Fine-tuning LLMs for Tourist Trajectory Prediction using Field Experiment Data

Recent research demonstrates a novel application of fine-tuned Large Language Models (LLMs) for predicting tourist behavior, specifically trajectories within a destination. This method addresses limitations of traditional prediction models by leveraging LLMs' capacity for integrating heterogeneous contextual data and common-sense reasoning, adapting general knowledge to specific local patterns. Validation at Wakayama Castle Park, Japan, indicates the effectiveness of this approach.

Why it matters

This development is crucial for entities involved in urban planning, tourism management, and infrastructure development, as it offers a sophisticated tool for anticipating visitor movements. Accurate prediction enables proactive resource allocation, optimized visitor flow, and more effective interventions, thereby enhancing operational efficiency and visitor experience within specific sites or broader regions.

Key insights

• Traditional methods for predicting tourist behavior struggle to generalize due to the context-dependent nature of visitor decisions (e.g., weather, fatigue).
• Large Language Models (LLMs) can address these limitations by encoding common-sense knowledge about human behavior from pretraining.
• LLMs enable reasoning about context-dependent decisions and flexibly integrate diverse information through natural language representation.
• Fine-tuning LLMs on local trajectory data adapts their general understanding to destination-specific patterns.
• The Llama-3.1-8B model, fine-tuned with 566 trajectories from Wakayama Castle Park, Japan, showed success in predicting tourist movements.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.20830