AI Tour Meeting: Group Travel Planning by LLM Agents

# AI Tour Meeting: Group Travel Planning by LLM Agents
Source: [https://arxiv.org/html/2607.18806](https://arxiv.org/html/2607.18806)
###### Abstract

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\.

![[Uncaptioned image]](https://arxiv.org/html/2607.18806v1/images/ai-tour-meeting-logo2.png)

AI Tour Meeting: Group Travel Planning by LLM Agents

Daisuke KikutaNTT, Inc\.daisuke\.kikuta@ntt\.com

## 1Introduction

Group travel planning aims to find an itinerary that balances multiple objectives among multiple participants, such as cost, travel time, and individual preferences\. Prior work has focused on supporting human groups in determining balanced itinerariesNguyen and Ricci \([2018](https://arxiv.org/html/2607.18806#bib.bib3)\); Chenet al\.\([2021](https://arxiv.org/html/2607.18806#bib.bib2)\); Chenget al\.\([2026](https://arxiv.org/html/2607.18806#bib.bib5)\)\. Meanwhile, in broader domains, recent studies have explored the use of persona\-based Large Language Model \(LLM\) agents to simulate human behaviorParket al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib8)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib6)\); Jianget al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib7)\); Hu and Collier \([2024](https://arxiv.org/html/2607.18806#bib.bib9)\)\. This emerging direction is also expected to be relevant to group travel planning\. For example, automatic evaluation in such simulation environments reduces the costs of large\-scale behavioral analysis in group travel planning and the evaluation of existing group travel recommender systems\. In application, users could interact with LLM agents representing group members who are unable to participate in the actual discussion and collaboratively determine an itinerary with them\.

To address this emerging need, we propose AI Tour Meeting, a group travel planning framework powered by multiple LLM agents\. Each agent is instantiated with a distinct persona and collaborates with the others to find an itinerary that satisfies their constraints and preferences through natural\-language discussion\. Our framework provides interfaces to orchestrate this process\. Compared with existing persona\-based agent simulationsParket al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib8)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib6)\)and general\-purpose multi\-agent frameworksLiet al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib11)\); Wuet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib10)\), our framework provides domain\-specific abstractions for group travel planning, including a structured itinerary representation, explicit proposal\-voting workflows, and constraint validation and metric monitoring specialized for travel planning\. This allows users to focus on their use cases without implementing the required functionality from scratch using a general\-purpose framework\.

In the following, we describe the technical details of our framework and validate whether it can successfully complete discussions using Qwen 3\.5 and GPT models of different sizes\. We then demonstrate its effectiveness as a simulation tool through analysis examples conducted using the framework\.

![Refer to caption](https://arxiv.org/html/2607.18806v1/x1.png)Figure 1:An overview of AI Tour Meeting\. The user first configures the meeting settings, including the global goals, constraints, participants, and workflow, and then starts the meeting\. The meeting alternates between conversation and voting phases whenever a new itinerary is proposed\. Once all participants are satisfied with the currently accepted itinerary, the meeting concludes\.
## 2Proposed Framework: AI Tour Meeting

Figure[1](https://arxiv.org/html/2607.18806#S1.F1)shows the overview of our framework: we first set up both LLM participants with distinct personas and a discussion workflow, then start the meeting among them aligned with the specified workflow\. The meeting begins with the conversation phase, in which participants take turns searching information, asking questions, and proposing itineraries\. Once an itinerary is proposed, it transitions to the voting phase, where the participants decide whether to accept or reject it\. This two\-phase process continues until all participants are satisfied with the currently accepted itinerary\.

In this section, we describe the definitions of itineraries and participants, the workflow options, the monitored metrics, and the user interface\.

### 2\.1Itinerary

An itinerary proposed during the meeting consists of information about destinations and the transportation between them, and is defined as follows:

R=\(d1,d2,…,dN\),R=\(d\_\{1\},d\_\{2\},\\dots,d\_\{N\}\),\(1\)
whereRRis the itinerary,did\_\{i\}is theii\-th destination\. Each destination includes a name, description, monetary cost, arrival time, duration of stay, transportation mode from the previous destination, transportation cost, and transportation duration\.

### 2\.2Participant

Each LLM agent \(participant\) is instantiated with a distinct persona\. The persona consists of the following attributes:Namespecifies their name;Backgroundspecifies their experience or situation;Personalityspecifies their stable traits;Preferencespecifies their preferences related to the tour;Goalspecifies their personal goals in the tour;Roleselects their role in the meeting from either facilitator or attendee;Tonespecifies their speaking tone such as formal or friendly;Explanation styleselects a strategy for convincing others to accept their proposal, from either subjective explanation, contrastive explanationKikutaet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib1)\)or both\. The former provides subjective opinions of the proposal based on their preferences\. The latter explains the objective pros and cons by comparing the proposed route with the current route in terms of objective metrics such as cost, time, and number of destinations\. We can select an LLM that instantiates the participant with model parameters such as temperature and seed\. Both commercial LLMs and local LLMs \(via Ollama or vLLM\) are supported, Context management methods, including auto\-compaction and history truncation, can be also configured individually for each participant\.

