Guiding LLM-Based Human Mobility Simulation with Mobility Measures from Shared Data

Large-scale human mobility simulation is critical for many science domains such as urban science, epidemiology, and transportation analysis. Recent works treat large language models (LLMs) as human agents to simulate realistic mobility trajectories by modeling individual-level cognitive processes. However, these approaches generate individual mobility trajectories independently, without any population-level coordination mechanism, and thus fail to capture the emergence of collective behaviors. To address this issue, we design M2LSimu, a mobility measures-guided multi-prompt adjustment framework that leverages mobility measures derived from shared data as guidance to refine individual-level prompts for realistic mobility generation. Our framework applies coarse-grained adjustment strategies guided by mobility measures, progressively enabling fine-grained individual-level adaptation while satisfying multiple population-level mobility objectives under a limited budget. Experiments show that M2LSimu significantly outperforms state-of-the-art LLM-based methods on two public datasets.

💡 Research Summary

**
The paper addresses a critical gap in large‑scale human mobility simulation: while recent works have leveraged large language models (LLMs) to generate realistic individual trajectories by modeling cognitive processes such as intention and reflection, they treat each agent independently. This independence prevents the emergence of population‑level collective behaviors that are well documented in urban science, epidemiology, and transportation research, such as the truncated‑power‑law distribution of travel distances and the Zipf‑like visitation frequency law.

To bridge this gap, the authors propose M2LSimu, a Mobility Measures‑guided Multi‑prompt Adjustment framework. The core insight is that key mobility measures (travel distance, radius of gyration, visitation frequency, etc.) are preserved even in coarse‑grained or aggregated public datasets. These measures therefore provide a privacy‑preserving “ground truth” that can guide the simulation without requiring full raw trajectories.

M2LSimu operates in an iterative loop. Initially, every individual receives a common base prompt plus a personalized user profile; the LLM generates a set of trajectories in parallel. The generated trajectories are evaluated against target mobility measures derived from shared data. The differences are fed to a secondary LLM, which produces coarse‑grained adjustment strategies (e.g., “increase short‑range trips for group 1”, “limit cross‑district exploration for group 2”). Each strategy is treated as an action in a Markov Decision Process (MDP) where the state is the current collection of prompts for the whole population.

Because the action space is large and each action requires re‑running the LLM simulation, the authors formulate the prompt‑adjustment problem as a multi‑objective optimization task and solve it with Monte Carlo Tree Search (MCTS). The reward function simultaneously rewards improvements in multiple mobility‑law objectives while penalizing regressions in others, ensuring balanced progress across spatial and temporal dimensions. A global action‑value estimator pre‑filters unlikely candidates, dramatically reducing computational overhead. The search is performed on a small representative subset of the population; the resulting prompt adjustments are then generalized to larger groups with similar profiles, further curbing cost.

Experiments on two public datasets (Beijing and Shanghai) compare M2LSimu against state‑of‑the‑art LLM‑based simulators such as LMmob. Evaluation metrics include complementary cumulative distribution functions of travel distance, cumulative distribution of radius of gyration, and the Zipf exponent of visitation frequency, measured via KL‑divergence and parameter estimation error. M2LSimu consistently outperforms baselines, achieving 11.29 %–64.08 % relative improvement across metrics. Notably, it corrects the over‑production of long‑distance trips and restores the heterogeneity of visitation patterns that baseline models miss.

A further ablation study shows that even when only statistical shared data (aggregated counts without individual trajectories) are available, the framework still yields measurable gains, confirming its practicality under strict data‑sharing constraints.

The paper’s contributions are threefold: (1) introducing mobility measures from shared data as external guidance for LLM prompt refinement, (2) designing a multi‑objective MDP‑based optimization pipeline powered by MCTS and a global value estimator, and (3) demonstrating that coordinated prompt adjustments can satisfy multiple population‑level mobility laws within a limited computational budget. The authors suggest future extensions to incorporate additional measures (e.g., time‑of‑day activity patterns, transport mode distributions) and to apply the framework to policy‑relevant simulations such as epidemic control or congestion mitigation.