How to Train a Real-World Silicon Concierge? Internalizing Complex Business Workflow to Only OneModel

如何训练现实世界的“硅基礼宾员”？将复杂业务流程内化为单一模型 (OneModel)

Abstract: Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer, and other components. These systems often fracture into a labyrinth of ad-hoc patches, leading to cascading errors and high latency.

摘要： 传统的工业智能体依赖于模块化流水线，包括路由（Router）、检索（Retriever）、规划（Planner）、执行（Executor）、响应（Responder）、审查（Reviewer）等组件。这些系统往往会演变成错综复杂的临时补丁迷宫，导致级联错误和高延迟。

We propose OneModel, an applicable paradigm shift from external workflows to internalized knowledge representation. Unlike modular systems that slice fluid user intents into static steps, OneModel consolidates complex business logic and SOPs directly into the model parameters.

我们提出了 OneModel，这是一种从外部工作流转向内化知识表示的可行范式转变。与将流畅的用户意图切割成静态步骤的模块化系统不同，OneModel 将复杂的业务逻辑和标准作业程序（SOP）直接整合到模型参数中。

Through Continual Pre-training (CPT) and logic-compilation SFT, we transform fragmented business rules into intuitive model reasoning within a unified attention space.

通过持续预训练（CPT）和逻辑编译监督微调（SFT），我们将碎片化的业务规则转化为统一注意力空间内直观的模型推理。

Deployed in our global financial service system, OneModel effectively breaks the trade-off between latency, accuracy, and complexity. Online A/B testing demonstrates an end-to-end latency reduction of more than 50 percent, from 18.7 seconds to 8.0 seconds, while the Intelligent Resolution Rate (IRR) increases from 64.3 percent to 83.3 percent.

OneModel 已部署在我们的全球金融服务系统中，有效打破了延迟、准确性和复杂性之间的权衡困境。在线 A/B 测试表明，端到端延迟降低了超过 50%，从 18.7 秒缩短至 8.0 秒，同时智能解决率（IRR）从 64.3% 提升至 83.3%。

The results show that OneModel can replace brittle engineering logic with internalized cognitive intuition, offering a scalable blueprint for transitioning industrial agents from complex, error-prone workflows to unified model architectures.

研究结果表明，OneModel 可以用内化的认知直觉取代脆弱的工程逻辑，为工业智能体从复杂、易错的工作流向统一模型架构的转型提供了一个可扩展的蓝图。