##### Actions

Thepersona\\mathrm\{persona\}is given to the LLM as a system prompt, and the LLM participant takes actions during the meeting while role\-playing according to the givenpersona\\mathrm\{persona\}\. The meeting proceeds in a turn\-based manner, and each participant may take multiple actions during their single turn\. The action at each stepsswithin a turn is defined as follows:

\(astype,msgs,as\)=pm​\(personam,Hs−1\),\(a^\{\\mathrm\{type\}\}\_\{s\},\\mathrm\{msg\}\_\{s\},a\_\{s\}\)=p\_\{m\}\(\\mathrm\{persona\}\_\{m\},H\_\{s\-1\}\),\(2\)wherepmp\_\{m\}is themm\-th LLM participant,astypea\_\{s\}^\{\\mathrm\{type\}\}is the action type,mgss\\mathrm\{mgs\}\_\{s\}is the message associated with the action,asa\_\{s\}is the taken action, andHs−1H\_\{s\-1\}is the meeting history, which includes system messages for managing the discussion workflowmsgsys\\mathrm\{msg\}\_\{\\mathrm\{sys\}\}and the past actions taken by all the participants\(\(a1type,msg1,a1\),…,\(as−1type,msgs−1,as−1\)\)\(\(a\_\{1\}^\{\\mathrm\{type\}\},\\mathrm\{msg\}\_\{1\},a\_\{1\}\),\\dots,\(a\_\{s\-1\}^\{\\mathrm\{type\}\},\\mathrm\{msg\}\_\{s\-1\},a\_\{s\-1\}\)\)\. The actionasa\_\{s\}varies depending on the selected action typeastypea^\{\\mathrm\{type\}\}\_\{s\}, as follows:

as=\{Search​\(qs\)if​astype=searchAsk​\(pm′,qs\)if​astype=askReflectif​astype=reflectPropose​\(Rs\)if​astype=proposeϕotherwiseScore​\(vs\)if​astype=scoringϕotherwise\.a\_\{s\}=\\left\\\{\\begin\{array\}\[\]\{ll\}\\mathrm\{Search\}\(q\_\{s\}\)&\\text\{if \}a^\{\\mathrm\{type\}\}\_\{s\}=\\mathrm\{search\}\\\\ \\mathrm\{Ask\}\(p\_\{m^\{\\prime\}\},q\_\{s\}\)&\\text\{if \}a^\{\\mathrm\{type\}\}\_\{s\}=\\mathrm\{ask\}\\\\ \\mathrm\{Reflect\}&\\text\{if \}a^\{\\mathrm\{type\}\}\_\{s\}=\\mathrm\{reflect\}\\\\\[4\.0pt\] \\mathrm\{Propose\}\(R\_\{s\}\)&\\text\{if \}a^\{\\mathrm\{type\}\}\_\{s\}=\\mathrm\{propose\}\\\\ \\phi&\\text\{otherwise\}\\\\\[4\.0pt\] \\mathrm\{Score\}\(v\_\{s\}\)&\\text\{if \}a^\{\\mathrm\{type\}\}\_\{s\}=\\mathrm\{scoring\}\\\\ \\phi&\\text\{otherwise\.\}\\end\{array\}\\right\.\(3\)
There are two types of actions: intermediate and terminal actions\. Intermediate actions may be taken multiple times within a turn, whereas once a terminal action is taken, the current turn ends\. Theintermediateaction types includesearch\\mathrm\{search\},ask\\mathrm\{ask\}, andreflect\\mathrm\{reflect\}\.search\\mathrm\{search\}issues a queryqsq\_\{s\}to retrieve information such as location or cost of destinations via web search\.ask\\mathrm\{ask\}asks another participantpm′p\_\{m^\{\\prime\}\}a questionqsq\_\{s\}and receives a response\.reflect\\mathrm\{reflect\}organizes and reviews the discussion conducted so far\. Theterminalterminalaction types differ between the conversation phase and the voting phase\. In theconversationphase, they includepropose\\mathrm\{propose\},satisfied\\mathrm\{satisfied\}, andpass\\mathrm\{pass\}\.propose\\mathrm\{propose\}proposes a new itineraryRsR\_\{s\}\.satisfied\\mathrm\{satisfied\}indicates thatpmp\_\{m\}is satisfied with the current route\.pass\\mathrm\{pass\}ends their turn without expressing a specific position\. In thevotingphase, they includeaccept\\mathrm\{accept\},reject\\mathrm\{reject\}, andscore\\mathrm\{score\}\.accept\\mathrm\{accept\}votes for the proposed itinerary,reject\\mathrm\{reject\}does not votes for the proposed itinerary, andscore\\mathrm\{score\}assigns a scorevsv\_\{s\}to the proposed route\.

### 2\.3Discussion workflow

A tour meeting consists of a global goal, participants, constraints, a turn rule, and a voting rule\. The global goal is a high\-level objective shared among all participants, e\.g\., planning a one\-day sightseeing tour in Tokyo\. Each participant advances the discussion in pursuit of their own personal goal while remaining aligned with the global goal\. Constraints restrict the itinerary in terms of travel date, budget, and time window\.

