COLA: Cross-city Mobility Transformer for Human Trajectory Simulation Yu Wang, Tongya Zheng, Shunyu Liu, Yuxuan Liang, Mingli Song This paper proposes COLA, a Cross-city mObiLity trAnsformer for human trajectory simulation, which addresses the challenge of data scarcity in urban mobility simulation. COLA leverages a model-agnostic transfer framework to transfer knowledge across cities, enabling the generation of realistic human mobility data for downstream tasks. The main challenges in cross-city mobility transfer include domain heterogeneity and subtle differences in long-tail frequency distributions of locations. COLA divides the Transformer into private modules for city-specific characteristics and shared modules for city-universal mobility patterns. It also employs a post-hoc adjustment strategy to calibrate the model's predictions for city-specific characteristics. Extensive experiments show that COLA outperforms state-of-the-art single-city baselines and cross-city baselines in human trajectory simulation. COLA is effective in generating high-quality synthetic data for practical applications, such as location prediction, and is robust to hyperparameter settings. The code is available at https://github.com/Star607/Cross-city-Mobility-Transformer.COLA: Cross-city Mobility Transformer for Human Trajectory Simulation Yu Wang, Tongya Zheng, Shunyu Liu, Yuxuan Liang, Mingli Song This paper proposes COLA, a Cross-city mObiLity trAnsformer for human trajectory simulation, which addresses the challenge of data scarcity in urban mobility simulation. COLA leverages a model-agnostic transfer framework to transfer knowledge across cities, enabling the generation of realistic human mobility data for downstream tasks. The main challenges in cross-city mobility transfer include domain heterogeneity and subtle differences in long-tail frequency distributions of locations. COLA divides the Transformer into private modules for city-specific characteristics and shared modules for city-universal mobility patterns. It also employs a post-hoc adjustment strategy to calibrate the model's predictions for city-specific characteristics. Extensive experiments show that COLA outperforms state-of-the-art single-city baselines and cross-city baselines in human trajectory simulation. COLA is effective in generating high-quality synthetic data for practical applications, such as location prediction, and is robust to hyperparameter settings. The code is available at https://github.com/Star607/Cross-city-Mobility-Transformer.