Source-linked AI summary
AI Tour Meeting: Group Travel Planning by LLM Agents
Daisuke Kikuta
TL;DR
Group travel planning requires balancing multiple participants’ objectives, but evidence on persona-based LLM agents for this setting remains limited. AI Tour Meeting provides a multi-agent discussion framework, and its validation and analyses show that it can support travel-planning discussions and reproduce effects of preference conflict and speaking order.
Problem
Evidence is limited on using persona-based LLM agents to simulate group travel planning discussions that balance participants’ costs, travel times, and preferences.
Method
AI Tour Meeting instantiates distinct-persona LLM agents that collaboratively discuss itineraries under shared goals, constraints, turn rules, and voting rules.
Results
Models performing comparably to or better than Qwen3.5-4B achieve 100% completion, while greater preference conflict lengthens discussions, lowers consensus and satisfaction, and raises victim rates.
Takeaways & Limitations
The framework supports research on LLM-agent behavior and automatic evaluation of recommender systems for group travel planning.
Takeaways & Limitations
Meeting-duration values are reference-only because duration largely depends on the machine used to deploy the LLMs.
Abstract
from arXiv · showhide
This paper proposes AI Tour Meeting, a group travel planning framework powered by multiple Large Language Model (LLM)-based agents. The agents are instantiated with distinct personas and collaboratively seek an itinerary that satisfies their constraints and preferences through natural language discussion. The framework enables easy and flexible orchestration of such discussions by providing interfaces for configuring agent personas, discussion workflows, monitoring, and LLM deployment. Its primary use case is a simulation tool for analyzing the behavior of multiple LLM agents during tour planning discussions. This paper demonstrates the utility of the framework by presenting system validation and several analytical results obtained by the framework.
1 Introduction
AI Tour Meeting is a multi-agent framework for group travel planning in which persona-based LLM agents discuss constraints and preferences to determine an itinerary. The paper validates discussion completion across Qwen 3.5 and GPT models and demonstrates the framework as a simulation tool.
- Motivation: Group travel planning seeks itineraries balancing multiple participants’ objectives, including cost, travel time, and individual preferences.Prior work has supported human groups in determining balanced itineraries.
- Framework: The framework supports LLM agents representing group members who cannot participate in the actual discussion.These agents collaboratively determine an itinerary with the participating group.
- Framework: AI Tour Meeting uses distinct persona-based LLM agents to collaboratively find itineraries satisfying their constraints and preferences through natural-language discussion.The framework is designed for group travel planning with multiple LLM agents.
- Validation: The paper validates whether the framework completes discussions using Qwen 3.5 and GPT models of different sizes.It also analyzes examples to demonstrate the framework’s effectiveness as a simulation tool.
2 Proposed Framework: AI Tour Meeting
AI Tour Meeting is a configurable multi-agent framework in which persona-driven LLM participants collaboratively construct and evaluate itineraries through alternating conversation and voting phases. It supports explicit goals, constraints, workflow rules, action types, and meeting termination conditions.
- Meeting Overview: Users configure global goals, constraints, participants, and workflows before agents alternate conversation and voting until the accepted itinerary satisfies all participants.The meeting may also terminate at a maximum number of turns or execution time, in addition to unanimous satisfaction.
- Itineraries: An itinerary records destinations and inter-destination transportation, including names, descriptions, costs, arrival times, stay durations, modes, and travel durations.The itinerary representation includes both destination-level and transportation-level information.
- Participants: Each LLM participant receives a distinct persona defining identity, background, personality, preferences, meeting role, tone, and explanation style.Explanation styles can be subjective, contrastive, or both, with contrastive explanations comparing routes using cost, time, and destination count.
- Actions: Agents act in turns using intermediate search, ask, and reflect actions, followed by phase-specific terminal actions such as propose, satisfied, accept, reject, and score.Intermediate actions may repeat within a turn, while a terminal action ends the current turn.
- Workflow Rules: The framework offers round robin, inviting, facilitating, random, parallel, balancing, and volunteer turn options, with rules configurable separately for conversation and voting.Balancing ensures one turn per participant per cycle, while volunteer allows participants to pass and skip their turns.
3 System Validation
The framework is validated across five LLMs using synthetic group-travel meetings with partially overlapping preferences and moderate, resolvable conflicts. Results show robust operation for models comparable to or better than Qwen3.5-4B, while weaker models produce more invalid actions and failures.
