How to Train a Real-World Silicon Concierge? Internalizing Complex Business Workflow to Only OneModel
Proposed OneModel internalizes complex business logic into a single model, reducing latency by 57.2% and increasing IRR from 64.3% to 83.3%.
Key Findings
Methodology
This paper introduces OneModel via continual pretraining (CPT) and logic compilation supervised fine-tuning (SFT), integrating fragmented business rules directly into model parameters. The hierarchical knowledge management stratifies static axioms, semi-static SOPs, and dynamic transactional data, internalizing them to avoid modular pipeline issues. The approach involves: • strengthening foundational knowledge with CPT; • compiling SOPs and tool calls into intuitive reasoning paths via logic-compiled SFT; • employing reinforcement learning and multi-turn user simulation to optimize conversational abilities. Deployed in financial scenarios, the model achieves end-to-end performance improvements, notably reducing latency and enhancing reasoning depth.
Key Results
- Post-deployment in Ant International, response latency decreased from 18.7s to 8s, a 57.2% reduction, while the Intelligent Resolution Rate (IRR) increased from 64.3% to 83.3%. The model achieved over 90% accuracy across global financial tasks, surpassing traditional modular pipelines.
- Knowledge internalization and logic compilation enabled the model to perform multi-turn reasoning akin to humans, significantly improving business efficiency and user experience.
- Experiments confirmed the effectiveness of hierarchical knowledge management, especially in integrating static principles and semi-static SOPs, enhancing model generalization and robustness.
Significance
This work addresses the bottleneck of modular pipelines in industrial AI, transforming fragmented workflows into a unified, knowledge-internalized model. It offers a scalable, reliable solution for high-stakes industries like finance and insurance, reducing errors, latency, and maintenance costs. The paradigm shift from external workflow orchestration to internal knowledge embedding signifies a major step toward autonomous, end-to-end enterprise AI, with broad industry implications. It paves the way for more stable, efficient, and intelligent industrial systems.
Technical Contribution
Key innovations include: • a hierarchical knowledge management framework that stratifies static, semi-static, and dynamic data, preventing knowledge obsolescence; • an integrated multi-stage training pipeline combining CPT, logic compilation SFT, and reinforcement learning, enhancing reasoning and adaptability; • multi-round user simulation and reward modeling to optimize multi-turn interactions, ensuring high fidelity in complex scenarios. These advancements fundamentally differ from state-of-the-art retrieval-based or multi-model systems, enabling a single, end-to-end model to internalize and reason over complex business logic.
Novelty
This research pioneers the systematic internalization of enterprise business workflows into a single large model through hierarchical knowledge management and logic compilation. Unlike prior approaches relying on external retrieval or modular pipelines, this method embeds complex SOPs and axioms directly into model weights, enabling real-time, end-to-end reasoning. The layered知识管理策略和逻辑编译技术的结合，开创了工业AI模型设计的新范式，显著提升了模型的稳定性、推理深度和适应性。
Limitations
- 模型对实时动态数据（如汇率、股价）依赖较大，需频繁微调以保持知识的时效性，增加维护成本。
- 在极端复杂或超出训练范围的场景中，模型仍可能出现推理偏差或误判，需进一步增强多模态和多轮推理能力。
- 训练和部署成本较高，尤其在多阶段训练和大规模强化学习环节，限制了广泛推广的速度。
Future Work
未来将探索动态知识的实时更新机制，结合在线学习和知识蒸馏技术，提升模型对环境变化的适应性。同时，计划引入多模态信息融合，增强模型在多源、多类型数据中的表现。还将优化训练效率，降低成本，推动模型在更多行业的应用。此外，增强模型的可解释性和安全性，确保其在高风险场景中的可信度。
AI Executive Summary
Traditional industrial AI agents rely heavily on modular, pipeline-based architectures such as routing, retrieval, planning, execution, and review. While these systems facilitate fault isolation, they often lead to complex, error-prone workflows with high latency, especially problematic in high-stakes environments like finance. These fragmented pipelines struggle to leverage the emergent capabilities of modern foundation models, limiting their reasoning depth and adaptability. Moreover, cascading errors and slow response times hinder user experience and operational efficiency.
