How a Constraint-Aware Multi-Agent AI Won the IJCAI‑2025 Travel Planning Challenge

Introduction

Alipay’s Industry Technology Department, Ant Group’s Basic Intelligent Technology team, and the research groups of Prof. Qian Hong and Prof. Li Bingdong from East China Normal University participated in the 34th International Joint Conference on Artificial Intelligence (IJCAI‑2025) Autonomous Travel Itinerary Planning Challenge. Using a self‑developed “large model + optimization solving” technology, they fulfilled user travel‑planning requests and secured the Original OS Track (natural language) championship and the DSL Track (constrained) runner‑up.

Background

The IJCAI‑2025 challenge aims to advance large‑model‑driven agents that can generate feasible, personalized travel plans from specific user requests such as “I want a 3‑day trip from Shanghai to Beijing, visit the Forbidden City, and stay within a budget of 5,000 CNY.”

Typical large‑model‑generated itineraries suffer from spatio‑temporal conflicts, outdated or inaccurate information, and shallow personalization because the problem is treated as pure text generation rather than multi‑constraint dynamic resource scheduling.

Solution

For the DSL track, which requires a verifiable domain‑specific language (DSL), we built a constraint‑aware multi‑agent framework that integrates a retrieval‑augmented fine‑tuned LLM to more accurately extract user constraints and intents. For the OS track, we extended this framework with a powerful DSL‑generation capability that automatically produces verifiable domain language for downstream validation. The framework consists of three core modules: Environment, Thinking, and Action.

1. Environment Module

The environment module serves as the interface between user requests and the large language model. By applying large‑scale supervised fine‑tuning (SFT) together with retrieval augmentation, it parses ambiguous expressions and dialect variations, aligns them with task requirements, and constructs verifiable constraints and domain language (implemented in Python for final constraint verification).

2. Thinking Module

The thinking module uses the fine‑tuned LLM to analyze the input, systematically infer task‑specific constraints, and ensure that implicit needs and personalized preferences are explicitly captured and remain consistent with the problem context.

3. Action Module

The action module orchestrates multiple specialized agents:

Inter‑city transportation agent – selects travel modes between origin and destination while balancing time, budget, and user preferences.

Attraction recommendation agent – aligns user preferences with geographic logic to produce feasible point‑of‑interest suggestions.

Route planning agent – solves a heuristic optimization problem to schedule daily itineraries, selecting POIs and minimizing total travel time.

Dining and hotel recommendation agent – inserts suitable restaurants and hotels based on preferences and budget.

Itinerary integration agent – stitches all POIs together, plans local paths between adjacent POIs, and outputs the final plan.

Correction & reflection controller – after each modification, validates budget, preferences, and other constraints using the Python‑based domain language.

These modules dramatically improve multi‑agent performance under vague constraints and provide theoretical insights and technical support for DSL‑based travel‑planning systems in broader domains.

Future Outlook

We will broaden the framework’s applicability to more complex constraint scenarios, continuously optimize module performance, and enhance system responsiveness and plan quality, thereby strengthening intelligent travel services for Alipay.

Alipay’s AI travel assistant is already live. As part of the national “Transportation Power” pilot, the team will contribute to high‑quality cultural‑tourism datasets, expert model training, and industry standard development to advance AI applications in travel and improve user experience.

Conference and Competition Websites

[1] IJCAI‑2025 official site: https://2025.ijcai.org/

[2] Autonomous Travel Itinerary Planning Challenge: https://chinatravel-competition.github.io/IJCAI2025/