TourPlanner: A Competitive Consensus Framework with Constraint-Gated Reinforcement Learning for Travel Planning

TL;DR

TourPlanner tackles the core challenges of travel itinerary generation: vast candidate POIs, reliance on a single reasoning path, and the joint optimization of hard and soft constraints. It fuses a Personalized Recall and Spatial Optimization (PReSO) preprocessing with a Competitive consensus Chain-of-Thought (CCoT) multi-path reasoning and a sigmoid-based Constraint-Gated Reinforcement Learning (RL) refinement. The framework demonstrates state-of-the-art performance on the TripTailor benchmark, achieving near-perfect feasibility, strong macro rationality, and improved spatial coherence across multiple LLM backbones. Its contributions include a robust multi-agent arbitration mechanism, spatially informed POI recall, and a curriculum-like RL gating strategy that progressively balances constraint satisfaction and personalization. Overall, TourPlanner advances practical, configurable, and scalable AI-assisted travel planning with strong potential for real-world deployment and personalization.

Abstract

Travel planning is a sophisticated decision-making process that requires synthesizing multifaceted information to construct itineraries. However, existing travel planning approaches face several challenges: (1) Pruning candidate points of interest (POIs) while maintaining a high recall rate; (2) A single reasoning path restricts the exploration capability within the feasible solution space for travel planning; (3) Simultaneously optimizing hard constraints and soft constraints remains a significant difficulty. To address these challenges, we propose TourPlanner, a comprehensive framework featuring multi-path reasoning and constraint-gated reinforcement learning. Specifically, we first introduce a Personalized Recall and Spatial Optimization (PReSO) workflow to construct spatially-aware candidate POIs' set. Subsequently, we propose Competitive consensus Chain-of-Thought (CCoT), a multi-path reasoning paradigm that improves the ability of exploring the feasible solution space. To further refine the plan, we integrate a sigmoid-based gating mechanism into the reinforcement learning stage, which dynamically prioritizes soft-constraint satisfaction only after hard constraints are met. Experimental results on travel planning benchmarks demonstrate that TourPlanner achieves state-of-the-art performance, significantly surpassing existing methods in both feasibility and user-preference alignment.

Figures (3)

• Figure 1: Overview of our TourPlanner Framework. Upon receiving a user query, the framework constructs candidate information via the PReSO workflow to support the CCoT process. The CCoT involves three main phases: (1) Agent Instantiation, followed by iterative cycles of (2) Parallel Proposal Generation and (3) Competition and Consensus Arbitration for each of the $d$ days. During arbitration, the system synthesizes the top-$k$ proposals into a daily plan. Finally, a Constraint-Gated RL module refines the consensus itinerary to simultaneously optimize both soft constraints (personalization) and hard constraints (feasibility).
• Figure 2: Comparison between prior methods.
• Figure 3: Recall Performance of PReSO versus TripTailor. Our workflow consistently outperforms TripTailor across various LLMs, achieving a higher proportion of identified ground-truth travel elements.