Multimodal Generative Retrieval Model with Staged Pretraining for Food Delivery on Meituan

Researchers developed a new AI model for Meituan's food delivery platform that improves search results. It uses a staged training process to better balance learning from images and text, preventing one type of data from dominating. The model also learns from compressed data signatures. Tests showed significant improvements in recommendation accuracy, leading to a 1.12% increase in platform revenue and a 1.02% higher click-through rate in real-world use.