AI Tour Meeting: Group Travel Planning by LLM Agents
Daisuke Kikuta
cs.AI, cs.CL, cs.MA
2026-07-21
A multi-agent travel-planning framework where persona agents discuss and vote to consensus; 2B models cannot sustain negotiation and 4B is the floor for 100% consensus.
Let multiple LLM agents plan a group trip like a committee, producing an itinerary everyone endorses. The hard part is not single-agent planning but group decision-making under preference conflict: one person wants to hike, another wants to shop, and the question is how they agree without one person being sacrificed the whole trip. Existing single-agent planners do not handle this kind of negotiation with opposing preferences, and general agent frameworks rarely treat consensus as an explicit termination condition.
Each agent carries a full persona (Name, Background, Personality, Preference, Goal, Role, Tone, Explanation-style), which decides whom it speaks for and what it cares about. The flow has two phases, discussion then voting, looping until consensus. In discussion, agents can search (look things up), ask (question others), or reflect (reconsider their stance), and can propose, express satisfied, or pass; in voting they accept, reject, or score. Five turn rules (Round-robin, Inviting, Facilitating, Random, Parallel) decide who speaks next, and five voting rules (Majority, Unanimous, Single-decider, Most-pleasure, Least-misery) decide what counts as passing. This is a configurable negotiation framework where itinerary, personas, actions, and rules are all swappable.
The key finding is a capability threshold:
| Model | Completion | Consensus |
| Qwen3.5-2B | 82% | 58% (fails) |
| 4B / 9B / gpt-oss-20b / gpt-5.4-mini | 100% | 100% |
Sharper preference conflict means more turns, harder consensus, and more victims: across aligned, mixed, and conflicting settings, turns run 10.3 / 12.1 / 25.9, consensus is 100 / 100 / 94%, and victim ratio (people whose preferences are ignored throughout) is 0.7 / 1.3 / 11.3%. The conflicting band drops consensus to 94% and victims to 11.3%, so even a good framework cannot suppress sharp opposition. Two structural findings stand out: later speakers have higher acceptance (more information makes agreement easier), and group sizes M = 3 / 5 / 10 all reach 100% consensus (size is not the bottleneck, capability is).
It offers a reproducible testbed for multi-agent group decision-making and quantifies a practical question, how capable a model must be to hold a negotiation: 2B is not enough, 4B is the floor. For people building multi-agent systems, this persona-discuss-vote scaffold can be lifted into other negotiation tasks. But this is fundamentally a simulation tool, not a capability breakthrough.
This is the thinnest of the six: a framework plus ablations, with no head-to-head against strong external baselines. The task domain is narrow (travel planning only), so generalization to harder negotiations is unproven. The license is NTT proprietary rather than open, so reproduction and commercial use hinge on reading the terms carefully.