Точка Синхронізації

Researchers and engineers at Meituan, China’s leading food delivery and retail platform, have published a technical paper on the arXiv preprint server in February 2025 detailing a new multimodal generative retrieval model designed to enhance search precision for food delivery services. The team developed this advanced architecture to solve the complex challenge of aligning diverse user queries with a vast inventory of food items and merchants by leveraging both visual and textual data. This innovation addresses the growing demand for more personalized and accurate search results in high-frequency e-commerce environments where traditional search methods often fail to capture the nuances of user intent. The paper highlights a critical flaw in existing mainstream retrieval systems, which typically utilize a "dual-tower" architecture. In these traditional models, queries and items are processed separately through dedicated towers, with optimization occurring simultaneously across internal and external tasks. However, the Meituan researchers discovered that this joint optimization process often leads to "modality imbalance," where one type of data—such as text—dominates the training process, causing the model to underutilize other critical information like high-resolution food imagery or localized merchant metadata. To overcome these limitations, the proposed model introduces a staged pretraining strategy that carefully balances the influence of different data modalities. By decoupling the training phases, the system ensures that both visual features and semantic text descriptions contribute effectively to the retrieval process. This approach allows the generative model to better predict and surface relevant results, even when user queries are vague or highly specific. The implementation of this technology at Meituan represents a significant step forward in applying generative AI to real-world logistics and consumer service platforms. Ultimately, this research suggests that the future of digital marketplaces lies in the sophisticated integration of multimodal data. For a platform like Meituan, which handles millions of orders daily, the ability to provide precise, context-aware retrieval can significantly improve user satisfaction and merchant visibility. The staged pretraining framework not only improves the accuracy of the search engine but also provides a more robust foundation for future AI-driven features in the competitive food delivery landscape.