Higher Satisfaction, Lower Cost: A Technical Report on How LLMs Revolutionize Meituan's Intelligent Interaction Systems

TL;DR

WOWService presents a scalable, multi-agent framework that integrates continual pre-training, lightweight supervised fine-tuning, direct preference optimization, and reinforcement learning to enhance industrial intelligent interaction systems. By coupling data-driven and knowledge-driven methods, and deploying a robust data loopback and evaluation platform, the approach achieves significant improvements in real-world Meituan interactions, highlighting reductions in dissatisfaction and improvements in personalized service. The contributions include a four-stage training pipeline, a dual-driven data-knowledge architecture, and an end-to-end evaluation framework with actionable metrics and a dynamic leaderboard. Practically, the work demonstrates how autonomous task management, proactive collaboration, and multi-modal understanding can deliver cost-efficient, scalable, and user-centric service in large-scale local lifestyle platforms. The proposed directions point toward richer tool use via agentic RL, deeper multi-modal integration, and genuinely personalized assistants that anticipate user needs.

Abstract

Enhancing customer experience is essential for business success, particularly as service demands grow in scale and complexity. Generative artificial intelligence and Large Language Models (LLMs) have empowered intelligent interaction systems to deliver efficient, personalized, and 24/7 support. In practice, intelligent interaction systems encounter several challenges: (1) Constructing high-quality data for cold-start training is difficult, hindering self-evolution and raising labor costs. (2) Multi-turn dialogue performance remains suboptimal due to inadequate intent understanding, rule compliance, and solution extraction. (3) Frequent evolution of business rules affects system operability and transferability, constraining low-cost expansion and adaptability. (4) Reliance on a single LLM is insufficient in complex scenarios, where the absence of multi-agent frameworks and effective collaboration undermines process completeness and service quality. (5) The open-domain nature of multi-turn dialogues, lacking unified golden answers, hampers quantitative evaluation and continuous optimization. To address these challenges, we introduce WOWService, an intelligent interaction system tailored for industrial applications. With the integration of LLMs and multi-agent architectures, WOWService enables autonomous task management and collaborative problem-solving. Specifically, WOWService focuses on core modules including data construction, general capability enhancement, business scenario adaptation, multi-agent coordination, and automated evaluation. Currently, WOWService is deployed on the Meituan App, achieving significant gains in key metrics, e.g., User Satisfaction Metric 1 (USM 1) -27.53% and User Satisfaction Metric 2 (USM 2) +25.51%, demonstrating its effectiveness in capturing user needs and advancing personalized service.

Figures (14)

• Figure 1: The comparative results of WOWService and the base model across 11 key metrics, which demonstrate that WOWService achieves notable superiority in business applications.
• Figure 2: Diagram of adaptive data mixtures optimization framework in continual pre-training.
• Figure 3: Architecture of the data loopback framework.
• Figure 4: Architecture of SRT for automated iteration using model self-sampled interaction data.
• Figure 5: Schematic Diagram of the Operable DPO Framework. Operational Personnel, trainers, and automated tools each participate in various coordination aspects throughout the entire process and collaborate with one another, thereby establishing an efficient and sustainable iterative process.