The tour meeting proceeds according to the specified turn rule, under which participants sequentially take the actions defined in Eq\. \(4\) multiple times within their turns\. This normal state is referred to as the conversation phase\. Once a participant proposes a itinerary \(i\.e\., terminates the turn withatype=proposea^\{\\mathrm\{type\}\}=\\mathrm\{propose\}\), the process transitions to the voting phase\. In this phase, all participants except the proposer cast votes on the proposed itinerary\. After collecting all votes, the acceptance or rejection of the proposed route is determined in accordance with the specified voting rule\. If the itinerary is accepted, the current itinerary is replaced with the proposed one; if it is rejected, the current itinerary remains unchanged\. The process then returns to the conversation phase, and the turns resume with the next participant after the proposer\.

When all participants consecutively indicate satisfaction with the current route \(i\.e\., terminate their turns withatype=satisfieda^\{\\mathrm\{type\}\}=\\mathrm\{satisfied\}\), consensus is considered reached, and the meeting concludes\. In practice, the termination condition of the meeting may also be constrained by a maximum number of turns or a maximum execution time, in addition to the unanimous agreement\.

Algorithm[1](https://arxiv.org/html/2607.18806#alg1)organizes the discussion workflow described above\. In the following, we describe turn rules and voting rules available in our framework\.

Input :Participants

𝒫=\{p1,p2,…,pM\}\\mathcal\{P\}=\\\{p\_\{1\},p\_\{2\},\\dots,p\_\{M\}\\\};

Turn rule

πturn\\pi\_\{\\mathrm\{turn\}\}; Voting rule

πvote\\pi\_\{\\mathrm\{vote\}\}
Output :The final tour itinerary

RfinalR\_\{\\mathrm\{final\}\}
1

2

Rfinal←ϕR\_\{\\mathrm\{final\}\}\\leftarrow\\phi;

H←ϕH\\leftarrow\\phi;

C←0C\\leftarrow 0
3

4while*C<MC<M*do

5

p←πturn​\(𝒫\)p\\leftarrow\\pi\_\{\\mathrm\{turn\}\}\(\\mathcal\{P\}\)
6while*atype∈\{search,ask,reflect\}a^\{\\mathrm\{type\}\}\\in\\\{\\mathrm\{search\},\\mathrm\{ask\},\\mathrm\{reflect\}\\\}*do

7

\(atype,msg,a\)←p​\(persona,H\)\(a^\{\\mathrm\{type\}\},\\mathrm\{msg\},a\)\\leftarrow p\(\\mathrm\{persona\},H\)
8

H←H\+\(atype,msg,a\)H\\leftarrow H\+\(a^\{\\mathrm\{type\}\},\\mathrm\{msg\},a\)
9if*atype=proposea^\{\\mathrm\{type\}\}=\\mathrm\{propose\}*then

10

R′←aR^\{\\prime\}\\leftarrow a
11end if

12

13end while

14if*atype=proposea^\{\\mathrm\{type\}\}=\\mathrm\{propose\}*then

15

𝒱←ϕ\\mathcal\{V\}\\leftarrow\\phi;

C←0C\\leftarrow 0
16for*\_←1\\\_\\leftarrow 1toM−1M\-1*do

17

p′←πturn​\(𝒫∖p\)p^\{\\prime\}\\leftarrow\\pi\_\{\\mathrm\{turn\}\}\(\\mathcal\{P\}\\setminus p\)
18while*atype∈\{search,ask,refect\}a^\{\\mathrm\{type\}\}\\in\\\{\\mathrm\{search\},\\mathrm\{ask\},\\mathrm\{refect\}\\\}*do

19

\(atype,msg,a\)←p′​\(persona,H\)\(a^\{\\mathrm\{type\}\},\\mathrm\{msg\},a\)\\leftarrow p^\{\\prime\}\(\\mathrm\{persona\},H\)
20

H←H\+\(atype,msg,a\)H\\leftarrow H\+\(a^\{\\mathrm\{type\}\},\\mathrm\{msg\},a\)
21

22end while

23

24if*atype∈\{accept,reject\}a^\{\\mathrm\{type\}\}\\in\\\{\\mathrm\{accept\},\\mathrm\{reject\}\\\}*then

25

𝒱←𝒱\+atype\\mathcal\{V\}\\leftarrow\\mathcal\{V\}\+a^\{\\mathrm\{type\}\}
26end if

27

28end for

29

30if*πvote​\(𝒱\)=accept\\pi\_\{\\mathrm\{vote\}\}\(\\mathcal\{V\}\)=\\mathrm\{accept\}*then

31

Rfinal←R′R\_\{\\mathrm\{final\}\}\\leftarrow R^\{\\prime\}
32

33end if

34

35else if*atype=satisfieda^\{\\mathrm\{type\}\}=\\mathrm\{satisfied\}*then

36

C←C\+1C\\leftarrow C\+1
37else

38

C←0C\\leftarrow 0
39end if

40

41end while

return

RfinalR\_\{\\mathrm\{final\}\}

Algorithm 1Tour meeting workflow![Refer to caption](https://arxiv.org/html/2607.18806v1/images/tourmeeting_overview2.png)Figure 2:The graphical user interface of AI Tour Meeting\. From left to right: the participant settings, the meeting settings, the ongoing meeting screen, and the analytics dashboard\.#### 2\.3\.1Turn rule

Our framework supports standard four turn rules, along with two additional options to make them more flexible\. There are five types of turn rules:Round robinrotates participants in a fixed cyclic order;Invitinglets the current speaker determine the next speaker;Facilitatingassigns an additional turn to a facilitator after each turn, who then determines the next speaker;Randomrotates participants randomly;Parallellets all participants vote simultaneously \(only for the voting phase\)\. Note that different turn rules may be applied separately to the conversation and voting phases\.