- Validation setup: 50 synthetic meetings with three participants were run across five LLMs: Qwen3.5-2B/4B/9B, gpt-oss-20b, and gpt-5.4-mini.Preferences partially overlap and contain moderate conflicts resolvable through adjustment; temperatures were 0.6 for local models and 1.0 for gpt-5.4-mini.
- Validation results: Completion reached 100%, while constraint-error and action-failure rates decreased as model performance improved.Completion measures meetings ending with consensus or the maximum turns and at least one accepted itinerary; constraint error measures newly proposed itineraries violating constraints.
- Validation results: Qwen3.5-2B produced invalid actions such as repeating messages and rejecting proposals based on a nonexistent “current route.”The reported invalid actions indicate weaker execution reliability for this model.
- Validation results: The framework runs properly with LLMs performing comparably to or better than Qwen3.5-4B.Meeting-duration values are reference-only because duration largely depends on the machine used to deploy the LLMs.
4 Analysis Examples
The framework is used to analyze how preference conflict and speaking order shape multi-agent travel-planning discussions. Greater conflict produces longer, less consensual discussions, while speaking-order effects emerge in mixed and conflicting settings.
- Preference conflict: The preference-conflict analysis generated 50 synthetic meetings for each of three settings: aligned, mixed, and conflicting.After each meeting, participant satisfaction was evaluated.
- Preference conflict: As preference conflict increased, turns and itinerary proposals increased, while consensus rate and average satisfaction decreased and victim rate increased.Victim rate is the proportion of participants with satisfaction scores of 4 or lower.
- Preference conflict: Representative consensus-building patterns included repeated rejection prompting compromise and mediation by another participant incorporating a rejected proposer’s preferences.In one case, Elena accepted an itinerary with a post-hoc satisfaction score of 3 despite remaining dissatisfied; in another, Leila proposed an itinerary on Noah’s behalf.
- Speaking order: The speaking-order analysis ran each meeting three times with Latin-square seat rotation across positions p1–p3 under a round-robin turn rule.The analysis covered three conflict levels while keeping other settings and constraints unchanged.
- Speaking order: Speaking position had no significant acceptance-rate effect in aligned, whereas second- and third-position proposals were significantly more accepted than first-position proposals in mixed and conflicting.The authors attribute this pattern to absent conflict in aligned and stronger preference conflicts in the other settings.
5 Conclusion
The paper proposes an LLM-based group travel planning framework and demonstrates its validity and utility through system-level evaluation and use cases. The framework supports research on LLM-agent behavior and automatic evaluation of group-travel recommender systems.
- Contributions: The paper proposes a new LLM-based framework for group travel planning.The framework is evaluated at the system level and through use cases.
- Research applications: The framework supports research on the behavior of LLM agents.
- Research applications: The framework enables automatic evaluation of recommender systems for group travel planning.
A License
The code is released under NTT’s proprietary license, restricting its use to research purposes; the authors state this is not a significant limitation because the framework primarily supports research simulation.
- A License: NTT’s proprietary license restricts code use to research purposes, which the authors consider non-limiting because the framework is primarily a research simulation tool.The framework is designed to support research on group travel planning.
B Related Work
Prior work addresses consensus-oriented group travel recommendation, LLM travel-planning benchmarks and improvements, and general LLM-based multi-agent simulation. However, existing studies focus mainly on individual travel or general-purpose orchestration, leaving group-travel discussion insufficiently addressed.
- Group travel planning: Earlier group travel systems supported consensus building through discussion monitoring and preference adaptation, or optimized itineraries using member relationships and POI information.These approaches include chat-based recommendation adaptation and attention mechanisms for integrating group-member and point-of-interest information.
- Travel planning with LLM agents: Recent studies benchmark and improve LLM travel-planning capabilities, but primarily target individual travel planning.The cited work includes travel-planning benchmarks and methods intended to improve LLM performance.
- LLM-based multi-agent simulation: LLM-based multi-agent simulation is widely studied, yet no prior work specifically addresses multi-agent discussion in group travel planning.The paper positions group-travel discussion as an application not covered by the cited multi-agent simulation literature.
- LLM-based multi-agent simulation: General-purpose multi-agent frameworks provide basic interaction orchestration, but users must implement domain-specific functionality themselves.The supplied passage identifies an implementation gap between generic orchestration tools and domain-specific group travel planning.
