CAD-Mob: A Unified Architecture for Human Mobility Prediction Integrating Large Model Reasoning, Causal Inference, and Diffusion Models

CAD-Mob proposes an innovative unified agent causal architecture that integrates large language model reasoning, causal inference, and diffusion models. It achieves zero-shot next-location prediction and sparse trajectory completion, opening up a new technical path for human mobility modeling. This architecture combines three cutting-edge technologies to enhance prediction accuracy, interpretability, and robustness.

Human mobility prediction is one of the core technologies in fields such as location services, intelligent transportation, and urban planning. Traditional prediction methods rely on statistical patterns from historical trajectory data but struggle to capture deep causal relationships and contextual semantics in complex mobility behaviors. With the rapid development of large language models (LLMs) and generative AI, researchers have begun exploring the integration of semantic understanding and causal reasoning capabilities into mobility modeling to improve prediction accuracy and interpretability.

AgentMove is the semantic understanding core of CAD-Mob. It leverages the strong reasoning capabilities of large language models to extract mobility intentions and contextual information from natural language descriptions. It can understand semantically rich behavior descriptions like "going to the gym after work" and convert them into structured mobility features, enabling the model to have zero-shot prediction capabilities—even when facing unseen location types or behavior patterns, it can make reasonable predictions based on common sense reasoning.

Mobility data often has selection biases and confounding factors. The causal inference layer filters out spurious correlations by identifying true causal effects, ensuring the model learns stable and transferable patterns. This layer uses advanced causal discovery algorithms and counterfactual reasoning techniques, allowing it to answer causal questions such as "How would the arrival time change if the user chose a different mode of transportation?" and significantly improving robustness in out-of-distribution scenarios.

ProDiff is the generative core of CAD-Mob, innovatively applying diffusion models to spatiotemporal trajectory data. It can generate complete and coherent mobility paths based on partially observed trajectory fragments, effectively solving the data sparsity problem. The progressive generation feature of diffusion models also allows fine-grained control over the generation process, outputting trajectories that better align with real human behavior patterns.

CAD-Mob excels in three key tasks: 1. Zero-shot next-location prediction: Using LLM common sense knowledge, it can predict new location types without large amounts of labeled data, solving the cold start problem; 2. Sparse trajectory completion: Reconstructing complete trajectories based on limited observation points, addressing data incompleteness issues like GPS interruptions; 3. Interpretable modeling: The causal inference layer provides interpretability for predictions, making it suitable for human-machine collaboration scenarios such as intelligent navigation recommendations and abnormal behavior detection.

The greatest innovation of CAD-Mob lies in the organic integration of three independent technologies: large language models provide semantic understanding and zero-shot capabilities, causal inference ensures robustness and interpretability, and diffusion models are responsible for high-quality trajectory generation. The modular design allows independent optimization and replacement of each component—for example, replacing the full-version LLM with a lightweight one, or adjusting the diffusion model's sampling strategy to adapt to real-time scenarios—flexibly meeting different application needs.

CAD-Mob marks a new stage in human mobility research—shifting from purely data-driven approaches to an agent paradigm that integrates knowledge, causality, and generative capabilities. With the development of multimodal large models and embodied intelligence, more powerful systems are expected in the future that can simultaneously understand multiple information sources such as text, images, and voice, achieving deep understanding and accurate prediction of human mobility behaviors.