Collab-REC: An LLM-based Agentic Framework for Balancing Recommendations in Tourism

TL;DR

Collab-REC introduces a multi-agent LLM framework for tourism recommendations that balances personalization, popularity, and sustainability. Three specialized agents generate city candidates while a non-LLM moderator grounds outputs to a 200-city knowledge base, computes a composite score, and iteratively merges proposals through multi-round negotiation to curb hallucinations and bias. Empirical evaluation on SynthTRIPS-based European city queries shows that multi-agent, multi-round negotiation yields higher relevance and greater diversity than baselines, with notable reductions in popularity bias and improved sustainability considerations. The work advances multi-stakeholder GenAI recommender design by demonstrating how modular agent roles, grounded moderation, and iterative compromise can produce more balanced, context-aware recommendations, albeit at increased computational cost.

Abstract

We propose Collab-REC, a multi-agent framework designed to counteract popularity bias and enhance diversity in tourism recommendations. In our setting, three LLM-based agents -- Personalization, Popularity, and Sustainability generate city suggestions from complementary perspectives. A non-LLM moderator then merges and refines these proposals via multi-round negotiation, ensuring each agent's viewpoint is incorporated while penalizing spurious or repeated responses. Experiments on European city queries show that Collab-REC improves diversity and overall relevance compared to a single-agent baseline, surfacing lesser-visited locales that often remain overlooked. This balanced, context-aware approach addresses over-tourism and better aligns with constraints provided by the user, highlighting the promise of multi-stakeholder collaboration in LLM-driven recommender systems.

Figures (6)

• Figure 1: Overview of the Collab-Rec workflow to generate city trip recommendations using multiple LLM agents. The non-LLM Moderator evaluates and combines the agent proposals, iterating through multiple rounds to refine the final recommendation set, which is then communicated to the user.
• Figure 2: Overview of the Collab-Rec Moderator core to generate city trip recommendations using multiple LLM agents. The moderator orchestrates the multi-round negotiation process, evaluates agent proposals, and aggregates them into a final recommendation list.
• Figure 3: Agent Dominance when split by the popularity levels of the queries. Sustainability and personalization agents tend to dominate for medium- and low-popularity queries, but the popularity agent takes a substantial lead for high-popularity queries, signaling a potential popularity bias inherent in pre-trained models. —--- (dotted line) represents the results from $\textit{gemini-2.5-flash}$ while --- (continuous line) represents the results from $\textit{gpt-o4-mini }$.
• Figure 4: City Distributions when split by the popularity levels of the recommended cities for 50 randomly sampled cities. For brevity, the x-axis represents the IATA codes of the respective cities. MAMI tends to provide lesser-known cities as a recommendation when compared to SASI and MASI.
• Figure 5: Agent Behavior metrics showing the agents' reliability score and hallucination rate over multiple rounds. —--- (dotted line) represents the results from $\textit{gemini-2.5-flash}$ while --- (continuous line) represents the results from $\textit{gpt-o4-mini }$.