AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AI Agents That Plan Group Trips Show How Negotiations Break Down

Group travel planning is a negotiation problem: each participant brings preferences, budgets, and time constraints, and a good itinerary has to balance all of them. Researchers are now testing whether large language models (LLMs)—the technology behind chatbots—can act as stand-ins for real group members in such discussions. In a new paper, NTT researcher Daisuke Kikuta proposes AI Tour Meeting, a framework in which multiple LLM-based agents, each given a distinct persona, collaborate through natural language to plan a one-day tour and vote on proposed itineraries.

Each agent is defined by attributes such as background, personality, preferences, personal goals, speaking style, and role in the meeting. During a conversation phase, agents search for information, ask each other questions, reflect on the discussion, and propose structured itineraries containing destinations, costs, arrival times, and transport details. After a proposal, the group enters a voting phase; depending on the chosen rule, the itinerary is accepted by majority vote, unanimous consent, or another mechanism. The meeting ends when every agent consecutively says they are satisfied with the current route.

The author validated the framework using 50 synthetic three-person meetings generated by GPT-5.4-mini, then reran them with several models. Qwen3.5-4B, Qwen3.5-9B, gpt-oss-20b, and gpt-5.4-mini completed 100% of meetings, meaning they ended with at least one accepted itinerary. The smallest model, Qwen3.5-2B, completed only 82% and reached consensus in 58% of meetings; it also produced invalid actions, including rejecting proposals by referring to a nonexistent “current route.” Constraint errors—proposals that violated budget or time rules—fell from 14.2% with the 2B model to 0.0% with Qwen3.5-9B. The author concludes that the framework needs models comparable to or better than Qwen3.5-4B.

The study then examined what happens when preferences conflict. Using Qwen3.5-9B agents, the author ran 50 meetings at each of three conflict levels: aligned, mixed, and conflicting. As conflict increased, average turns rose from 10.3 to 25.9, and proposals rose from 2.3 to 6.5. Consensus fell from 100% in aligned and mixed settings to 94% in conflicting settings—a drop of six percentage points—while the share of participants scoring their satisfaction at 4 or lower on a 10-point scale rose from 0.7% to 11.3%, a change of 10.6 percentage points. One example shows the mechanism: in a mixed-preference meeting, participant Elena’s proposals were rejected nine times in a row; after 15 proposals, Jisoo urged Elena to accept a route so the group could finish. Elena agreed, but her post-meeting satisfaction score was 3, indicating major personal goals went unmet. The authors describe this as compromise driven by social pressure rather than genuine agreement.

The paper also reports that in mixed and conflicting settings, proposals from the second and third speakers were accepted significantly more often than those from the first speaker. The authors attribute this to later speakers having more feedback from earlier proposals and votes, while first speakers often had to ask questions before making a successful proposal.

A key limitation is that all reported meetings were synthetic, generated and evaluated by LLMs rather than tested with human groups. The results therefore show how AI agents behave in this setting, not that the behavior matches human deliberation. Large runs also carry real costs: in an additional validation, 10-agent meetings averaged 471.5 minutes and about 3.07 million input tokens on the authors’ setup.

The broader value of the work is a flexible testbed for studying multi-agent negotiation—and, potentially, a way for an absent group member to be represented by an agent that voices their interests. What it does not do is prove that AI agents plan trips better than people; instead, it offers a controlled way to watch how they compromise, stall, and sometimes agree for the wrong reasons.