钉钉AI实现国际顶级基准、实际应用落地的双重突破

Core Insights - Dingtalk-DeepResearch, developed by the DingTalk team, achieved a significant milestone by ranking second globally and first domestically in the DeepResearch Bench test with a score of 48.49, surpassing major systems like OpenAI and Claude [1][3] - The system has been successfully applied in complex scenarios such as manufacturing and supply chain, demonstrating industry-leading accuracy and robustness in handling multi-modal data, thus facilitating intelligent upgrades for enterprises [3][7] System Design and Functionality - The core of Dingtalk-DeepResearch is a multi-agent deep research framework designed for real enterprise scenarios, effectively integrating deep research generation, heterogeneous table parsing, reasoning, and multi-modal report generation within a single system [4] - The system employs a three-layer architecture (task-oriented agent layer, core engine layer, data layer) to support parallel processing of complex tasks and multi-stage reasoning, enabling automatic parsing of complex factory production tables into clear and insightful analysis reports [4] Continuous Learning and Evolution - Unlike traditional static architectures, the framework features an online learning mechanism guided by entropy and memory awareness, allowing agents to continuously evolve without manual intervention, akin to employees improving skills through practice [5] - The system can autonomously learn and remember user preferences regarding report formats and styles, aligning future outputs with user needs, thereby facilitating knowledge reuse and efficiency across teams and the entire company [5] Quality Assurance Mechanism - To ensure the accuracy and reliability of generated content, Dingtalk-DeepResearch includes the DingAutoEvaluator assessment system, which conducts multi-dimensional quality checks on each report, covering data accuracy, logical coherence, and tool usage standards [6] - Any identified issues are automatically fed back into the training process to optimize the model, creating a continuous improvement loop from generation to evaluation to optimization [6] Practical Applications and Value Creation - Dingtalk-DeepResearch has been stably applied in various real business scenarios, providing value by quickly analyzing complex table data across departments in the supply chain and converting raw equipment operation data into visual analysis reports for predictive maintenance in manufacturing [7] - The CTO of DingTalk, Zhu Hong, stated that Dingtalk-DeepResearch combines adaptive optimization and multi-modal reasoning to form a flexible enterprise-level AI framework aimed at addressing complex and evolving real business tasks, accelerating the deployment of cutting-edge AI technologies in practical production needs [7]