Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dingtalk DeepResearch: A Unified Multi Agent Framework for Adaptive Intelligence in Enterprise Environments

Mengyuan Chen, Chengjun Dai, Xinyang Dong, Chengzhe Feng, Kewei Fu, Jianshe Li, Zhihan Peng, Yongqi Tong, Junshao Zhang, Hong Zhu

TL;DR

The paper tackles the challenge of enterprise-scale information synthesis that requires deep research, cross-document reasoning, and multimodal report generation. It proposes a unified, three-layer, multi-agent framework (Agent Studio, Core, Data Layer) with DingAutoEvaluator to enable evaluation-driven continuous optimization and secure workflow integration in real-world settings. A central contribution is the large-scale documentary reinforcement learning pipeline (Doc-RM, SFT for structured formats, static and live RL, online DPO) plus a structure-aware tabular reasoning engine (NL2SQL with robust retrieval) guided by an automated online evaluation flywheel. The framework demonstrates practical impact through production deployment, end-to-end showcases, and forthcoming service availability, enabling adaptive, multi-modal enterprise intelligence within the Dingtalk ecosystem.

Abstract

We present Dingtalk DeepResearch, a unified multi agent intelligence framework for real world enterprise environments, delivering deep research, heterogeneous table reasoning, and multimodal report generation.

Dingtalk DeepResearch: A Unified Multi Agent Framework for Adaptive Intelligence in Enterprise Environments

TL;DR

The paper tackles the challenge of enterprise-scale information synthesis that requires deep research, cross-document reasoning, and multimodal report generation. It proposes a unified, three-layer, multi-agent framework (Agent Studio, Core, Data Layer) with DingAutoEvaluator to enable evaluation-driven continuous optimization and secure workflow integration in real-world settings. A central contribution is the large-scale documentary reinforcement learning pipeline (Doc-RM, SFT for structured formats, static and live RL, online DPO) plus a structure-aware tabular reasoning engine (NL2SQL with robust retrieval) guided by an automated online evaluation flywheel. The framework demonstrates practical impact through production deployment, end-to-end showcases, and forthcoming service availability, enabling adaptive, multi-modal enterprise intelligence within the Dingtalk ecosystem.

Abstract

We present Dingtalk DeepResearch, a unified multi agent intelligence framework for real world enterprise environments, delivering deep research, heterogeneous table reasoning, and multimodal report generation.

Paper Structure

This paper contains 18 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Large-scale and Multi-stage Documentary Reinforcement Learning on Both Static and Live Queries
  3. Adaptive Online Learning via Entropy-Guided Memory Retrieval for Planning and Self-Evolving
  4. Structure-Aware Heterogeneous Table Parsing, Retrieval and Reasoning Verified in Enterprise Scenarios
  5. Data Ingestion.
  6. Structural Parsing.
  7. Semantic Understanding.
  8. Tabular Reasoning.
  9. Table Retrieval.
  10. DingAutoEvaluator: Automated Online Evaluation for Data Flywheel and Continuous Optimization
  11. Showcases
  12. Complicated Tabular Parsing, Retrieval and Reasoning
  13. Case A
  14. Case B
  15. Semantically Aligned Vision–Language Fusion for Multimodal Document Generation
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Dingtalk-DeepResearch's performance on Deep Research Benchmark du2025deepresearchbenchcomprehensivebenchmark and Researcher Bench xu2025researcherbenchevaluatingdeepai.
  • Figure 2: The Dingtalk-DeepResearch framework is a multi‑agent architecture for advanced real‑world problem solving, comprising: (1) Dingtalk-DeepResearch Agent Studio – professional agents for deep research, tabular processing, and data analytics, alongside customizable personal agents; (2) Dingtalk-DeepResearch Core – featuring context compression, reasoning & planning, long/short‑term memory, human‑in‑the‑loop control, a self‑evolution engine, and integrated tools for code execution, web search, file and tabular retrieval, multimodal processing, and enterprise ecosystem connectivity, including automatic linkage to relevant files, messages, and tasks within Dingtalk domains; when user‑granted permissions are enabled, the system can also securely connect to personal work documents and related resources; powered by LLMs with CPT, SFT, and RL training; (3) Dingtalk-DeepResearch Data Layer – a unified data backbone encompassing knowledge graphs, databases, caches, and multimodal datasets (dialogue, audio, image, video, graph, text, tabular) across business, industry, personal, and synthetic sources, enabling intelligent correlation and retrieval of diverse corporate and sector‑specific data.
  • Figure 3: A real-world example form primarily containing inventory, shipping plans, and transportation information for a specific material (P/N: C01, Painted Upper Back Cover, typically used for electronic device casings). The note explains that due to the uncertainty of booking times, there may be a delay of 1-3 weeks between the pickup date and the actual sailing date. Therefore, accurate warehouse arrival times are crucial to ensure a smooth supply chain.
  • Figure 4: This table presents part of a 1,200‑row weekly production record from a real manufacturing plant case, one of eight similar table files in total. MY and HLD refer to specific product models, and the summary rows at the bottom show the total weekly output for all products.