LLMs as Orchestrators: Constraint-Compliant Multi-Agent Optimization for Recommendation Systems

TL;DR

This work addresses constrained multi-objective recommendation under hard business constraints by formulating the problem as a constrained multi-objective optimization with objective vector $\mathbf{f}(L)=[f_1(L),f_2(L),f_3(L)]$ and constraints $g_j(L)\le 0$. It introduces DualAgent-Rec, a dual-agent framework with an Exploitation Agent (constraint-dominated search) and an Exploration Agent (unconstrained Pareto search) whose collaboration is guided by an LLM-based coordinator and reinforced by an adaptive $\epsilon$-relaxation mechanism that ensures feasibility at convergence. Empirical results on Amazon Reviews 2023 demonstrate 100% constraint satisfaction and a 4–6% improvement in Pareto hypervolume over strong baselines, while maintaining competitive accuracy-diversity trade-offs. The approach offers interpretable coordination and deployment-friendly guidelines, effectively bridging academic optimization methods and production-ready recommendation systems.

Abstract

Recommendation systems must optimize multiple objectives while satisfying hard business constraints such as fairness and coverage. For example, an e-commerce platform may require every recommendation list to include items from multiple sellers and at least one newly listed product; violating such constraints--even once--is unacceptable in production. Prior work on multi-objective recommendation and recent LLM-based recommender agents largely treat constraints as soft penalties or focus on item scoring and interaction, leading to frequent violations in real-world deployments. How to leverage LLMs for coordinating constrained optimization in recommendation systems remains underexplored. We propose DualAgent-Rec, an LLM-coordinated dual-agent framework for constrained multi-objective e-commerce recommendation. The framework separates optimization into an Exploitation Agent that prioritizes accuracy under hard constraints and an Exploration Agent that promotes diversity through unconstrained Pareto search. An LLM-based coordinator adaptively allocates resources between agents based on optimization progress and constraint satisfaction, while an adaptive epsilon-relaxation mechanism guarantees feasibility of final solutions. Experiments on the Amazon Reviews 2023 dataset demonstrate that DualAgent-Rec achieves 100% constraint satisfaction and improves Pareto hypervolume by 4-6% over strong baselines, while maintaining competitive accuracy-diversity trade-offs. These results indicate that LLMs can act as effective orchestration agents for deployable and constraint-compliant recommendation systems.

Figures (5)

• Figure 1: DualAgent-Rec framework architecture with three-stage pipeline. Input Processing: User history, item catalog, and business constraints (fairness, seller coverage, new item exposure) are encoded. Dual-Agent Optimization: The LLM Coordinator dynamically allocates resources ($\alpha$) between the Exploitation Agent (CDP-based selection) and Exploration Agent (doubled mutation rate). Bidirectional knowledge transfer exchanges elite solutions. Adaptive $\epsilon$-relaxation gradually tightens constraints. Output Selection: Pareto-optimal feasible solutions are filtered to produce the final recommendation list.
• Figure 2: Optimization convergence curves across methods. DualAgent-Rec achieves faster initial convergence and higher final hypervolume compared to variants without LLM coordination or with single population structure. The dual-agent architecture provides early diversity advantage through the exploration agent.
• Figure 3: Constraint satisfaction progress over optimization generations. The adaptive $\epsilon$-relaxation mechanism enables gradual convergence toward 100% feasibility while maintaining search diversity in early generations. The self-calibrating mechanism allows initial exploration of promising infeasible regions.
• Figure 4: Pareto front comparison between DualAgent-Rec and Single Population baseline across the accuracy-diversity trade-off space. The dual-agent framework achieves better spread and coverage, particularly in high-diversity regions where the Exploration Agent discovers solutions missed by unified search.
• Figure 5: Ablation study heatmap showing hyperparameter sensitivity. Population size and constraint strictness have the largest impact on optimization quality.