Technical Summary: LLMs as Orchestrators for Constraint-Compliant Multi-Agent Optimization

Problem Statement

E-commerce recommendation systems face a critical challenge: they must simultaneously optimize multiple conflicting objectives (accuracy, diversity, novelty, fairness) while strictly adhering to hard business constraints. In production environments, constraints such as seller coverage, category fairness, and new-item exposure are non-negotiable; violating them even once is unacceptable.

Prior research has largely treated these constraints as soft penalties or regularization terms, or focused on item scoring and interaction via Large Language Models (LLMs). These approaches often lead to persistent constraint violations in deployment or result in Pareto fronts dominated by infeasible solutions. While multi-objective evolutionary algorithms exist, they struggle to balance solution quality with strict feasibility. Furthermore, existing LLM-based recommenders typically operate at the level of content understanding or user interaction, leaving the potential of LLMs as high-level optimization coordinators for constrained search largely unexplored.

Methodology: DualAgent-Rec

The authors propose DualAgent-Rec, a framework that leverages an LLM to orchestrate a dual-agent evolutionary optimization process. The system is designed to guarantee 100% constraint satisfaction while optimizing the trade-off between accuracy and diversity.

1. Problem Formulation

The recommendation task is formulated as a constrained multi-objective optimization problem:

• Objectives: Maximize relevance ($f_1$), diversity ($f_2$, via intra-list distance), and novelty ($f_3$, via inverse popularity).
• Hard Constraints:
  • Category Fairness: Gini coefficient of category distribution $\le \theta_{fair}$.
  • Seller Coverage: Number of distinct sellers $\ge \theta_{seller} \cdot k$.
  • New Item Exposure: Number of items added in the last 30 days $\ge \theta_{new} \cdot k$.

2. Dual-Agent Architecture

The framework decomposes the search space into two specialized agents that evolve populations in parallel:

• Exploitation Agent: Focuses on refining high-quality solutions within the feasible region. It employs Constraint Domination Principle (CDP) for selection, strictly prioritizing feasible solutions over infeasible ones and using constraint violation magnitude as a tie-breaker. It uses DE/pbest/1 variation to accelerate convergence.
• Exploration Agent: Maintains a separate population and performs unconstrained Pareto search with a doubled mutation rate. It ignores constraints during selection to traverse underexplored regions of the objective space, acting as a diversity reservoir to prevent premature convergence to suboptimal feasible basins.
• Bidirectional Knowledge Transfer: Elite solutions are exchanged between agents based on crowding distance, ensuring that the exploitation agent benefits from novel discoveries while the exploration agent receives guidance from high-quality feasible solutions.

3. LLM-Based Coordination

A central innovation is the use of an LLM as a dynamic resource allocator. Instead of fixed heuristics, the LLM coordinator:

• Observes: A structured summary of optimization metrics (hypervolume, improvement rates), constraint statistics (feasibility ratio, violation magnitude), and user context.
• Decides: Outputs an allocation ratio $\alpha \in [0, 1]$ that determines the population sizes of the Exploitation and Exploration agents ($|P_{exploit}| = \alpha N$).
• Adapts: Increases $\alpha$ (favoring exploitation) when feasibility is stable and progress slows; decreases $\alpha$ (favoring exploration) when constraint violations persist or hypervolume stagnates. This provides interpretable, context-aware orchestration without hand-crafted scheduling rules.

4. Adaptive Constraint Handling

To reconcile early exploration with strict final feasibility, the system uses a self-calibrating $\epsilon$-relaxation strategy:

• Constraints are initially relaxed ($\epsilon_0$) based on the 80th percentile of initial violations, allowing ~20% of early solutions to be treated as feasible.
• The relaxation parameter $\epsilon_t$ decays smoothly over generations ($\epsilon_t = \epsilon_0 \cdot \gamma^{t/T_{max}}$), gradually tightening constraints until exact feasibility is enforced at convergence.

Key Contributions

1. Dual-Agent Architecture: A novel separation of exploration and exploitation into specialized agents, improving Pareto front coverage and optimization efficiency under competing objectives.
2. LLM Orchestration: The introduction of an LLM as a high-level coordinator that adaptively manages resource allocation based on optimization dynamics and constraint satisfaction, replacing brittle rule-based schedules.
3. Adaptive Constraint Strategy: A self-calibrating $\epsilon$-relaxation mechanism that guarantees hard constraint satisfaction in final solutions while preserving exploration efficiency during early stages.
4. Empirical Validation: Extensive experiments demonstrating that the framework achieves 100% constraint satisfaction while outperforming strong baselines in Pareto hypervolume and convergence speed.

Experimental Results

Experiments were conducted on the Amazon Reviews 2023 dataset across three diverse categories: All_Beauty (small), Electronics (medium), and Clothing_Shoes_Jewelry (large).

• Constraint Satisfaction: DualAgent-Rec achieved 100% feasibility across all categories, validating the effectiveness of the adaptive $\epsilon$-relaxation and CDP mechanisms.
• Pareto Performance: The framework improved the Pareto Hypervolume (HV) by 4–6% compared to single-population baselines and rule-based allocation variants.
• Accuracy-Diversity Trade-off: While the unconstrained variant achieved higher diversity, it failed feasibility (72%). DualAgent-Rec maintained competitive diversity (0.247) and accuracy (NDCG@10: 0.581) while strictly satisfying constraints.
• Ablation Insights:
  • Population Size: Smaller populations (50) offered faster runtime (5.2s) suitable for latency-sensitive apps, while larger populations (200) maximized diversity for offline batch processing.
  • Constraint Strictness: This was the most impactful factor; stricter constraints reduced HV but maximized accuracy, offering a tunable trade-off for deployment.
  • LLM Impact: The LLM coordinator provided modest but consistent HV improvements (~1%) over fixed-rule allocation, with significant qualitative benefits in interpretability and automatic adaptation.

Significance and Claims

The paper claims to bridge the gap between academic multi-objective optimization research and real-world production requirements. Its primary significance lies in demonstrating that LLMs can effectively act as orchestration agents for complex, constrained optimization problems, moving beyond their traditional roles in scoring or interaction.

The authors assert that DualAgent-Rec offers a principled framework where constraint satisfaction is treated as a "first-class objective," ensuring that every deployed recommendation list is feasible by design rather than in expectation. The work provides practical deployment guidelines, showing that the framework can be tuned for different scenarios (e.g., real-time vs. offline) while maintaining the non-negotiable requirement of hard constraint compliance.