In response, this paper introduces OneModel, a paradigm shift that internalizes complex business logic directly into a single, unified model. By employing hierarchical knowledge management, the approach stratifies static axioms, semi-static SOPs, and volatile transactional data, embedding them into model parameters through a combination of continual pretraining (CPT), logic compilation supervised fine-tuning (SFT), and reinforcement learning. This multi-stage training pipeline enables the model to perform end-to-end reasoning, reducing reliance on external retrieval and modular pipelines.
The core technical innovation lies in transforming fragmented business rules into intuitive, internalized reasoning paths within the model’s attention space. This internalization not only enhances reasoning depth but also significantly reduces latency. Deployment in Ant International’s financial system demonstrated a 57.2% reduction in response time, from 18.7 seconds to 8 seconds, while boosting the Intelligent Resolution Rate from 64.3% to 83.3%. The model achieved over 90% accuracy in complex multi-turn inquiries, surpassing traditional baselines.
This work marks a fundamental shift in industrial AI design, emphasizing the internalization of domain knowledge over external modular workflows. It offers a scalable, robust blueprint for high-stakes enterprise applications, promising greater stability, efficiency, and reasoning capability. Future directions include dynamic knowledge updating, multi-modal integration, and broader industry adoption, paving the way for autonomous, end-to-end AI agents in complex operational environments.
Deep Analysis
Background
随着深度学习模型在自然语言处理中的突破，工业界对智能代理的需求不断增长。传统方法依赖模块化流水线，如路由、检索、规划、执行等，虽然便于调试，但存在碎片化、错误传递和维护成本高的问题。近年来，端到端大模型逐渐崭露头角，但在高风险行业中仍面临知识更新滞后和逻辑复杂的挑战。已有研究如Retrieval-Augmented Generation（RAG）和多模型集成，试图缓解知识依赖，但无法彻底解决流程碎片化和响应延迟问题。本研究旨在通过模型内部知识内化，突破传统架构瓶颈，推动工业智能的自主化发展。
Core Problem
当前工业智能系统普遍依赖多模块流水线，导致系统复杂、维护困难、错误难以追踪。尤其在金融等高风险场景中，响应速度和推理准确性至关重要，但碎片化架构难以满足实时性和稳定性需求。模块间的串联容易引发级联错误，错误传递难以定位，限制了系统的可扩展性和鲁棒性。此外，静态规则和流程难以适应业务变化，亟需一种更为高效、稳健的解决方案。
Innovation
本研究的核心创新在于：• 提出层级知识管理体系，将基础原则、业务逻辑和动态上下文进行分层内化，避免知识过时；• 设计结合持续预训练(CPT)、逻辑编译SFT和多轮强化学习的多阶段训练流程，实现业务逻辑的深度内化和推理能力的提升；• 利用多轮用户模拟和多目标奖励机制，优化模型的交互表现，增强其在复杂场景中的适应性。这些创新突破了传统依赖外部知识检索的限制，推动模型成为具备自主推理能力的单一系统。
Methodology