In addition, when theBalancingoption is enabled, speaking turns are organized into cycles such that all participants take turns once per cycle to ensure fairness\. When theVolunteeroption is enabled, participants are allowed to select thepass\\mathrm\{pass\}as a terminal action and skip their turn \(i\.e\., only those who wish to speak take a turn\)\.

#### 2\.3\.2Voting rule

Our framework also supports standard five voting rules\. InMajorityandUnanimous, participants cast a binary vote \(accept or reject\) on the proposed itinerary\. The proposed itinerary replaces the current itinerary if it receives a majority of accept votes in majority voting, or if all participants vote accept in unanimous voting\. InSingle decider, a specified participant alone cast the binary vote, and the current itinerary is updated accordingly\. InMost pleasureandLeast misery, participants assign a score to the proposed itinerary according to 10\-point rubrics\. The proposed itinerary replaces the current itinerary if, in the former, the sum of all participants’ scores is greater than or equal to that of the current itinerary, or, in the latter, the minimum score across participants is greater than or equal to that of the current itinerary\.

fromtour\_meeting\.cliimportbuild\_meeting

meeting=build\_meeting\(

title="One\-DayTokyoTour",

global\_goals="Planafunone\-dayTokyotour\.",

participants=\[

\{

"name":"Alice",

"background":"Ahistoryenthusiastwho\.\.\.",

"goals":"VisitSenso\-jiand\.\.\.",

"model\_name":"vllm/0/Qwen/Qwen3\.5\-9B",

\.\.\.

\},

\.\.\.

\],

constraints=\{"budget":"¥10K","start":\.\.\.\},

settings=\{"max\_turns":100,"turn\_rule":\.\.\.\},

\)

asyncio\.run\(meeting\.run\_cli\(\)\)

Figure 3:Minimal example of Python APIModelCompletion\(%\)↑\\uparrowConsensus\(%\)Constraint err\.\(%\)↓\\downarrowAction fail\.\(%\)↓\\downarrowTurnsDuration111Since meeting duration largely depends on the machine used to deploy the LLMs, these values are for reference only\.\(min\)Tokens \[K\]\(in / out\)UsableQwen3\.5\-2B8282585814\.214\.21\.31\.361\.561\.572\.572\.571007100/114114✗Qwen3\.5\-4B1001001001003\.13\.10\.50\.515\.715\.737\.037\.0240240/4444✓Qwen3\.5\-9B1001001001000\.00\.00\.10\.112\.812\.833\.633\.6134134/3737✓gpt\-oss\-20b1001001001000\.40\.40\.20\.220\.420\.411\.811\.8315315/2525✓gpt\-5\.4\-mini10010080800\.00\.00\.10\.134\.534\.56\.36\.3535535/4040✓Table 1:System validation results across different models and model sizes\. These results suggest that our framework requires an LLM with performance comparable to or higher than Qwen3\.5\-4B\.

### 2\.4Monitoring and metrics

Our framework collects various metrics to analyze both system\-level performance and the dynamics of discussions\. The system\-level metrics include agents’ token usage, the number of retries, and the number of time violations at destinations in the proposed itineraries\. The discussion\-dynamics metrics include the number of meeting steps and turns, the message history, the distribution of executed action types, the distribution of votes and scores, and the details of the proposed itineraries\. Users can leverage these metrics to analyze discussions flexibly from multiple perspectives\. Basic analyses, such as computing representative statistics and distributions for each metric, can be performed using built\-in functions\.

### 2\.5User interface

Figure[2](https://arxiv.org/html/2607.18806#S2.F2)shows the snapshots of meeting settings and meeting dialogue, and analytics dashboard in the GUI of AI Tour Meeting\. Participants and meetings can be easily configured through pull\-down selections and text inputs, and the meeting can be browsed in a chatbot\-style dialogue view\. The analytics dashboard allows users to visually explore some basic analytical metrics\. Our framework also supports Python API\. As shown in Figure[3](https://arxiv.org/html/2607.18806#S2.F3), users can flexibly configure and launch meetings at scale with only a few lines of code\.

## 3System Validation

First, we validate whether our framework runs robustly across different models and model sizes\.

##### Setup

We use gpt\-5\.4\-miniOpenAI \([2026](https://arxiv.org/html/2607.18806#bib.bib15)\)to generate 50 synthetic meetings with three participants whose preferences partially overlap and contain moderate conflicts that can be resolved through adjustment\. We then run these meetings using five different LLMs to instantiate the participants: Qwen3\.5\-2B/4B/9BQwen Team \([2026](https://arxiv.org/html/2607.18806#bib.bib13)\), gpt\-oss\-20bOpenAI \([2025](https://arxiv.org/html/2607.18806#bib.bib14)\), gpt\-5\.4\-mini \(medium effort\)\. The temperature is set to 0\.6 for the local models and 1\.0 for gpt\-5\.4\-mini\. Local models are served with vLLMKwonet al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib12)\)on a single RTX A6000 GPU \(48GB\)\. All meetings are configured with round\-robin turns, majority voting, a 100\-turn limit, a time window of 09:00–18:00, and a budget of $100 per participant\.

