TRIP: A Bi-level Travel Routing Intelligent Planner Informed from Human Planning Behaviors

Abstract

Tourism has become a popular activity with the improvement of living standards, often requiring travelers to search for guides and plan routes that match their preferences while considering constraints like budget and distance, which can be both time-consuming and mentally taxing. Recently, large language models (LLMs) have shown strong reasoning abilities in complex tasks, making the development of an LLM agent for planning personalized travel itineraries that meet these constraints highly valuable. However, existing travel planning agents typically rely on structured POI resources, generic planning workflows, or limited external knowledge, and they often struggle to jointly satisfy subjective preferences and objective hard constraints in realistic planning scenarios. To address this, we propose TRIP, a bi-level Travel Routing Intelligent Planner, which consists of macro-planning based on the user’s personalized requirements and micro-planning to consider hard constraints. The macro-planning utilizes hybrid retrieval techniques to search for real-world travel guides that are most similar to the user’s personalized requirements. The micro-planning dynamically fine-tunes the itineraries by invoking relevant tools and reflecting on the user’s hard constraints. This approach is closely aligned with human-like planning behaviors. Experimental results on real-world and simulated datasets demonstrate that the proposed agent generates travel itineraries that better satisfy personalized requirements and hard constraints. The travel planning system has been deployed for user testing at the Hangzhou West Lake Scenic Area. 1,047 participants provided an average satisfaction score of 3.5 out of 5 for the generated itineraries, which is significantly higher compared to existing tourism planning methods.