AI Tour Meeting: Group Travel Planning by LLM Agents
This paper introduces "AI Tour Meeting," a framework that utilizes multiple Large Language Model agents with distinct personas to collaboratively plan group travel itineraries through natural language discussion, serving as a flexible tool for simulating and analyzing multi-agent behaviors.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers don't just answer questions but actually have personalities, opinions, and the ability to hold a conversation. This is the realm of Large Language Models (LLMs). Think of them as incredibly well-read digital librarians who have read almost everything on the internet. Usually, we talk to them one-on-one, like asking a librarian for a book recommendation. But what happens when you ask a whole group of these digital librarians to work together? That's where Multi-Agent Systems come in. It's like setting up a roundtable where every computer has a specific role and a unique personality, tasked with solving a problem together. While we often use these tools to help real humans plan trips, this paper asks a fascinating question: What if the computers are the travelers? Could a group of AI agents, each with their own distinct likes and dislikes, actually sit down, argue, compromise, and agree on a perfect vacation plan without a single human needing to type a word?
The paper "AI Tour Meeting" by Daisuke Kikuta introduces a new framework designed to simulate exactly this scenario. It's a digital playground where multiple AI agents, each dressed in a unique "persona" (like a history buff, a budget-conscious student, or a foodie), gather to plan a group trip. The system doesn't just let them chat; it forces them into a structured meeting with rules, voting, and specific goals. The researchers found that when these AI agents are given the right instructions and a strong enough "brain" (a powerful LLM), they can successfully navigate complex disagreements, propose new ideas, and reach a consensus on an itinerary, just like a group of friends planning a weekend getaway.
The Digital Travel Club
Imagine you are organizing a trip for a group of friends, but instead of real people, you have a group of AI characters. One is "Alice," who loves history and hates spending money. Another is "Bob," who wants to see everything in one day and doesn't care about the cost. The paper's framework, AI Tour Meeting, is the software that runs this meeting. It sets the stage: "Okay, everyone, we have 24 hours and a budget of $100. Let's plan a day in Tokyo."
The magic happens in how these AI agents interact. They don't just spit out a list of places; they hold a conversation. They take turns speaking, asking questions, and making proposals. If Alice suggests visiting a museum, Bob might say, "That's cool, but it's too far from the food market I want to visit." The system then moves into a voting phase. Everyone votes on whether to accept the new plan or stick with the old one. If the plan gets enough votes, it becomes the new "current itinerary." This cycle of chatting and voting continues until everyone is happy (or at least, everyone agrees to stop arguing).
The Rules of the Game
To make sure the AI doesn't just ramble on forever, the framework gives them a strict rulebook.
- Personas: Each agent is given a detailed backstory. They have a name, a personality (like "cautious" or "sociable"), and specific goals (like "I must visit the temple").
- Turns: The meeting follows a schedule. It could be a round-robin (everyone speaks in order), or one person can invite the next speaker.
- Actions: During their turn, an agent can do a few things:
- Search: Look up facts (like "How much is the train ticket?").
- Ask: Ask another agent a question.
- Propose: Suggest a new travel route.
- Vote: In the voting phase, they can Accept or Reject a proposal.
- Satisfied: If they like the current plan, they can say, "I'm good with this," which helps end the meeting.
The system is smart enough to check if a plan breaks the rules. If an agent suggests a trip that costs $500 when the budget is $100, the system catches that error. It also tracks how many times the group has to argue before they agree.
What the Computers Actually Did
The researchers tested this system to see if it actually works. They set up meetings with different types of AI models (some smaller, some bigger) and watched what happened.
The Results:
- It Works (Mostly): When they used powerful AI models (like Qwen3.5-4B or larger), the meetings almost always finished successfully. The agents managed to agree on a plan 100% of the time in many tests.
- The "Brain" Matters: Smaller, weaker AI models (like the 2B version) struggled. They sometimes got confused, repeated the same sentence, or made up facts. The paper suggests you need a fairly smart AI (at least as good as Qwen3.5-4B) to run this without glitches.
- Conflict Takes Time: The researchers created scenarios where the agents had conflicting goals. When everyone agreed on what they wanted, the meeting was quick (about 10 turns). But when they had big disagreements, the meetings dragged on (up to 26 turns on average), and the agents had to propose many different routes before finding one that worked.
- The "Victim" Effect: In the most difficult scenarios, where the agents had very different tastes, some agents ended up unhappy. About 11% of the time, an agent was so unhappy with the final plan that they gave it a low score (4 or lower out of 10). This mirrors real life: sometimes, to get a group to agree, someone has to compromise and give up what they really wanted.
Why This Matters
This isn't just a fun game of digital tourists. The paper suggests this framework is a powerful tool for research. By simulating these meetings, scientists can study how groups make decisions without needing to hire hundreds of real people. They can test how different rules (like voting methods or speaking orders) affect the outcome.
For example, they found that in a group with conflicting opinions, the person who speaks first often has a harder time getting their plan accepted. The people who speak later have the advantage of hearing everyone else's complaints and can tweak their proposal to make everyone happier. This is a discovery about group dynamics that the AI helped uncover.
The paper also hints at a future where you could use this technology to plan trips for real groups. Imagine you are planning a vacation with friends who live in different time zones or can't make the call. You could create AI agents that represent your friends' personalities and preferences. These agents could chat with your own agent to figure out a plan that everyone would like, so when you finally talk to your friends, you already have a perfect itinerary ready to go.
In short, AI Tour Meeting shows that if you give computers the right personalities and rules, they can learn to be good teammates, navigating the messy, complicated world of group decision-making just like we do. It's a step toward a future where AI doesn't just do our work for us, but helps us understand how we work together.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.