Improving Spatial and Temporal Awareness of Large Language Models for Personalized Travel Planning

ACL ARR 2025 February Submission7466 Authors

Abstract: Travel itinerary planning requires coordinating multiple tasks, like flight and hotel bookings, while meeting spatial and temporal constraints. Although Large Language Models (LLMs) show promise in tackling complex planning tasks, they often struggle with maintaining geographical consistency, managing real-time travel constraints, and providing accurate logistical details. We present TravLinkTo, a system designed to enhance the spatial and temporal awareness of LLMs for travel itinerary planning. TravLinkTo ensures geographical coherence by leveraging city graphs to optimize routes across cities. It integrates LLM-generated content with real-time data from external sources, such as flight schedules and restaurant booking systems, through an iterative self-refinement process that utilizes automated query generation, execution, and parameterized planning templates. We evaluate TravLinkTo on the TravelPlanner dataset. Experimental results show that TravLinkTo significantly outperforms existing LLM-based methods, enhancing both the quality and efficiency of travel planning.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented, retrieval; knowledge augmented;commonsense reasoning; applications;

Contribution Types: NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 7466