TLDR: This research introduces RETAIL, a new dataset for real-world travel planning that addresses implicit queries, environmental factors, and detailed plan generation. It also proposes TGMA, a topic-guided multi-agent framework that significantly improves the feasibility and detail of AI-generated travel plans, outperforming existing models on complex real-world scenarios.
Travel planning, a seemingly straightforward task, often becomes complex when dealing with real-world scenarios. While large language models (LLMs) have made strides in automating travel plan generation, they still fall short in several key areas. A new research paper introduces a novel dataset called RETAIL and a framework named TGMA, aiming to bridge this gap and bring LLM-based travel planning closer to real-world utility.
The researchers identified three primary challenges with existing travel planning systems powered by LLMs. Firstly, these systems typically assume users provide explicit and detailed queries. In reality, people often express their travel requirements implicitly, with preferences evolving through conversation. Secondly, current solutions frequently overlook crucial environmental factors like weather conditions or real-time traffic, and they struggle to adapt to dynamic user preferences. This limits the practicality and feasibility of the generated plans. Lastly, existing systems tend to produce basic itineraries, lacking the rich, all-in-one details—such as specific timing, images of points of interest (POIs), ticket pricing, and user-generated content (UGC) enriched room information—that are essential for a truly practical travel plan.
To address these limitations, the researchers constructed RETAIL, which stands for REal-world Tourism All-in-one Interactive pLanning dataset. This comprehensive dataset is designed to support decision-making for implicit queries, while also covering explicit queries and scenarios requiring plan revisions. RETAIL incorporates environmental awareness to ensure plans are feasible under real-world conditions and includes detailed POI information for creating all-in-one travel plans. The dataset is extensive, comprising 10,182 real-world cases, covering 24 major Chinese cities, and integrating information for over 60,000 points of interest, along with transportation and weather data.
Building upon the RETAIL dataset, the team proposed TGMA, a topic-guided multi-agent framework. TGMA is designed to improve the generation of real-world travel plans through two main components. The first is a Topic-Guided Interaction Logic, which dynamically selects appropriate conversation topics based on the dialogue context. This ensures natural and relevant interactions, guiding users from initial implicit requirements to specific preferences. The second component is a Multi-Agent Architecture, which breaks down complex planning tasks into manageable steps. This architecture includes three specialized agents:
Intent Detection Agent
This agent transforms unstructured conversational input into structured representations, extracting key travel parameters like location, time constraints, traveler details, accommodation preferences, and budget. It also captures user modifications throughout the conversation.
Overall Plan Agent
This agent coordinates high-level planning decisions, systematically processing day allocation, attraction planning, and transit arrangements. It uses an integrated toolset to balance planning objectives with operational constraints, creating coherent and executable travel recommendations.
Also Read:
- Enhancing LLM Perspective-Taking: A Look at Structured Thought-Action Sequences
- Advancing Personalized AI: Memory Mechanisms for Complex Reasoning
Detailed Plan Agent
The final agent enriches the high-level plans with comprehensive, executable details. It integrates nearby POI information, attraction tickets, accommodation options, and restaurant recommendations, generating optimized daily arrangements that consider travel efficiency and user comfort.
The experimental results highlight the significant challenges in real-world travel planning. Even the strongest existing models achieved a mere 1.0% pass rate on the RETAIL dataset, underscoring the complexity of the task. In stark contrast, TGMA demonstrated substantially improved performance, achieving a 2.72% pass rate. This improvement, while seemingly small, represents a significant leap in a highly challenging domain, offering promising directions for future research.
Further analysis revealed the critical impact of various constraints on the final pass rate of travel plans. For instance, “Time Interval” and “POI Validation” were identified as major commonsense constraints affecting plan feasibility, suggesting the need for better activity scheduling and verification of points of interest. Among user preferences, “Hotel Type” and “Required Sites” had the most significant influence, indicating that plans often struggle to meet specific accommodation and must-visit site requirements. The research also emphasized the importance of decision-making support, multimodal information (like user reviews and visual content), and environmental awareness in enhancing travel planning quality.
In conclusion, this work sheds light on the critical limitations of current LLM-based travel planning systems and offers a robust solution. By introducing the RETAIL dataset and the TGMA framework, the researchers have provided valuable resources and insights for developing more practical and user-aligned travel planning systems. For more details, you can refer to the full research paper here: RETAIL: Towards Real-world Travel Planning for Large Language Models.
