BUA-IRIP-LLM Releases Behavior2Trip: A New Benchmark for Personalized Travel Planning via User Behavior Trajectory

Summary:The BUA-IRIP-LLM team has introduced a new benchmark named Behavior2Trip, which aims to infer user preferences from past behavior trajectories for personalized travel planning. The benchmark consists of 11,400 instances, with each instance representing an average of 39.8 user behaviors across 14 attributes and 5 preference dimensions. The team also proposed the B2T-Agent, a reinforcement learning-based model that leverages user behavior trajectories for preference-aligned retrieval. Experiments

---

Core Breakthroughs

The BUA-IRIP-LLM team has released a new benchmark named Behavior2Trip, addressing the limitations of existing travel planning agents in personalized recommendations. The benchmark achieves breakthroughs in the following ways:

• Data Scale and Diversity: The dataset is collected from one of China's largest online travel platforms, comprising 11,400 instances, with each instance representing an average of 39.8 user behaviors across 14 attributes and 5 preference dimensions.
• Task Definition: The team introduced the "Behavior-Aware Travel Planning" task, which infers user preferences directly from past behaviors without relying on explicit user instructions or interactions.
• Model Innovation: The B2T-Agent, a reinforcement learning-based agent, was proposed. It leverages user behavior trajectories for preference-aligned retrieval and integrates with external tools for efficient information retrieval.

Technical Highlights

1. Data Construction: The Behavior2Trip benchmark features a meticulously constructed dataset that provides high-quality data support for research.
2. Reinforcement Learning Application: The B2T-Agent utilizes reinforcement learning to make dynamic decisions and optimizations in complex task environments.
3. Performance Advantage: In experiments, the B2T-Agent significantly outperformed GPT-4.1 and other baselines in the hardest tasks, demonstrating its strong capabilities in personalized travel planning.
4. Generalization Ability: The B2T-Agent also performed well in the TravelPlanner benchmark, proving its generalization ability across different task scenarios.

Industry Impact

• Enhanced Personalized Services: This research provides a new technical path for the travel planning industry, significantly improving the quality of personalized services and user satisfaction.
• AI Application Expansion: The method can be extended to other fields that require personalized recommendations, such as e-commerce, news recommendation, etc.
• Developer Recommendations: For developers, the Behavior2Trip benchmark and B2T-Agent model provide a powerful tool for developing smarter and more personalized travel planning systems.

Developer Recommendations

1. Data Utilization: Developers are advised to fully utilize the Behavior2Trip benchmark dataset for model training and evaluation.
2. Model Optimization: It is recommended to try combining other advanced AI technologies (such as multimodal learning, graph neural networks, etc.) to further optimize the performance of the B2T-Agent.
3. Application Expansion: Explore the possibility of applying this method to other fields, such as smart homes, virtual assistants, etc.

---

Source: ArXiv cs.AI (2026-08-27)

— END —

Tags: #BUA-IRIP-LLM #Behavior2Trip #Personalized Recommendation #Reinforcement Learning #Travel Planning