MARS: Multi-Agent Re-ranking for Repeat-Order Food Delivery Recommendation

As large language models (LLMs) become increasingly prevalent in recommendation systems, a critical question remains: how much of their performance is due to the inherent strength of the pre-trained backbone versus the surrounding retrieval and filtering infrastructure? This paper investigates this by introducing MARS, a modular multi-agent framework designed for repeat-order food delivery recommendation.

MARS employs a coarse-to-fine recommendation strategy that decomposes the task into two distinct stages. First, an 'Analyzer' agent predicts likely cuisines based on temporal, behavioral, and geographic context, informed by global preference signals from a LightGCN model. Second, a 'Critic' agent ranks candidate vendors within the filtered cuisine set using local peer evidence derived from Swing similarity. The framework uses a hub-and-spoke architecture where a 'Manager' agent orchestrates the workflow, handles state preparation, and applies deterministic fallback rules to ensure robust output.

The study evaluates MARS on two real-world datasets from Delivery Hero (DHRD-SE and DHRD-SG). The results show that MARS consistently outperforms strong non-LLM baselines, including graph-based and food-delivery-specific models. The authors find that stronger pre-trained backbones significantly improve ranking accuracy when paired with the same retrieval evidence. Furthermore, enabling inference-time 'thinking' or reasoning traces consistently boosts performance, suggesting that test-time computation is a vital component for LLM-based recommenders in complex, constrained environments.

This work provides a reproducible, transparent blueprint for building hybrid recommendation systems. By demonstrating that modular, agentic pipelines can leverage the reasoning capabilities of LLMs without requiring end-to-end task-specific training, the authors offer a scalable path for deploying LLMs in production environments where interpretability and modularity are essential.