DingTalk AI Achieves Dual Breakthroughs: Reaching a World-Class Benchmark and Real-World Application; Its System Capabilities Rank Second Globally and First in China!

Recently, DingTalk-DeepResearch, a deep research system developed by the DingTalk team, achieved a breakthrough in an international authoritative benchmark. In the DeepResearch Bench test, it scored 48.49 points, ranking second globally and first in China, surpassing mainstream systems such as OpenAI and Claude.

Real-world Applications Across Multiple Scenarios: Intelligent Handling of Complex Tasks

The system has been successfully applied in complex scenarios such as manufacturing and supply chain management. It maintains industry-leading accuracy and robustness in tasks involving complex heterogeneous tables, multi-stage reasoning, and multimodal generation, helping enterprises efficiently process multimodal data and achieve intelligent upgrades.

This advancement in DingTalk's deep research system marks the first dual breakthrough—achieving top-tier international benchmarks while also enabling real-world production deployment—signaling that China's enterprise-grade AI technology has entered the global first tier.

Multi-Agent Architecture: Supporting Deep Collaborative Research

The core of Dingtalk-DeepResearch is a multi-agent deep research framework designed for real-world enterprise scenarios. It effectively integrates deep research generation, heterogeneous table parsing and reasoning, and multimodal report generation into a single system.

This design is akin to bringing together team members with different specialized skills into one system: some excel at analyzing tabular data, others focus on report generation, and still others coordinate tool calls. Through a three-layer architecture—the task-oriented agent layer, the core engine layer, and the data layer—the system supports parallel processing and multi-stage reasoning for complex tasks. For example, it can automatically parse factory production tables containing multiple nested and merged cells and transform them into clear, insightful analysis reports.

Ongoing Evolution Mechanism: Enabling Adaptive Learning

To handle dynamic changes in enterprise scenarios, the framework employs an entropy-guided, memory-aware online learning mechanism that allows agents to continuously evolve—much like employees who improve their skills through repeated practice without manual intervention. This mechanism ensures that the system can automatically extract insights from historical interactions, gradually adapting to different enterprises' business processes and users' operational styles.

For instance, when a user repeatedly modifies the format of an AI-generated report, the system autonomously learns and remembers the user's preferences regarding format, style, and key points, proactively aligning future outputs with the user's needs. On DingTalk's enterprise AI platform, these personalized preferences can be distilled into capabilities and shared across teams or even the entire company, enabling the reuse and enhancement of organizational knowledge.

Closed-loop Evaluation System: Ensuring Reliable Output Quality

To ensure the accuracy and reliability of generated content, Dingtalk-DeepResearch incorporates the DingAutoEvaluator evaluation system. This system conducts multidimensional "quality checks" on every report generated, covering data accuracy, logical consistency, and adherence to tool usage standards. If issues are detected, the system automatically feeds relevant cases back into the training process to refine the model, creating a continuous improvement loop that spans generation, evaluation, and optimization.

Stable Deployment Across Multiple Industries: Delivering Tangible Value

Currently, Dingtalk-DeepResearch is stably deployed in several real-world business scenarios, delivering tangible value. In the supply chain domain, the system can quickly analyze complex cross-departmental table data, providing intelligent recommendations for procurement strategies. In manufacturing, it can automatically convert raw equipment operation data into visualized analysis reports, supporting decision-making for fault prediction and maintenance. All core functions have been validated through international benchmark tests, ensuring the technology's reliability and leading-edge status.

Zhu Hong, DingTalk's CTO, stated, "By combining adaptive optimization with multimodal reasoning, Dingtalk-DeepResearch forms a flexible, deployable enterprise-grade AI framework designed to handle complex and ever-evolving real-world business tasks. This technology is accelerating its integration into products such as AI search, AI tables, automated workflows, and agent platforms, bringing cutting-edge AI closer to actual production needs and delivering AI solutions that truly create value for enterprises."

DomTech is DingTalk's officially designated service provider in Macau, specializing in providing DingTalk services to a wide range of customers. If you'd like to learn more about DingTalk platform applications, feel free to contact our online customer service, call +852 95970612, or email cs@dingtalk-macau.com. We have an outstanding development and operations team with extensive market service experience, ready to provide you with professional DingTalk solutions and services!