DaiLLM: An Agent Framework Based on a Large Language Model and Reinforced by AI to Generate Daily Activity Trajectories

ABSTRACT Simulating human mobility is crucial for urban planning, environmental exposure assessment, and social behavior studies. However, existing methods often rely on rules and assumptions, making it difficult to provide robust results and explicit decision‐making processes when generating plausible daily activity trajectories from sparse location records such as check‐ins. To bridge these gaps, we propose a large language model (LLM)‐based agent framework that innovatively incorporates the Chain‐of‐Thought (CoT) approach and Reinforcement Learning from AI Feedback (RLAIF) mechanisms, namely Daily mobility LLM (DaiLLM). This framework consists of three modules: the Profiler derives user characteristics from historical trajectory data; the Generator simulates movement through the CoT process, which integrates constraints from Profiler outputs; and the Discriminator employs RLAIF to evaluate the similarity between generated results and real data, providing feedback to optimize the results. Experiments demonstrate that: (1) our method improves accuracy by 34.2% compared to traditional approaches in daily activity trajectory generation; (2) the CoT approach enhances activity logic and results in a 26.1% increase through task decomposition and constraint integration; (3) the RLAIF mechanism explicitly provides an optimization process through Latent Dirichlet Allocation (LDA) analysis; and (4) our method excels in transfer learning and long‐term generation across different city contexts. This study provides a robust, transparent, and cost‐effective solution for plausible daily activity trajectory generation from sparse check‐in data, which will establish a new paradigm for mobility behavior analysis in data‐scarce scenarios.