A Constraint-Aware Multi-Agent Optimization Framework with Robust Domain Specific Language Generation for Travel Planning

Abstract

Today’s LLMs, equipped with advanced reasoning capabilities and the ability to invoke external tools, are giving rise to proactive language agents that can perceive, plan, and act in complex real-world contexts. Travel planning exemplifies this paradigm: it demands not only the construction of a feasible itinerary but also the nuanced alignment with users’ explicit constraints and implicit preferences. However, existing methods are hindered by two key challenges: first, directly capturing constraints from user input is unreliable; and second, the fact that LLMs lack built-in spatial reasoning capabilities, making it difficult for them to identify efficient or shortest paths. To bridge these gaps, this paper proposes a constraint-aware multi-agent optimization framework for travel planning. At the core of our approach is NL2C2DSL (Natural Language to Constraints to Domain-Specific Language), a pipeline that leverages LLMs to extract both explicit and implicit user constraints from natural language queries and translates them into a structured domain-specific language (DSL) through pattern-based mapping. Building on this formalized representation, we design a collaborative multi-agent framework comprising specialized agents for intercity transportation, attraction recommendation, path planning, hotel and restaurant selection, itinerary synthesis, and refinement/validation. Through coordinated interaction among specialized agents, our framework jointly constructs a coherent and comprehensive travel plan. Extensive experiments across different phase dataset validate our framework’s superior travel planning performance. The proposed framework has won the championship on theAutonomous Travel Itinerary Planning Challenge at IJCAI’25 competition. Our codes are available at https://github.com/zhuyiyi-123/CAMTP.

Keywords

travel planning / decision making / constraint optimization / multi-agent / large language model

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.