Collab-REC framework overview: three specialist LLM agents (Popularity, Sustainability, Personalization) iteratively propose city candidates across rounds, with a deterministic Moderator merging proposals into a collective offer.
Figure 1. Overview of the Collab-REC framework. Three specialist LLM agents (Personalization, Popularity, Sustainability) iteratively propose city recommendations. A deterministic moderator grounds, scores, and aggregates proposals into a collective offer, broadcasting structured feedback until convergence.

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 iterative constrained refinement, ensuring each agent’s viewpoint is incorporated while penalizing spurious or repeated responses.


Extensive offline experiments on European city queries using LLMs from different sizes and model families demonstrate that Collab-Rec enhances diversity and overall relevance compared to a single-agent baseline, surfacing lesser-visited locales that are often overlooked. This balanced, context-aware approach better reflects a broader range of user and system-level considerations, highlighting the potential of multi-stakeholder collaboration in LLM-driven recommender systems.


Code, data, and other artifacts are available at github.com/ashmibanerjee/collab-rec.

Multi-Agent Systems Tourism Recommendation Popularity Bias Diversity LLM Agents Hallucination Control Sustainability Multi-Stakeholder
900
Tourism Queries
200
European Cities
6
LLM Backbones
3
Specialist Agents

Agentic Framework

Collab-REC decomposes multi-stakeholder tourism recommendation into role-specialized agents coordinated by a transparent, deterministic moderator. The framework iterates until convergence or patience-based early stopping.

Moderator Core diagram showing the internal steps: Initial Feedback (hallucination identification, previous rejection verifier) → Grounding & Assessment (Agent Success, Hallucination Rate, Agent Reliability) → Collective Offer and Collective Rejections fed back to LLM Agents A1, A2, A3.
Figure 2. Moderator Core (M). Each round the moderator identifies hallucinated or rejected cities, scores agent lists via grounded diagnostics (Agent Success, Hallucination Rate, Agent Reliability), and produces an updated Collective Offer Φt and rejection set Φ′t that constrain agents in the next round.

Specialist Agents

Agent a1
  Personalization Agent

Focuses on the user-centric perspective, prioritizing explicit filters and query-specific preferences such as budget, travel month, and personal interests (e.g., museums, nature, nightlife).

Agent a2
  Popularity Agent

Emphasizes the popularity dimension and is configured to mitigate short-head concentration by proposing less popular cities when the query suggests a preference for less crowded destinations.

Agent a3
  Sustainability Agent

Prioritizes sustainability-related attributes such as walkability, seasonality, and air-quality indicators, promoting environmentally preferable destinations to mitigate overtourism.

Deterministic Moderator: A non-LLM controller that (i) validates proposals against the city catalog, (ii) computes grounded diagnostic scores via rank-discounted scalarization, (iii) constructs the collective offer Φt, and (iv) broadcasts structured feedback — including a collective rejection set — to constrain agents in subsequent rounds. Two rejection strategies are supported: Aggressive and Majority.

Contributions

Problem Formulation

We formalize multi-stakeholder, multi-constraint tourism recommendation as a grounded multi-objective ranking problem, where feasibility is defined with respect to an explicit destination catalog and constraint satisfaction.

Framework Design

We introduce Collab-Rec, a modular multi-agent architecture with a transparent, deterministic moderator that scores candidates under multiple objectives and iteratively conditions agent outputs via structured feedback.

Efficiency-Aware Moderation

We propose and empirically validate a patience-based early stopping protocol, capturing most quality gains by ~4–5 rounds while substantially reducing inference time for API-served models.

Large-Scale Evaluation

We evaluate 900 queries across six LLM families (claude-sonnet-4-5, gemini-2.5-flash, gpt-oss-20b, gemma-3-12b, olmo3-7b-instruct, gemma-3-4b), with statistical testing and convergence behavior analysis.

Hallucination Control

Structured output constraints and catalog validation ensure agents output valid city names. The moderator penalizes invalid outputs and accumulates rejection sets across rounds to progressively reduce hallucinations.

Reproducibility

We release code, prompts, and evaluation artifacts to enable reproduction and extension. All prompts used by the three agents are included in the paper appendix.

Results & Discussion

Collab-REC consistently improves moderator success (constraint satisfaction under catalog validation) over the single-agent baseline (SASI) and the single-round multi-agent baseline (MASI) across all six LLM families. Quality gains plateau by approximately 4–5 rounds, validating the patience-based early stopping protocol.
Collab-REC shifts recommendation distributions toward the long tail of the city catalog, reducing Gini concentration and improving entropy and catalog coverage compared to both baselines. The Sustainability and Popularity agents together surface destinations that are systematically overlooked by single-agent prompting.
Agent reliability evolves from exploration (higher churn in early rounds) to stabilization as the collective offer converges. Hallucination rates decrease with grounded moderator feedback: rejection sets prevent repeated invalid outputs, and the structured output schema reduces syntactic invalidity.
Overhead scales approximately linearly with the number of rounds. Patience-based early stopping is shown to capture most quality gains while substantially reducing total inference calls, making the framework practical for API-served LLMs where token cost is a concern.
Ablations removing individual scoring components (success, reliability, or hallucination terms) show that each term contributes to the overall moderator score. Aggressive versus Majority rejection strategies differ in their convergence speed and diversity trade-offs, with Majority rejection being more permissive and typically achieving comparable quality with fewer rejected cities.

Acknowledgment

We thank the Google Developer Experts Program for their generous support through Google Cloud credits.

BibTeX

@misc{banerjee2025collabrec,
  title         = {Collab-REC: An LLM-based Agentic Framework for
                    Balancing Recommendations in Tourism},
  author        = {Banerjee, Ashmi and Satish, Adithi and
                    Aisyah, Fitri Nur and W{\"o}rndl, Wolfgang and
                    Deldjoo, Yashar},
  year          = {2025},
  eprint        = {2508.15030},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2508.15030}
}