Go to ICLR 2025 Conference homepageChinaTravel: A Real-World Benchmark for Language Agents in Chinese Travel Planning
Recent advances in Large Language Models (LLMs), particularly in language reasoning and tool-use capabilities have sparked the rapid development of \emph{Language Agents} to assist humans across various real-world applications. Among these, travel planning stands out as a significant domain, presenting both academic challenges and practical value due to its inherent complexity and real-world relevance. However, existing travel plan benchmarks do not test language agents with human users or their ability to follow customized requirements, both of which are vital for deploying them in real-world applications. In this paper, we propose ChinaTravel, a new benchmark tailored to authentic Chinese travel requirements, aiming to provide a more realistic evaluation framework for future language agents. We collect the travel requirements through questionnaires and employ an efficient and faithful evaluation process with 46 metrics covering feasibility, constraint satisfaction, and preference comparison. Moreover, we identify three challenges in the real-world deployments of travel planning, including \emph{constraint recognition}, \emph{concept openness}, and \emph{customized preference}. The empirical studies show that even state-of-the-art neural-symbolic agents succeed in 51.3% constraint validation of human queries. Our findings point to the need for methods that can improve the ability of agents to understand diverse intentions or keep track of constraints with emerging concepts from human requirements.