##### Results

Table[1](https://arxiv.org/html/2607.18806#S2.T1)shows the validation results\. We report the following metrics:completiondenotes the proportion of meetings that ended upon reaching consensus or the maximum number of turns, with at least one itinerary accepted;consensusdenotes the proportion of meetings that ended upon reaching consensus;constraint errordenotes the proportion of newly proposed itineraries that violated the constraints;action failuredenotes the proportion of actions that failed after more than three retries; andturns,duration, andtokensdenote the corresponding average values\. The results show that all models except for Qwen3\.5\-2B achieve a 100% completion rate and the rates of constraint errors and action failures decrease as model performance improves\. Qwen3\.5\-2B produces several invalid actions, including repeating the same message and rejecting proposals by referring to a nonexistent “current route\.” Overall, our framework runs properly with LLMs that perform comparably to or better than Qwen3\.5\-4B\.

AlignedMixedConflictingTurns10\.3±7\.6010\.3\_\{\\pm 7\.60\}12\.1±9\.8012\.1\_\{\\pm 9\.80\}25\.9±22\.725\.9\_\{\\pm 22\.7\}Proposals2\.3±2\.302\.3\_\{\\pm 2\.30\}2\.7±2\.502\.7\_\{\\pm 2\.50\}6\.5±5\.306\.5\_\{\\pm 5\.30\}Consensus rate100%100\\%100%100\\%94%94\\%Satisfaction8\.93±1\.478\.93\_\{\\pm 1\.47\}8\.39±1\.708\.39\_\{\\pm 1\.70\}7\.13±2\.067\.13\_\{\\pm 2\.06\}Victim rate0\.7%0\.7\\%1\.3%1\.3\\%11\.3%11\.3\\%Table 2:Meeting results under controlled preference conflict among LLM participants\.![Refer to caption](https://arxiv.org/html/2607.18806v1/x2.png)\(a\)A pattern of prompting compromise\.
![Refer to caption](https://arxiv.org/html/2607.18806v1/x3.png)\(b\)A pattern of mediation by another participant\.

Figure 4:Examples of consensus\-building patterns\.

## 4Analysis Examples

In this section, we demonstrate the utility of our framework by providing two analysis examples conducted using our framework\.

### 4\.1Preference conflict

##### Setup

This example analyzes how meeting dynamics change depending on the degree of preference conflict among participants\. We generate 50 synthetic meetings for each of three settings:alignedassigns aligned preferences among the three participants;mixedassigns mixed preferences as in the system validation;conflictingassigns conflicting preferences that require at least one participant to compromise on at least one aspect\. After each meeting, we also evaluate each participant’s satisfaction with the final itinerary on a 1–10 scale using an LLM\-as\-a\-judge\. All meetings use the inviting turn rule and Qwen\-3\.5\-9B as participants, with all other settings and constraints unchanged from the system validation\.

##### Results

Table[2](https://arxiv.org/html/2607.18806#S3.T2)shows that as the degree of preference conflict increases, the number of turns and itinerary proposals both increase, resulting in longer discussions\. Meanwhile, as preference conflict increases, the consensus rate and average satisfaction score decrease, while thevictim rate—the proportion of participants with a satisfaction score of 4 or lower—increases\. These results suggest that discussions among LLM participants reproduce a natural phenomenon observed in human deliberation: greater conflict leads to longer discussions and makes it harder to find a proposal that satisfies everyone\.

Figure[4](https://arxiv.org/html/2607.18806#S3.F4)shows two representative patterns of the consensus\-building process under preference conflict\. Figure[4\(a\)](https://arxiv.org/html/2607.18806#S3.F4.sf1)shows a case in which Elena’s proposals are repeatedly rejected by the other two participants\. Eventually, Jisoo urges Elena to conclude the meeting, and Elena compromises to end the discussion\. Elena assigns the final itinerary a score of 3 in the post\-hoc evaluation, indicating that she agreed despite remaining dissatisfied\. In contrast, Figure[4\(b\)](https://arxiv.org/html/2607.18806#S3.F4.sf2)shows a case in which, after Noah’s proposal was rejected twice, Leila intervened and proposed an itinerary on his behalf that incorporated his preferences\.

### 4\.2Speaking order

![Refer to caption](https://arxiv.org/html/2607.18806v1/x4.png)Figure 5:Proposal acceptance rates by preference conflict level and speaking position\. Error bars are Wilson 95% confidence intervals on the total rate\.##### Setup

This example analyzes how speaking order affects the proposal acceptance rate\. We run each of the 50 meetings at three different conflict levels under a round\-robin turn rule with a Latin\-square seat rotation: each meeting is executed three times, rotating the participants through the speaking positionsp1p\_\{1\}–p3p\_\{3\}\. The other settings and constraints are the same as the previous analysis example\.

