COMPASS
收藏资源简介:
COMPASS是一个用于评估智能体在真实旅行规划场景中战略性地协调工具和优化用户偏好的多轮交互基准。数据集涵盖了20个美国国家公园的交通、住宿和票务等真实旅行数据库,以及一个全面反映商业预订平台的工具生态系统。数据集包含281个任务,分为三个难度级别,包括仅酒店、酒店加航班、酒店航班加许可证,旨在评估智能体在满足硬约束的同时优化用户偏好的能力。数据集适用于解决智能体在实际任务中优化用户偏好的问题。
COMPASS is a multi-turn interaction benchmark for evaluating AI Agents' ability to strategically coordinate tools and optimize user preferences within real-world travel planning scenarios. The dataset includes real travel databases pertaining to transportation, accommodation and ticketing for 20 U.S. National Parks, as well as a comprehensive tool ecosystem that reflects the landscape of commercial booking platforms. It contains 281 tasks divided into three difficulty tiers, with task types covering hotel-only bookings, hotel plus flight bookings, and hotel, flight plus permit bookings. The benchmark aims to evaluate agents' capacity to optimize user preferences while satisfying hard constraints. This dataset is applicable to addressing the problem of optimizing user preferences for AI Agents in practical task settings.

- 1通过哈佛大学、苹果公司 · 2025年



