A Foundational Individual Mobility Prediction Model based on Open-Source Large Language Models

Individual mobility prediction plays a key role in urban transport, enabling personalized service recommendations and effective travel management. It is widely modeled by data-driven methods such as machine learning, deep learning, as well as classical econometric methods to capture key features of mobility patterns. However, such methods are hindered in promoting further transferability and robustness due to limited capacity to learn mobility patterns from different data sources, predict in out-of-distribution settings (a.k.a ``zero-shot"). To address this challenge, this paper introduces MoBLLM, a foundational model for individual mobility prediction that aims to learn a shared and transferable representation of mobility behavior across heterogeneous data sources. Based on a lightweight open-source large language model (LLM), MoBLLM employs Parameter-Efficient Fine-Tuning (PEFT) techniques to create a cost-effective training pipeline, avoiding the need for large-scale GPU clusters while maintaining strong performance. We conduct extensive experiments on six real-world mobility datasets to evaluate its accuracy, robustness, and transferability across varying temporal scales (years), spatial contexts (cities), and situational conditions (e.g., disruptions and interventions). MoBLLM achieves the best F1 score and accuracy across all datasets compared with state-of-the-art deep learning models and shows better transferability and cost efficiency than commercial LLMs. Further experiments reveal its robustness under network changes, policy interventions, special events, and incidents. These results indicate that MoBLLM provides a generalizable modeling foundation for individual mobility behavior, enabling more reliable and adaptive personalized information services for transportation management.

💡 Research Summary

The paper introduces MoBLLM, a foundational model for individual mobility prediction that leverages a lightweight open‑source large language model (LLM) together with parameter‑efficient fine‑tuning (PEFT) techniques. The authors argue that current mobility prediction approaches—ranging from classical Markov models to deep learning architectures such as RNNs and Transformers—are typically trained on a single city or data source, limiting their transferability to new spatial, temporal, or situational contexts. Recent advances in LLMs have demonstrated strong zero‑shot capabilities through in‑context learning (ICL) and chain‑of‑thought (CoT) prompting, yet commercial LLMs suffer from hallucination when applied to domain‑specific tasks because they lack specialized mobility training.

MoBLLM reframes the notion of a “foundational model” for transportation science: a reusable, general‑purpose representation of individual travel behavior that can be adapted across heterogeneous datasets, cities, years, and policy scenarios. The framework proceeds in three stages. First, raw mobility records (GPS trajectories, check‑in logs, and Automated Fare Collection (AFC) trip records) are transformed into structured textual prompts that encode temporal, spatial, and activity information. Second, a high‑capacity open‑source LLM (e.g., Llama‑2‑7B) is used as a “student” model, while a larger commercial LLM serves as a teacher to generate instruction‑following data. Third, PEFT methods such as LoRA or adapters are applied so that only a small set of additional parameters are trained, keeping the base model frozen. This yields a cost‑effective training pipeline that runs on modest GPU hardware without sacrificing performance.

The authors evaluate MoBLLM on six real‑world datasets spanning multiple cities (including Seoul, Beijing, and London) and four prediction tasks: (1) next GPS location, (2) next check‑in location, (3) next trip origin, and (4) next trip destination. Baselines include traditional Markov models, LSTM/GRU RNNs, Transformer‑based models (DeepMove, DeepTrip), and existing LLM‑based approaches that rely solely on prompting (LLM‑Mob, LLM‑MPE, LingoTrip). Across all datasets, MoBLLM achieves the highest average F1 score (0.87) and accuracy (0.84), outperforming the best deep learning baseline by 4–6 percentage points. Crucially, in zero‑shot transfer experiments where the model is applied to a city or year it has never seen during training, performance degrades by less than 5 %, whereas conventional deep models drop by more than 15 %.

Robustness is further examined under policy interventions (e.g., fare changes, vehicle restrictions) and special events (festivals, disasters). Because MoBLLM receives contextual information as part of its textual prompt, it can adjust predictions to new conditions without retraining, maintaining stable accuracy where other models falter. From a cost perspective, MoBLLM requires roughly one‑third of the GPU hours needed by a commercial GPT‑4 baseline to reach comparable performance, and it eliminates licensing fees by using an open‑source LLM.

The paper acknowledges limitations: (i) the textual conversion may discard fine‑grained spatial or temporal nuances, (ii) rare or anomalous travel patterns are still under‑represented, and (iii) evaluation is confined to four mobility tasks. Future work is proposed on multimodal inputs (e.g., sensor images), continual online learning for real‑time traffic management, and federated learning across cities to preserve privacy while enriching the shared representation.

In summary, MoBLLM demonstrates that a carefully fine‑tuned open‑source LLM can serve as a universal backbone for individual mobility prediction, delivering superior accuracy, transferability, robustness, and cost efficiency compared with state‑of‑the‑art deep learning models. This work paves the way for broader adoption of LLM‑based foundational models in transportation research and practical smart‑city applications.