ArXiv Introduces OneModel: Revolutionizing Industrial Agent Architecture for End-to-End Efficiency

Summary:ArXiv introduces OneModel, a novel approach that revolutionizes traditional industrial agent architecture by consolidating complex business logic and standard operating procedures (SOPs) directly into model parameters, replacing the conventional modular pipeline. This shift results in a significant reduction in end-to-end latency and an improvement in the Intelligent Resolution Rate (IRR) from 64.3% to 83.3%, as demonstrated by online A/B testing. The latency drops from 18.7 seconds to 8.0 secon

---

Key Breakthroughs

OneModel introduces a paradigm shift in industrial agent architecture through the following innovations:

• Knowledge Internalization: Consolidates complex business logic and standard operating procedures (SOPs) directly into model parameters, eliminating the need for external modular workflows.
• Continual Pre-training (CPT) and Logic-Compilation SFT: These techniques transform fragmented business rules into intuitive model reasoning within a unified attention space.
• Unified Attention Space: Enables efficient knowledge representation and reasoning in a single attention space.

Technical Highlights

• Performance Improvement: Online A/B testing demonstrates a more than 50% reduction in end-to-end latency, from 18.7 seconds to 8.0 seconds.
• Intelligent Resolution Rate (IRR) Enhancement: The IRR increases from 64.3% to 83.3%, indicating a significant improvement in the model's decision-making capabilities for complex tasks.
• Scalability: The method provides a scalable solution for transitioning from complex, error-prone workflows to a unified model architecture.

Industry Impact

OneModel represents a significant advancement in the field of industrial agents, with implications for:

• Efficiency: By reducing latency and increasing IRR, OneModel can significantly enhance the efficiency and accuracy of industrial processes.
• Simplified Architecture: Replacing traditional modular workflows simplifies system architecture and reduces maintenance costs.
• Wide Application: Suitable for industries such as finance, manufacturing, and supply chain management where efficient decision-making and low latency are crucial.

Developer Recommendations

• Focus on Continual Pre-training (CPT): Developers should explore CPT to leverage its advantages in knowledge internalization.
• Investigate Logic-Compilation SFT: This technique is crucial for transforming complex business rules into model reasoning.
• Evaluate System Compatibility: When integrating OneModel into existing systems, assess its impact on current workflows and conduct thorough testing.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-08-24)

— END —

Tags: #OneModel #Industrial Agents #Continual Pre-training #Logic-Compilation SFT #Unified Attention Space