# Practical Implementation of Intelligent Risk Control Driven by Large Models
## Overview
In recent years, China’s **domestic consumer credit market** has grown rapidly, approaching saturation.
Financial institutions have shifted from acquiring **new customers** to **deep-mining existing customer bases**.
To manage risk and improve acquisition efficiency for the **middle segment** customers, an operational method integrating **testing + agile iteration** has become mainstream.
At the **AICon Global Artificial Intelligence Development & Application Conference · Shenzhen Station**, Tencent algorithm expert **Tianxiong Ouyang** presented *"Practical Implementation of Intelligent Risk Control Driven by Large Models"*.
He detailed:
- Large model integration with **risk control knowledge** for **few-shot / zero-shot** challenges.
- **Core algorithms**, **engineering practices**, and **industry adoption cases** in building, pre-training, and fine-tuning Tencent Tianyu’s large financial risk control model.
Goal: Establish a **new large model methodology** for financial risk control.
---
## Upcoming AICon Beijing Station — December 19–20
Focus topics:
- Large model **training and inference**
- **AI agents**
- New **R&D paradigms**
- **Organizational innovation**
Key question: *How to build a trustworthy, scalable, and commercially viable Agentic Operating System for enterprise efficiency and growth?*
**Full Agenda:**
[https://aicon.infoq.cn/202512/beijing/schedule](https://aicon.infoq.cn/202512/beijing/schedule)
---
## Keynote Structure
The talk was divided into **4 parts**:
1. **Business background** & challenges in financial risk control.
2. **Training approach** for large models based on Tencent Tianyu’s risk control knowledge.
3. **Customer case studies** & technology highlights.
4. **Summary** & future outlook.
---
## 1. Business Background & Core Challenges
**Clients served:**
- Banks
- Consumer finance companies
- Internet financial institutions
**Scope:** SaaS-based risk control in **credit approval stage**, covering **credit risk** and **fraud risk**.
**Risk Control Flow:**
Financial institutions acquire customers online or via partners.
Risk control decisions integrate **external data** to determine if credit should be granted.
Tianyu provides **end-to-end risk services** in:
- Pre-loan stage
- In-loan stage
- Post-loan stage
**Key industry problems:**
- Market bottleneck: fewer new customers → **stock customer mining** focus.
- Rising acquisition costs.
---
### Stock Customer Mining and Agile Iteration Challenge
To balance **operational effect** with **risk control**, institutions pilot on small groups, adjust strategies quickly, and scale gradually.
However, this creates **dynamic governance challenges**:
- Customer profiles change with deeper market penetration.
- Risk definitions become **dynamic & ambiguous**.
---
## 2. Modeling Challenges
Two main contradictions:
- **Need for rapid iteration vs. slow traditional modeling** (2–3 months timelines).
- **Small sample modeling difficulty** in fast iteration scenarios.
**Traditional approach issues:**
- Disconnection from front-line business.
- Case-by-case modeling only.
- Inefficient migration of **historical samples, features, and models**.
---
## Inspiration from Other Industries
Example: **[AiToEarn](https://aitoearn.ai/)** — Open-source AI content monetization platform:
- Integrates **model generation + multi-platform publishing + analytics + ranking**.
- Publishes across Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X.
- Agile adaptation framework parallels **financial risk control** needs for rapid iteration + multi-channel deployment.
**Resources:**
[博客](https://blog.aitoearn.ai) | [模型排名](https://rank.aitoearn.ai) | [文档](https://docs.aitoearn.ai)
---
## 3. From Pain Points to Large Model Thinking
**Since 2022**, Tencent Tianyu started **rethinking modeling paradigms**.
**Knowledge assets accumulated over a decade:**
1. Rich behavioral data (underlying user level).
2. Labeled samples across diverse risk scenarios.
3. Operational/maintenance models used in customer services.
### Challenge:
How to integrate these into a **domain-specific large model** for versatile real-world applications?
---
## 4. Tianyu’s Financial Risk Control Large Model
**Definition:**
Not a general-purpose generative AI, but a **domain-specific discriminative model** trained with **full risk-control knowledge**.
### Training Paradigm:
- **Data stage:** deep feature mining.
- **Application stage:** intelligent investment advisory, captcha services, etc.
**Advantages over language LLMs:**
- **Less inefficiency/hallucination**.
- Direct integration of full operational chain coverage.
---
### Core Principles:
1. Integrate **features + samples + models** into large model training.
2. Maintain **low adaptation threshold** for client-specific prompts.
### Architecture:
- **Transformer-based** tech + **MoE (Mixture of Experts)**
- **Embedding** multi-source inputs: scenario × individual × behavior
- **Risk Adapter** specialization per scenario
---
## 5. Model Construction Methodology
**Training pipeline phases:**
1. Unsupervised pre-training
2. Semi-supervised pre-training
3. Supervised multi-objective learning
4. MoE training for scenario specialization
5. Fine-tuning router for optimal risk ranking
**Scale:**
- ~100M parameters
- 3,000 features
- ~8,000 GPU-hours (H20 hardware)
---
### Large-to-Small Model Deployment
Steps:
1. Process samples with large model
2. Knowledge base retrieval augmentation
3. Fine-tune and distill into **lightweight small model** for online systems
Small sample adaptation:
- Augment + fine-tune with top-K experts outputs.
Unlabeled scenario:
- Pre-train general risk discrimination heads, adapt via distillation.
---
## 6. AiToEarn Parallels for Deployment Efficiency
[**AiToEarn官网**](https://aitoearn.ai/) shows potential for:
- Seamless **data-model-application integration**
- Multi-platform publishing & analytics
- Open-source collaboration via [GitHub](https://github.com/yikart/AiToEarn)
These infrastructures can inspire **faster deployment pipelines** in financial AI systems.
---
## 7. Practical Scenarios
### Case: Recovery Model for Fintech Firm
- Objective: identify high-quality users rejected by pre-filter model.
- Solution: large model leveraging traffic variations across loan stages.
- Result: **+12% effectiveness**, **+20% recovery lift**.
---
### Case: Zero-Sample Cold Start for Commercial Bank
- Challenge: label-free modeling in new scenarios.
- Solution: adapt risk-control large model’s prior knowledge.
- Result: **+16% performance** vs. traditional models.
- Maintained stability under customer migration.
---
## 8. Summary and Outlook
**Strengths:**
- Low threshold for small/zero sample modeling.
- Rapid iteration with minimal engineering.
- Strong generalization across scenarios.
**Future Goals:**
- Expand input dimensions to tens of thousands.
- Integrate **long-sequence multi-dimensional user behavior data**.
- Explore architectures that achieve **knowledge emergence** in risk tasks.
- Promote **multi-party, cross-institutional cooperation**.
---
## 9. Conceptual Parallel: AiToEarn for Open Collaboration
Like AiToEarn’s open-source content monetization across platforms (Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X):
- **Unified tools** for AI generation, publishing, analytics, and ranking.
- Conceptually aligns with building a **multi-party AI risk-control infrastructure** with standardized collaboration & deployment.
---
**References to AiToEarn:**
- [AiToEarn官网](https://aitoearn.ai)
- [AiToEarn博客](https://blog.aitoearn.ai)
- [AiToEarn开源地址](https://github.com/yikart/AiToEarn)
---