##### Results

Figure[5](https://arxiv.org/html/2607.18806#S4.F5)shows no statistically significant difference across speaking positions inaligned\. Inmixedandconflicting, however, proposals from the second and third participants have significantly higher acceptance rates than those from the first participant\. This may be because participants have no preference conflict inaligned, so proposals are likely to be accepted regardless of speaking order\. On the other hand, inmixedandconflicting, preference conflicts make it difficult for the first participant to propose an itinerary that satisfies the preferences of the others\. Later participants can use more feedback from earlier proposals and voting results to propose itineraries that better reflect the preferences of others, resulting in higher acceptance rates\. Additionally, we found that most accepted proposals from the first participant were made after asking the others for input\. In contrast, participants who spoke later could often make accepted proposals without asking, relying instead on information from the preceding discussion\.

## 5Conclusion

In this paper, we proposed a new LLM\-based group travel planning framework and demonstrated its validity and utility through system\-level evaluation and use cases\. We believe that our framework supports research on the behavior of LLM agents and automatic evaluation of recommender systems for group travel planning\. Beyond these uses, it may also enable new applications, such as recommender systems where LLM agents act on behalf of group members who are unable to participate and provide ideas from their perspectives \(see Appendix[C](https://arxiv.org/html/2607.18806#A3)\)\.

## References

- L\. Chen, J\. Cao, H\. Chen, W\. Liang, H\. Tao, and G\. Zhu \(2021\)Attentive multi\-task learning for group itinerary recommendation\.Knowl\. Inf\. Syst\.63\(7\),pp\. 1687–1716\.External Links:ISSN 0219\-1377,[Link](https://doi.org/10.1007/s10115-021-01567-3),[Document](https://dx.doi.org/10.1007/s10115-021-01567-3)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1)\.
- X\. Cheng, Y\. Hu, L\. Zheng, Z\. Pan, X\. Li, and Y\. Liu \(2026\)GroupTravelBench: benchmarking llm agents on multi\-person travel planning\.External Links:2605\.25200,[Link](https://arxiv.org/abs/2605.25200)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1)\.
- J\. Choi, J\. Yoon, J\. Chen, S\. Jha, and T\. Pfister \(2026\)ATLAS: constraints\-aware multi\-agent collaboration for real\-world travel planning\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=mIYGiBf9Pm)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px2.p1.1)\.
- T\. Hu and N\. Collier \(2024\)Quantifying the persona effect in LLM simulations\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 10289–10307\.External Links:[Link](https://aclanthology.org/2024.acl-long.554/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.554)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1)\.
- H\. Jiang, X\. Zhang, X\. Cao, C\. Breazeal, D\. Roy, and J\. Kabbara \(2024\)PersonaLLM: investigating the ability of large language models to express personality traits\.InFindings of the Association for Computational Linguistics: NAACL 2024,K\. Duh, H\. Gomez, and S\. Bethard \(Eds\.\),Mexico City, Mexico,pp\. 3605–3627\.External Links:[Link](https://aclanthology.org/2024.findings-naacl.229/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-naacl.229)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1)\.
- D\. Kikuta, H\. Ikeuchi, K\. Tajiri, and Y\. Nakano \(2024\)RouteExplainer: an explanation framework for vehicle routing problem\.InAdvances in Knowledge Discovery and Data Mining,pp\. 30–42\.External Links:ISBN 9789819722594,ISSN 1611\-3349,[Link](http://dx.doi.org/10.1007/978-981-97-2259-4_3),[Document](https://dx.doi.org/10.1007/978-981-97-2259-4%5F3)Cited by:[§2\.2](https://arxiv.org/html/2607.18806#S2.SS2.p1.1)\.
- W\. Kwon, Z\. Li, S\. Zhuang, Y\. Sheng, L\. Zheng, C\. H\. Yu, J\. Gonzalez, H\. Zhang, and I\. Stoica \(2023\)Efficient memory management for large language model serving with pagedattention\.InProceedings of the 29th Symposium on Operating Systems Principles,SOSP ’23,New York, NY, USA,pp\. 611–626\.External Links:ISBN 9798400702297,[Link](https://doi.org/10.1145/3600006.3613165),[Document](https://dx.doi.org/10.1145/3600006.3613165)Cited by:[§3](https://arxiv.org/html/2607.18806#S3.SS0.SSS0.Px1.p1.1)\.
- G\. Li, H\. A\. A\. K\. Hammoud, H\. Itani, D\. Khizbullin, and B\. Ghanem \(2023\)CAMEL: communicative agents for ”mind” exploration of large language model society\.InThirty\-seventh Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=3IyL2XWDkG)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p2.1)\.
- B\. Liu, J\. Ge, and J\. Wang \(2025\)Vaiage: a multi\-agent solution to personalized travel planning\.External Links:2505\.10922,[Link](https://arxiv.org/abs/2505.10922)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px2.p1.1)\.
- T\. N\. Nguyen and F\. Ricci \(2018\)A chat\-based group recommender system for tourism\.Information Technology & Tourism18\(1\),pp\. 5–28\.External Links:[Document](https://dx.doi.org/10.1007/s40558-017-0099-y)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1)\.
- OpenAI \(2025\)Gpt\-oss\-120b & gpt\-oss\-20b model card\.External Links:2508\.10925,[Link](https://arxiv.org/abs/2508.10925)Cited by:[§3](https://arxiv.org/html/2607.18806#S3.SS0.SSS0.Px1.p1.1)\.
- OpenAI \(2026\)Introducing GPT‑5\.4 mini and nano\.Note:Accessed 2026\-07\-16External Links:[Link](https://openai.com/index/introducing-gpt-5-4-mini-and-nano/)Cited by:[§3](https://arxiv.org/html/2607.18806#S3.SS0.SSS0.Px1.p1.1)\.
- J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein \(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology,UIST ’23,New York, NY, USA\.External Links:ISBN 9798400701320,[Link](https://doi.org/10.1145/3586183.3606763),[Document](https://dx.doi.org/10.1145/3586183.3606763)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p2.1)\.
- Qwen Team \(2026\)Qwen3\.5: towards native multimodal agents\.External Links:[Link](https://qwen.ai/blog?id=qwen3.5)Cited by:[§3](https://arxiv.org/html/2607.18806#S3.SS0.SSS0.Px1.p1.1)\.
- H\. Singh, N\. Verma, Y\. Wang, M\. Bharadwaj, H\. Fashandi, K\. Ferreira, and C\. Lee \(2024\)Personal large language model agents: a case study on tailored travel planning\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track,F\. Dernoncourt, D\. Preoţiuc\-Pietro, and A\. Shimorina \(Eds\.\),Miami, Florida, US,pp\. 486–514\.External Links:[Link](https://aclanthology.org/2024.emnlp-industry.37/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-industry.37)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px2.p1.1)\.
- Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu, A\. H\. Awadallah, R\. W\. White, D\. Burger, and C\. Wang \(2024\)AutoGen: enabling next\-gen LLM applications via multi\-agent conversations\.InFirst Conference on Language Modeling,External Links:[Link](https://openreview.net/forum?id=BAakY1hNKS)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p2.1)\.
- J\. Xie, K\. Zhang, J\. Chen, T\. Zhu, R\. Lou, Y\. Tian, Y\. Xiao, and Y\. Su \(2024\)TravelPlanner: a benchmark for real\-world planning with language agents\.InForty\-first International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=l5XQzNkAOe)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px2.p1.1)\.
- X\. Zhou, H\. Zhu, L\. Mathur, R\. Zhang, H\. Yu, Z\. Qi, L\. Morency, Y\. Bisk, D\. Fried, G\. Neubig, and M\. Sap \(2024\)SOTOPIA: interactive evaluation for social intelligence in language agents\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=mM7VurbA4r)Cited by:[Appendix B](https://arxiv.org/html/2607.18806#A2.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p1.1),[§1](https://arxiv.org/html/2607.18806#S1.p2.1)\.

