities into 1km x 1km grids using a "spiral indexing" strategy. The city center is always tokenized as <LOC_0>, with indices spiraling outward. This ensures that <LOC_0> consistently represents "downtown" across different cities. Trajectories are then serialized into text (e.g., "arrival time is $t_i$, location is <LOC_i>") for processing by Large Language Models (LLMs).
2. Task Vector Construction
A "task vector" is defined as the difference in weights between a fine-tuned model and the base model. MobTA isolates the parameter shift between bus routes and real mobility in a source city: $$\Delta V = V^{src\_mob} - V^{src\_bus}$$ This $\Delta V$ represents the "human flexibility" behavior missing from bus schedules.
3. Zero-Shot Arithmetic
The learned shift is applied to the target city's bus-trained vector to estimate the target mobility vector: $$\hat{V}^{tgt\_mob} = V^{tgt\_bus} + \mu\Delta V$$ By adding these weights, the LLM learns to generate flexible human trajectories in the target city without ever seeing its real mobility data.
Experimental Evaluation
The experiments utilize real-world urban datasets from Shanghai (SH), Wuxi (WX), and Singapore (SG). The authors collected massive real mobility trajectory datasets (346K to 18.4M records) to serve as source data and ground-truth evaluation benchmarks.
Performance Comparison
MobTA was compared against several baselines, including traditional generative models (TimeGeo, MoveSim) and state-of-the-art cross-city transfer models (COLA). As shown in the results, MobTA significantly outperforms existing methods across various spatial metrics, including Distance (Dist.), Radius (Rad.), and Location diversity (Loc.).
Key Insights
- Transferability of the "Gap": The discrepancy between bus timetables and real mobility is a structured, behavioral deviation that remains remarkably consistent across different urban environments.
- Weight Algebra for Spatial Data: The study demonstrates that arithmetic on LLM weights can effectively transfer complex spatial behaviors. Subtracting the "bus" concept from the "mobility" concept leaves a pure behavioral vector that can be applied to new city structures.
- Robustness to Instruction Tuning: Theoretical and empirical analysis proves that MobTA works equally well on both Base and Instruction-Tuned LLMs. Because trajectory generation relies on a narrow parameter subspace (spatial tokens), it does not interfere with parameters responsible for general chat or safety alignment.
Conclusion
MobTA introduces a novel paradigm for zero-shot trajectory generation by leveraging task arithmetic. By treating urban mobility as a combination of structural public transit and flexible human behavior, it achieves near-supervised performance without requiring sensitive target-city data. Future work will explore dynamic scaling of arithmetic coefficients and the integration of richer metadata, such as topological graph relationships between bus stations, to further enhance generation fidelity.
Summary updated