C Human Participant (Interactive Mode)
AI Tour Meeting supports human participation in agent meetings through chat-based actions matching those available to LLM participants. Users can manually edit and have an LLM refine itinerary proposals, including to incorporate perspectives of absent friends represented by persona-based agents.
- Human participation: Human users participate through chat-based actions that mirror those available to LLM participants.They interact during their assigned turns using the actions defined in Eq. (2).
- Chat interactions: Users can vote, ask participants questions, propose itineraries, and manually modify destinations through the chat interface.Figure 6 presents chat-box snapshots for these interaction modes.
- Proposal refinement: The “Generate with AI” button lets users interact with an LLM to refine their itinerary proposals.This provides AI-assisted refinement after users propose and edit destinations.
- Absent participant representation: Users can represent absent friends as persona-based LLM participants and incorporate their perspectives through itinerary discussions.The agents stand in for friends unable to attend the meeting.
D Additional Validation: Participant Size
The framework was additionally validated on synthetic meetings with 5 and 10 participants using GPT-5.4-mini generation and Qwen3.5-9B evaluation. It maintained complete meetings as group size increased, while constraint errors and discussion turns rose modestly.
- Setup: The validation generated 50 synthetic meetings for each size—5 and 10 participants—with GPT-5.4-mini and evaluated them using Qwen3.5-9B.The evaluation used the same settings as the system validation.
- Results: 100% completion rate was maintained as the number of participants increased to 5 and 10.The result indicates that meetings continued to complete successfully at both larger group sizes.
- Results: Constraint error rate increased slightly with group size but remained sufficiently low, partly because itineraries tended to include more destinations.The reported increase was associated with the greater number of destinations in larger-group itineraries.
- Results: The number of discussion turns grew slightly faster than the number of participants.This describes the scaling trend observed as participant size increased.
E Context Management
The framework builds each participant’s input context from persona and meeting settings, phase markers, action logs, and current-turn instructions, with retry handling and configurable compaction methods for managing context length.
- Input context: Each participant’s context begins with a system prompt for persona and meeting settings, then alternates phase markers and action logs as memory.During a turn, current actions and phase instructions are appended before the model generates the next action; completed-turn actions are then recorded.
- Input context: Phase markers separate conversation and voting, while action logs record all participants’ actions in speaking order and assign roles by action source.The system prompt uses the system role; phase markers and instructions use the user role; the participant’s actions use the assistant role, and others’ actions use the user role.
- Retry loop: Invalid JSON or itinerary time/cost violations append an error message to the context and trigger regeneration, with a maximum of three retries.The retry loop applies when an action output is malformed or violates a proposed itinerary constraint.
- Context compaction: The framework supports summarization, turn-window truncation, and token-based truncation to compact participant context.Summarization inserts a specified summary after the system prompt, while the other methods retain only recent turns or tokens and remove earlier context.
F Prompts
The framework uses customizable, persona-encoded system prompts and phase-specific instructions to coordinate discussion, itinerary proposals, voting, questions, speaker invitations, and post-meeting satisfaction evaluation. Runtime placeholders and structured output schemas support flexible prompt instantiation and consistent agent responses.
- Prompt organization: The main prompts cover persona encoding, conversation and voting phases, participant invitations, and post-meeting satisfaction evaluation.Users can customize the system prompt, while curly-brace placeholders such as {name} and {format_instructions} are filled at runtime.
- Persona prompt: The system prompt instructs each participant to role-play a specified persona and help the group reach a shared itinerary satisfying everyone’s interests.Persona fields include the participant’s name and meeting role, such as facilitator or attendee.
- Conversation phase: Conversation-phase instructions let agents ask participants, reflect on discussion, or propose routes using a structured output schema.Proposal instructions require an ordered destination list, conversational explanation, and arrival-time consistency with stay durations and travel times.
- Voting and questions: Voting-phase and question-answering prompts reuse structured actions while directing agents to evaluate a proposer’s route or respond to another participant’s question.The voting prompt identifies the proposer through {proposer_name}, and the answering prompt identifies the questioner through {asker_name}.
- Turn management and evaluation: Invitation and post-meeting prompts select the next speaker strategically and ask each persona to assess how well the final route satisfies personal goals.The invitation rule requires choosing exactly one available participant and setting that name in next_speaker.