\# ParticipantsCompletion\(%\)↑\\uparrowConsensus\(%\)Constraint err\.\(%\)↓\\downarrowAction fail\.\(%\)↓\\downarrowTurnsDuration\(min\)Tokens \[K\]\(in / out\)ValidityM=3M=31001001001000\.00\.00\.10\.112\.812\.833\.633\.6134134/3737✓M=5M=51001001001001\.61\.60\.80\.825\.425\.4149\.3149\.3659659/9999✓M=10M=1010010096962\.42\.40\.30\.368\.668\.6471\.5471\.530743074/406406✓Table 3:Additional system validation with Qwen3\.5\-9B across participant sizesM∈\{3,5,10\}M\\in\\\{3,5,10\\\}## Appendix ALicense

Our code is licensed under NTT’s proprietary license, which restricts the use of our code to research purposes only\. Since our framework is designed primarily as a simulation tool to support research on group travel planning, this restriction does not pose a significant limitation\.

## Appendix BRelated Work

##### Group travel planning

Nguyen and Ricci \([2018](https://arxiv.org/html/2607.18806#bib.bib3)\)proposed a system that supports consensus building by monitoring chat discussions and adapting recommendations to evolving user preferences\.Chenet al\.\([2021](https://arxiv.org/html/2607.18806#bib.bib2)\)proposed a method that integrates member relationships and POI information through attention mechanisms to optimize group itinerary recommendation\. More recent work has investigated whether LLMs can help group travel planning while resolving preference conflicts among group membersChenget al\.\([2026](https://arxiv.org/html/2607.18806#bib.bib5)\)\. These works focus on supporting humans, whereas our work explores the use of LLM agents as proxies for humans in group travel planning\.

##### Travel planning with LLM agents

Recent work has proposed benchmarks for evaluating the travel\-planning capabilities of LLMsXieet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib16)\); Singhet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib19)\), as well as methods for improving their performanceLiuet al\.\([2025](https://arxiv.org/html/2607.18806#bib.bib18)\); Choiet al\.\([2026](https://arxiv.org/html/2607.18806#bib.bib17)\)\. However, these studies primarily focus on individual travel planning\.

##### LLM\-based multi\-agent simulation

LLM\-based multi\-agent simulation has been widely studiedParket al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib8)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib6)\); Jianget al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib7)\); Hu and Collier \([2024](https://arxiv.org/html/2607.18806#bib.bib9)\)\. However, to our knowledge, no prior work has specifically addressed such discussion in group travel planning\. From an implementation perspective, although general\-purpose multi\-agent frameworksLiet al\.\([2023](https://arxiv.org/html/2607.18806#bib.bib11)\); Wuet al\.\([2024](https://arxiv.org/html/2607.18806#bib.bib10)\)provide basic functionality for orchestrating agent interactions, users must implement domain\-specific components for group travel planning\. Our framework provides these components by default, allowing users to focus on their use cases\.

