TLDR: Meituan’s LongCat Interaction Team developed WOWService, an intelligent interaction system leveraging Large Language Models (LLMs) and multi-agent architectures. It addresses common challenges in customer service like data quality, multi-turn dialogue, and adaptability. Through a multi-stage training pipeline (Continual Pre-Training, Supervised Fine-Tuning, Preference Learning, Reasoning Enhancement) and specialized agents, WOWService significantly improves user satisfaction and reduces operational costs, as demonstrated by its deployment on the Meituan App.
In today’s fast-paced digital world, enhancing customer experience is paramount for business success. As service demands grow in both scale and complexity, companies are increasingly turning to advanced technologies to provide efficient, personalized, and round-the-clock support. Generative artificial intelligence (AI) and Large Language Models (LLMs) are at the forefront of this transformation, empowering intelligent interaction systems to meet these evolving needs.
However, developing and deploying such systems comes with its own set of challenges. These include the difficulty of creating high-quality data for initial training, which can hinder a system’s ability to learn and adapt while increasing labor costs. Multi-turn dialogues often suffer from poor intent understanding and rule compliance, leading to suboptimal performance. Rapid changes in business rules can also affect a system’s flexibility and ability to expand at a low cost. Furthermore, relying on a single LLM can be insufficient for complex scenarios, highlighting the need for multi-agent frameworks that enable effective collaboration. Lastly, the open-ended nature of multi-turn dialogues makes quantitative evaluation and continuous improvement difficult due to the lack of unified ‘golden answers’.
Introducing WOWService: Meituan’s Solution for Intelligent Interactions
To tackle these multifaceted challenges, the LongCat Interaction Team at Meituan has introduced WOWService, an intelligent interaction system specifically designed for industrial applications. By integrating LLMs with sophisticated multi-agent architectures, WOWService enables autonomous task management and collaborative problem-solving. The system focuses on key areas such as data construction, enhancing general capabilities, adapting to specific business scenarios, coordinating multiple AI agents, and automating evaluation processes.
WOWService has already been successfully deployed on the Meituan App, demonstrating significant improvements in key performance indicators. For instance, User Satisfaction Metric 1 (USM 1) saw a 27.53% reduction, indicating improved efficiency, while User Satisfaction Metric 2 (USM 2) increased by 25.51%, reflecting enhanced personalized service. These results underscore WOWService’s effectiveness in understanding user needs and delivering a superior customer experience.
The Technical Backbone: A Multi-Stage Training Pipeline
WOWService’s robust performance is built upon a comprehensive training pipeline that continuously strengthens the LLMs’ domain-specific skills. This pipeline consists of four main stages:
- Continual Pre-Training (CPT): This stage focuses on building a strong foundational model by incorporating large-scale dialogue interaction data. It addresses the challenge of balancing general language understanding with domain-specific expertise, ensuring that the model retains its broad capabilities while excelling in specialized interactions.
- Supervised Fine-Tuning (SFT): Here, the foundational model is aligned with specific domain knowledge and intelligent interaction styles. A key innovation is a lightweight SFT paradigm that prioritizes data quality over quantity, significantly reducing training costs and accelerating iteration cycles. This stage also unifies dialogue generation and tool command execution into a single step, reducing latency and ensuring consistency.
- Preference Learning (DPO): This stage is crucial for aligning the models with human-friendly outputs. Using Direct Preference Optimization (DPO), the system learns from comparative human feedback, where responses are ranked as ‘preferred’ or ‘dispreferred’. This helps the model generate more acceptable and empathetic responses, addressing common issues like unclarified user demands, redundant inquiries, hallucinations, and incorrect solutions.
- Reasoning Enhancement: This final stage focuses on improving both the model’s general conversational ability and its adherence to knowledge-based rules. It involves enhancing multi-constraint knowledge following and reinforcing reasoning through a hybrid data- and knowledge-driven approach. Additionally, it aims to make dialogues more human-like and emotionally intelligent, using a refined Reward Model and response rewriting techniques to improve quality and emotional expressiveness.
Beyond a Single Brain: The Power of Multi-Agent Architecture
Recognizing that a single AI agent often falls short in complex scenarios, WOWService employs a sophisticated multi-agent architecture. This design allows specialized sub-agents to handle distinct functions and collaborate effectively, preventing overload on any individual system. The primary agent maintains conversational control while dynamically invoking specialized agents as needed, treating them as executable tools to gather information or handle specific parts of a request.
Key specialized agents include:
- Outbound-Call Agent: This agent manages intelligent outbound calling, from parsing requests and placing calls to retrieving outcomes and reporting results to the user. Its introduction significantly improved the correctness of information conveyed, adherence to procedures, and understanding of user intent.
- Agent of Proactive Collaboration: This agent enhances user engagement by proactively identifying and confirming potential user needs at the start of a conversation. It adapts across multiple service scenarios, seamlessly switching between them based on user replies and relevant operational instructions.
- Agent of Multi-Modal Understanding: This agent is vital for interpreting non-textual user inputs, such as speech, images, and videos. By accurately recognizing these diverse inputs, it assists the primary dialogue model in planning subsequent steps, expanding the system’s ability to handle a broader range of complex service scenarios.
Measuring Success: A Robust Evaluation Framework
To ensure continuous improvement and maintain high service quality, Meituan has developed a comprehensive evaluation framework for intelligent interaction systems. This framework supports an end-to-end workflow, incorporating both manual and automated evaluation methods. It periodically assesses state-of-the-art models and AI agents, publishing results on a dynamic leaderboard to guide algorithmic development.
The evaluation process includes constructing diverse evaluation sets based on real-world scenarios, designing scenario-specific metrics, and executing evaluations through human assessors and dedicated model-based automated scoring systems. This rigorous approach ensures that WOWService’s performance is consistently monitored and optimized.
Also Read:
- A New Framework for AI-Powered Customer Support Efficiency
- Boosting E-commerce Search Accuracy: A Framework for Transferring LLM Reasoning to Lightweight Models
The Future of Intelligent Interactions
The LongCat Interaction Team envisions intelligent interaction systems evolving into truly intelligent, proactive, and user-centric assistants. Building on the foundation of WOWService, future work will focus on empowering tool use through agentic reinforcement learning, advancing multi-agent collaboration and multi-modal integration (including text, speech, image, and video understanding), and creating deeply personalized, user-centric assistants that anticipate user needs and act as trusted partners in daily life.
This innovative approach by Meituan is not just about improving efficiency; it’s about fundamentally transforming how users interact with digital services, leading to higher satisfaction and lower operational costs. You can read the full technical report here.
