The Meituan LongCat Interaction team released the WOWService large model interaction system white paper, revealing technical details that have been fully launched in Meituan's intelligent customer service: through "data + knowledge dual-driven" and a four-stage training system, the customer service resolution rate improved by 9% and user satisfaction increased by 12% in complex business scenarios, with training and annotation volume being only 10% of traditional solutions.

Core Framework  

1. Data and Knowledge Dual-Driven: Joint training with structured business rules and real conversation logs, achieving an accuracy of 96% on knowledge points such as SKU, promotions, and after-sales service.  

2. Multi-Agent Collaboration: The main Agent is responsible for task distribution, while sub-Agents specialize in scenarios like refunds, address changes, and invoices, reducing average response time by 27%  

3. Self-Optimization Loop: High-rated online conversations are extracted daily, automatically annotated, and fed back into training, achieving "a small iteration every 7 days."

Four-Stage Training Process  

- Continuous Pre-training: 50 billion token domain-specific corpus, enabling the base model to be familiar with local life terminology  

- Supervised Fine-tuning: 10% annotated data achieves the effect of traditional 100% annotation, saving 75% cost  

- Direct Preference Optimization (DPO): Using positive and negative sample pairs to calibrate the style of "politeness + efficiency"  

- Reinforcement Learning (RL): Online real-time rewards = resolution rate + user rating, allowing the model to automatically align with commercial goals.

Business Outcomes  

The white paper shows that WOWService has been implemented across six business lines at Meituan, including takeout, in-store, hotel, and travel. During peak promotion periods, it can handle more than 8,000 QPS, saving 18% in overall customer service manpower, and achieving an 84% first-time resolution rate in complex scenarios, significantly better than the original base model.

Open Source and Future Plans  

The team plans to open-source the lightweight version WOWService-Lite (<7B parameters) and the multi-agent framework in Q1 2026 for community secondary development; meanwhile, they will collaborate with the Chinese Computer Federation to release the "Local Life Large Model Benchmark," promoting industry standardization.