WOWService - Technical Report on Meituan's Large-Scale Interactive System

What is WOWService?

WOWService, released by Meituan's LongCat team, is an intelligent interaction system that optimizes user experience and reduces costs by integrating large-scale language models (LLMs) and a multi-agent architecture. The system addresses five major challenges in intelligent interaction: cold-start data construction, multi-turn dialogue performance, changes in business rules, the limitations of a single LLM, and the challenge of quantitative evaluation in multi-turn dialogues, proposing innovative solutions. WOWService employs a four-stage training pipeline: Continuous Pre-training (CPT), Supervised Fine-tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning (RL), significantly improving model performance in business scenarios. Its multi-agent architecture achieves efficient processing of complex tasks through collaboration between a master agent and multiple specialized sub-agents. The system introduces a dual-driven mechanism of data and knowledge, further enhancing model performance through Self-Optimizing Training (SRT) and hybrid data-knowledge-driven methods.

Main functions of WOWService

• Data Construction and OptimizationWOWService employs a dual-driven strategy of data and knowledge, improving model training efficiency and performance while reducing data annotation costs through adaptive data hybrid optimization and high-quality data filtering.
• Multi-turn dialogue performance improvementThe system significantly enhances its ability to understand intent, follow rules, and extract solutions in multi-turn dialogues through a four-stage training process: continuous pre-training, supervised fine-tuning, direct preference optimization, and reinforcement learning.
• Business rule adaptation and self-evolutionWOWService has a self-optimizing training mechanism that can quickly adapt to changes in business rules, achieve continuous model evolution and optimization, and reduce business expansion costs.
• Multi-agent collaborationThe system introduces a multi-agent architecture, where the main agent is responsible for global dialogue control and specialized sub-agents handle specific tasks, thereby improving the service quality and flexibility of the system in complex scenarios.
• Automated evaluation and feedbackA comprehensive evaluation framework has been constructed, covering the basic performance of the model and the end-to-end agent effect evaluation, and the model can be continuously improved through data feedback.
• Multimodal interaction supportWOWService supports multimodal input, such as voice and images. By understanding intelligent agents through multimodal methods, it improves the system's ability to understand non-text input and expands application scenarios.

Technical Principles of WOWService

• Data and knowledge dual-driveWOWService integrates data-driven and knowledge-driven approaches, combining high-quality data and structured business knowledge to enhance the model's ability to follow business rules and knowledge points, thereby improving the model's performance in complex business scenarios.
• Four-stage training pipelineThe training process employs four stages: Continuous Pre-training (CPT), Supervised Fine-tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning (RL), to gradually improve the model's generality, business adaptability, and inference capabilities.
• Multi-agent architectureThe system introduces a main intelligent agent and multiple specialized sub-intelligent agents. The main intelligent agent is responsible for global dialogue control, while the sub-intelligent agents handle specific tasks. Through dynamic invocation and information integration, it achieves efficient processing of complex tasks.
• Self-optimization trainingBy automatically filtering high-quality service cases and rewriting low-quality dialogues, a self-evolving data loop is built, enabling the model to continuously optimize and adapt to new business needs.
• Hybrid data knowledge-drivenBy combining the flexibility of data-driven approaches with the adaptability of knowledge-driven rules, and through dynamic knowledge retrieval and reasoning enhancements, the accuracy and reliability of models in complex business processes can be improved.
• Reinforcement learning and preference optimizationBy leveraging reinforcement learning and direct preference optimization techniques, the model can be adjusted based on human feedback to generate outputs that better align with human preferences, thereby enhancing the user experience.
• Automated evaluation frameworkEstablish a comprehensive evaluation system that ranges from basic model performance to end-to-end agent effectiveness, supporting automated evaluation and continuous optimization to ensure the stability and efficiency of the model in practical applications.

WOWService project address

• arXiv technical paper: https://arxiv.org/pdf/2510.13291

Application scenarios of WOWService

• Meituan Intelligent Customer Service SystemWOWService has been successfully deployed in Meituan App's intelligent customer service system to handle user inquiries, provide 24/7 uninterrupted service, and significantly improve user satisfaction and problem-solving efficiency.
• Multi-business scenario coverageIt is widely used in dozens of business scenarios within Meituan, such as food delivery, hotel booking, and tourism, to provide personalized services according to different business needs.
• Complex business process processingIn complex business processes involving multiple roles (such as users, merchants, and delivery personnel), WOWService can flexibly schedule and coordinate resources to ensure efficient information transmission and service execution.
• Multimodal interaction supportIt supports multiple input methods such as voice and image, providing users with a more natural and convenient interactive experience, and is suitable for scenarios that require visual or auditory assistance.
• Proactive Collaboration and RecommendationsIt can proactively uncover users' potential needs, provide personalized recommendations and solutions, and improve user engagement and satisfaction.
• Intelligent outbound calling serviceEnables intelligent outbound calling functions, such as order confirmation and satisfaction surveys, to improve operational efficiency and user response speed.