[2603.04750] HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel

Computer Science > Artificial Intelligence arXiv:2603.04750 (cs) [Submitted on 5 Mar 2026] Title:HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel Authors:The Viet Bui, Wenjun Li, Yong Liu View a PDF of the paper titled HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel, by The Viet Bui and 2 other authors View PDF HTML (experimental) Abstract:Sequential LLM agents fail on long-horizon planning with hard constraints like budgets and diversity requirements. As planning progresses and context grows, these agents drift from global constraints. We propose HiMAP-Travel, a hierarchical multi-agent framework that splits planning into strategic coordination and parallel day-level execution. A Coordinator allocates resources across days, while Day Executors plan independently in parallel. Three key mechanisms enable this: a transactional monitor enforcing budget and uniqueness constraints across parallel agents, a bargaining protocol allowing agents to reject infeasible sub-goals and trigger re-planning, and a single policy trained with GRPO that powers all agents through role conditioning. On TravelPlanner, HiMAP-Travel with Qwen3-8B achieves 52.78% validation and 52.65% test Final Pass Rate (FPR). In a controlled comparison with identical model, training, and tools, it outperforms the sequential DeepTravel baseline by +8.67~pp. It also surpasses ATLAS by +17.65~pp and MTP by +10.0~pp. On FlexTravelBench multi-turn...