## Appendix CHuman Participant \(Interactive Mode\)

AI Tour Meeting also allows human users to participate in meetings\. At their turn, they can interact with the LLM participants through the chat box by executing the same actions defined in Eq\. \([2](https://arxiv.org/html/2607.18806#S2.E2)\) as the LLM participants\. Figure[6](https://arxiv.org/html/2607.18806#A3.F6)shows snapshots of the chat box during voting, when asking another participant a question, and when proposing an itinerary\. When proposing an itinerary, users can manually modify destinations\. By clicking the “Generate with AI” button, they can interact with an LLM and have it refine their proposal\. A use case of this feature is to represent friends who are unable to attend the meeting as LLM participants based on their personas\. By discussing the itinerary with these agents, users can incorporate the absent participants’ perspectives into the tour planning\.

![Refer to caption](https://arxiv.org/html/2607.18806v1/x5.png)Figure 6:Chat boxes for human participants\.
## Appendix DAdditional Validation: Participant Size

##### Setup

We additionally validate whether the framework correctly completes meetings as the number of participants increases\. We generate 50 synthetic meetings for each participant size \(5 and 10 participants\) using GPT\-5\.4\-mini, and evaluate them with Qwen3\.5\-9B under the same settings as the system validation\.

##### Result

Table[3](https://arxiv.org/html/2607.18806#S5.T3)shows that our framework maintains a 100% completion rate as the number of participants increases to 5 and 10\. Although the constraint error rate increases slightly as the group size increases, because itineraries tend to include more destinations, it remains sufficiently low\. Additionally, we found that the number of turns grows slightly faster than the number of participants\.

## Appendix EContext Management

##### Input context

Figure[7](https://arxiv.org/html/2607.18806#A6.F7)shows the structure of the input context of a participant\. The context starts with a system prompt defining the participant’s persona and the meeting settings, after which phase markers and action logs are appended alternately as the memory\. During the participant’s turn, the actions taken so far in the current turn and the phase instructions are appended to this memory, and the model generates the next action\. Once the turn ends, all the actions are recorded in the action logs\.

The phase marker is a system message that indicates the boundary between the conversation and voting phases\. Action logs store the actions \(Eq\.[2](https://arxiv.org/html/2607.18806#S2.E2)\) of all participants in each phase in speaking order\. When constructing the LLM input, the system prompt is assigned thesystemrole, while phase markers and phase instructions are assigned theuserrole\. The participant’s own previous actions are assigned theassistantrole, whereas actions from the other participants are assigned theuserrole, making it easier for the model to distinguish between itself and the other participants\. Note that each participant action is also prefixed with the speaker’s name\. The phase instruction is temporary context inserted immediately before action generation\. It describes the actions available in the current phase and specifies the JSON schema for the output\.

##### Retry loop

If an action output contains invalid JSON or violates a time or cost constraint in the proposed itinerary, the corresponding error message is temporarily appended to the end of the context, and the model is prompted to generate the output again\. In this paper, the maximum number of retries is set to three\.

##### Context compaction

Our framework supports three context compaction methods: summarization, turn\-window truncation, and token\-based truncation\. In the summarization method, a specified portion of the context is summarized and inserted immediately after the system prompt, followed by the remaining previous\-turn context in its original form\. The turn\-window method retains only the most recent fixed number of turns and removes all earlier turns from memory\. The token\-based truncation method retains only the most recent context up to a specified number of tokens and removes the rest\.

## Appendix FPrompts

Figures[8](https://arxiv.org/html/2607.18806#A6.F8)–[14](https://arxiv.org/html/2607.18806#A6.F14)list the main prompts used by our framework: the system prompt that encodes the persona \(Section[2\.2](https://arxiv.org/html/2607.18806#S2.SS2)\), the phase instructions in the conversation and voting phases, the prompts for inviting other participants, and the post\-meeting satisfaction evaluation prompt used in Section[4\.1](https://arxiv.org/html/2607.18806#S4.SS1)\. The system prompt can be customized by users\. Placeholders in curly braces \(e\.g\.,\{name\},\{format\_instructions\}\) are filled at runtime\.

![Refer to caption](https://arxiv.org/html/2607.18806v1/x6.png)Figure 7:Input context of a participant\.`System prompt`Figure 8:System prompt for a participant\.`Phase instruction for actions in the conversation phase Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 9:Phase instructions for actions in the conversation phase \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\.`Phase instruction for itinerary proposal in the conversation phase Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 10:Phase instruction for itinerary proposal in the conversation phase \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\. when the model selects theproposeaction in Figure[9](https://arxiv.org/html/2607.18806#A6.F9), this instruction is inserted into the context, prompting the model to output an itinerary according to the specified output schema\.`Phase instruction for actions in the voting phase Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 11:Phase instruction for actions in the voting phase \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\.`Phase instruction for answering questions from other participants Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 12:Phase instruction for answering questions from other participants \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\.`Inviting turn rule: next\-speaker prompt Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 13:Phase instruction for inviting the next speaker, which is used when the turn rule is set to invitation \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\.`Post\-meeting satisfaction evaluation prompt Output schema embedded in \{format\_instruction\} \(simplified version\)`Figure 14:Post\-meeting satisfaction evaluation prompt \(top\) and the simplified output schema embedded in\{format\_instructions\}\(bottom\)\.