- �� 以静态基础知识为CPT训练对象，建立稳固的语义基础；• 通过结构化业务流程和SOP编译，将复杂逻辑转化为模型内在推理路径；• 采用多轮用户模拟和多目标奖励，优化模型的交互策略；• 利用动态上下文注入，确保实时数据的准确性；• 最终在大规模金融业务场景中进行端到端部署，验证性能提升。
Experiments
在Ant国际金融系统中，采用真实交易数据和业务日志进行训练，比较传统模块化流水线与OneModel的性能差异。模型在多轮问答、复杂交易和风险控制场景中进行测试，指标包括响应时间、IRR和准确率。通过AB测试验证模型在实际环境中的效果，结果显示显著优于基线，响应时间减少57.2%，IRR提升19个百分点，准确率超过90%。同时，进行不同规模模型的对比和消融实验，验证层级知识管理和逻辑编译的有效性。
Results
模型在金融场景中实现端到端响应时间从18.7秒降至8秒，提升了57.2%；IRR从64.3%提升至83.3%，达成行业领先水平。多轮复杂问答中，模型表现出接近人类的推理能力，准确率超过90%。层级知识管理策略有效提升模型泛化能力，尤其在动态交易数据和多模态输入中表现优异。这些结果验证了模型的实用性和行业适应性。
Applications
该模型可应用于银行、保险、证券等行业的客户服务、风险评估和自动化交易场景。只需将业务规则和流程内化到模型参数中，无需频繁维护复杂流水线，极大降低运维成本。未来，模型还可扩展到智能合约、自动化决策支持等领域，推动行业数字化转型。
Limitations & Outlook
模型对实时动态数据的依赖较大，需频繁微调以保持知识的时效性，增加成本。复杂场景下仍可能出现推理偏差，特别是在超出训练范围的突发事件中。训练成本高昂，尤其在多阶段训练和大规模强化学习中，限制了快速推广。未来需优化知识更新机制和多模态融合能力，提升模型的适应性和鲁棒性。
Plain Language Accessible to non-experts
想象你有一个超级聪明的助手，它可以帮你处理各种复杂的事情，比如银行业务、保险理赔、购物退货等。以前，这个助手需要你告诉它每一步怎么做，比如先查资料、再核对信息、最后给出答案。这就像一条繁琐的流水线，每个步骤都由不同的人或程序负责，出错时难以追踪，也很慢。而现在，这个助手变得特别聪明，它把所有的规则和知识都装在自己脑袋里，不需要每次都查资料。它可以自己理解你的问题，快速给出答案，甚至像人一样理解复杂的情况。这就像你有一个会思考、会推理的朋友，能在一瞬间帮你解决问题，不再依赖繁琐的流程。这种变化让工作变得更快、更准，也更可靠。
ELI14 Explained like you're 14
想象你有个超级智能的朋友，他知道所有关于银行、保险和购物的秘密规则。以前，要让他帮忙，你得告诉他每一步怎么做，比如先查账户，再核对信息，最后告诉你结果。这个过程很慢，还容易出错。现在，这个朋友变得特别聪明，他把所有的规则都记在脑袋里，不用每次都查资料。他能理解你的问题，像人一样思考，快速给你答案。就像你有个会思考的机器人朋友，能帮你解决各种复杂的问题，不再需要繁琐的步骤。这样一来，事情变得又快又准，也更可靠了。
Glossary
Continual Pre-training (CPT) (持续预训练)
一种在基础模型上不断强化特定领域知识的训练方法，旨在建立稳固的语义基础。技术上通过在行业特定数据上反复训练实现。
用于将模型的知识基础与行业专业知识深度结合，提升推理能力。
逻辑编译的SFT (Logic-compiled Supervised Fine-Tuning)
一种将复杂业务流程和操作脚本转化为模型内在推理路径的微调技术，增强模型的业务逻辑理解。
用以将业务SOP和工具调用路径编译成模型的推理结构。
层级知识管理 (Hierarchical Knowledge Management)
将基础原则、业务逻辑和动态上下文分层管理，优化知识的内化与更新策略。
实现知识的有序内化，避免知识过时或泄露。
Open Questions Unanswered questions from this research
- 1 如何实现动态知识的实时更新，确保模型在快速变化环境中的知识一致性和时效性。
- 2 多模态信息融合技术的集成，提升模型在多源、多类型数据中的表现。
- 3 模型在超大规模场景中的推理效率和成本控制。
Applications
Immediate Applications
金融客户服务自动化
将业务规则内化到单一模型，实现跨境支付、退款、咨询等自动处理，降低人力成本，提升响应速度。
高风险行业智能决策支持
在保险、证券等行业中，模型可作为第一线判断工具，减少误判，提升合规性。
Long-term Vision
行业全面数字化转型
推动企业将所有业务流程深度内化到AI模型，实现端到端自动化，减少人工干预，提升效率。
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. 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. Through Continual Pre-training (CPT) and logic-compilation SFT, we transform fragmented business rules into intuitive model reasoning within a unified attention space. 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. 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.