CityWeave: Weaving User Needs and World Constraints for Personalized and Reliable Mobility Planning
CityWeave training and grounding framework. Source: paper.Abstract
CityWeave is a multimodal agent framework for personalized and reliable mobility planning under user and world constraints. It structures reasoning around Who, When, Where, and How, uses User-World Grounding to verify preferences and route feasibility, and trains the agent with supervised fine-tuning and reinforcement learning. Evaluations on 180,000 real-world planning records show strong gains in final pass rate, commonsense compliance, personalization, and reliability.
Publication
Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, 12173-12182
研究概览
CityWeave 面向真实的门到门出行规划，同时考虑用户画像、时间窗口、地图拓扑和交通方式等约束。方法将规划过程拆解为 Who、When、Where 和 How 四个结构化槽位，再通过 User-World Grounding 模块检查个性化需求与导航可行性。
核心方法与发现
- 以 3W1H 结构约束多模态智能体的长链路推理和工具调用。
- 联合优化用户偏好、地点存在性、路线连通性、换乘合法性和时刻表约束。
- 在 18 万条真实出行规划记录上进行训练和评测，显著提升最终方案通过率、常识约束通过率、个性化和可靠性。
资料来源
Ao Wang
Master’s Graduate
Qiang Xia
Ph.D. Student
Yi Zhou
Ph.D. Student
