MARS: Multi-Agent Re-ranking for Food Delivery Recommendations

Researchers have introduced a novel system known as MARS (Multi-Agent Re-ranking System) that enhances food delivery recommendations for repeat orders by leveraging large language models within a structured approach. MARS integrates global preference signals based on LightGCN, local peer evidence through Swing, geospatial filtering, and LLM reasoning driven by prompts, taking into account behavioral, temporal, and geographic factors. The operation unfolds in two phases: predicting cuisine and ranking vendors. Its performance was assessed using two actual benchmarks from Delivery Hero, namely DHRD-SE and DHRD-SG. This study explores the potential of pre-trained LLMs when paired with lightweight collaborative retrieval and contextual filtering in a controlled hybrid environment.