What is WOWService
WOWService is an intelligent interaction system released by Meituan's LongCat team. By integrating large language models (LLMs) and a multi-agent architecture, it optimizes user experience and reduces costs. The system proposes innovative solutions to five major challenges in the field of intelligent interaction: cold-start data construction, multi-turn dialogue performance, business rule changes, the limitations of a single LLM, and the difficulty of quantitative evaluation for multi-turn dialogues. WOWService adopts a four-stage training pipeline: continual pre-training (CPT), supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning (RL), significantly improving the model's performance in business scenarios. Its multi-agent architecture enables efficient handling of complex tasks through collaboration between a main agent and multiple specialized sub-agents. The system introduces a data-and-knowledge dual-driven mechanism, further improving model performance through self-optimizing training (SRT) and a hybrid data-knowledge-driven approach.
Main Features of WOWService
-
Data Construction and Optimization: WOWService adopts a data-and-knowledge dual-driven strategy, improving model training efficiency and performance and reducing data annotation costs through adaptive data mixing optimization and high-quality data selection.
-
Improved Multi-Turn Dialogue Performance: Through a four-stage training process of continual pre-training, supervised fine-tuning, direct preference optimization, and reinforcement learning, the system significantly enhances intent understanding, rule adherence, and solution extraction capabilities in multi-turn dialogues.
-
Business Rule Adaptation and Self-Evolution: WOWService has a self-optimizing training mechanism that can quickly adapt to changes in business rules, achieve continuous evolution and optimization of the model, and reduce business expansion costs.
-
Multi-Agent Collaboration: It introduces a multi-agent architecture, in which the main agent is responsible for global dialogue control and specialized sub-agents handle specific tasks, improving service quality and flexibility in complex scenarios.
-
Automated Evaluation and Feedback: A comprehensive evaluation framework has been built, covering model basic performance and end-to-end agent effectiveness evaluation, achieving continuous model improvement through data feedback.
-
Multimodal Interaction Support: WOWService supports multimodal inputs such as speech and images. Through multimodal understanding agents, it improves the system's ability to understand non-text inputs and expands application scenarios.
Technical Principles of WOWService
-
Data-and-Knowledge Dual-Driven: WOWService integrates data-driven and knowledge-driven approaches. By combining high-quality data with structured business knowledge, it strengthens the model's ability to adhere to business rules and knowledge points, improving model performance in complex business scenarios.
-
Four-Stage Training Pipeline: It adopts a four-stage training process consisting of continual pre-training (CPT), supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning (RL), progressively improving the model's general capabilities, business adaptation capabilities, and reasoning capabilities.
-
Multi-Agent Architecture: It introduces a main agent and multiple specialized sub-agents. The main agent is responsible for global dialogue control, while the sub-agents handle specific tasks, enabling efficient handling of complex tasks through dynamic invocation and information integration.
-
Self-Optimizing Training: By automatically selecting high-quality service cases and rewriting low-quality dialogues, it builds a self-evolving data loop, enabling the model to continuously optimize and adapt to new business needs.
-
Hybrid Data-Knowledge-Driven: Combining the flexibility of data-driven approaches with the rule adaptability of knowledge-driven approaches, it enhances the model's accuracy and reliability in complex business processes through dynamic knowledge retrieval and reasoning augmentation.
-
Reinforcement Learning and Preference Optimization: By 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, improving user experience.
-
Automated Evaluation Framework: Establishing a comprehensive evaluation system from basic model performance to end-to-end agent effectiveness, supporting automated evaluation and continuous optimization to ensure the model's stability and efficiency in real-world applications.
WOWService Project Address
- arXiv Technical Paper：https://arxiv.org/pdf/2510.13291
WOWService Application Scenarios
-
Meituan Intelligent Customer Service System: WOWService has been successfully deployed in the intelligent customer service system of the Meituan App, handling user inquiries and providing 24/7 uninterrupted service, significantly improving user satisfaction and problem resolution efficiency.
-
Multi-Business Scenario Coverage: It is widely used in dozens of business scenarios within Meituan, such as food delivery, hotel booking, travel, etc., providing personalized services according to different business needs.
-
Complex Business Process Handling: In complex business processes involving multiple roles (such as users, merchants, and delivery riders), WOWService can flexibly schedule and coordinate resources to ensure efficient information transmission and service execution.
-
Multimodal Interaction Support: It supports multiple input methods such as voice and images, providing users with a more natural and convenient interaction experience, and is suitable for scenarios that require visual or auditory assistance.
-
Proactive Collaboration and Recommendations: It can proactively uncover users' potential needs and provide personalized recommendations and solutions, enhancing user engagement and satisfaction.
-
Intelligent Outbound Call Service: It implements intelligent outbound call functions, such as order confirmation and satisfaction surveys, improving operational efficiency and user response speed.