WebWeaver – Alibaba Tongyi’s Open-Source Dual-Agent Framework

What is WebWeaver？

WebWeaver is a novel dual-agent framework developed by Alibaba’s Tongyi team, part of the Tongyi DeepResearch family, designed for open-ended deep research. WebWeaver simulates the human research process by splitting tasks between two agents: a Planning Agent (responsible for exploration and outline generation) and a Writing Agent (responsible for content synthesis). Its core innovation lies in dynamic outline optimization, treating the research outline as a “living document” that iteratively cycles between searching and refining, allowing research directions to evolve dynamically with new discoveries. Using a hierarchical memory-based synthesis approach, WebWeaver constructs reports section by section, ensuring coherence, accuracy, and depth grounded in sources. In open-ended deep research benchmarks, WebWeaver achieves state-of-the-art performance and has created the WebWeaver-3k dataset, enabling smaller models to attain expert-level research capabilities.

Key Features of WebWeaver

• Efficient Information Retrieval and Integration: Rapidly retrieves relevant information from vast web sources and organizes it into a logical structure, providing users with a comprehensive and accurate knowledge base.

• Dynamic Outline Optimization: Treats research outlines as “living documents,” continuously adjusting and refining them based on new findings, allowing research directions to adapt flexibly and avoid fixed thinking patterns.

• Hierarchical Report Synthesis: Constructs reports in a hierarchical, section-by-section manner. Each section retrieves the most relevant evidence from the memory bank, ensuring coherence and accuracy while avoiding information loss common in long-text generation.

• Enhancing Small Model Capabilities: Using the WebWeaver-3k dataset, complex research skills are transferred to smaller models, enabling them to achieve expert-level research performance and lowering the barrier to high-quality research.

Technical Principles of WebWeaver

• Dual-Agent Framework: Comprises a Planning Agent for exploration and outline generation and a Writing Agent for report synthesis. This division of labor simulates human research processes, improving efficiency and quality.

• Dynamic Iterative Mechanism: The Planning Agent continuously performs web searches during research, comparing results with the outline and optimizing it in a dynamic loop. This enables timely integration of new information, avoiding the rigidity of traditional static outlines.

• Memory Bank Utilization: The Writing Agent retrieves the most relevant evidence from a curated memory bank during report generation. The memory bank stores preprocessed and annotated information retrieved from the web.

• Reinforcement Learning and Optimization: WebWeaver continuously optimizes search and generation strategies through reinforcement learning. Feedback signals guide behavioral adjustments, improving search efficiency and report quality. Its adaptive optimization mechanism enables it to handle diverse research topics and tasks effectively.

Project Links

• GitHub Repository: https://github.com/Alibaba-NLP/DeepResearch/tree/main/WebAgent/WebWeaver

• arXiv Paper: https://arxiv.org/pdf/2509.13312

Application Scenarios of WebWeaver

• Academic Research: Quickly integrates literature, generates reviews and draft papers, supporting efficient scholarly research.

• Enterprise Decision Support: Helps companies gather market and industry information for market research, strategic planning, and investment decision-making.

• Education: Assists teachers in sourcing teaching materials and course design; aids students in learning and capstone projects, enhancing educational outcomes.

• Media and Journalism: Enables journalists and media personnel to rapidly gather news background and expert insights, improving reporting quality and feature planning.

• Government and Public Policy: Assists government agencies in collecting socio-economic information, supporting policy formulation, public affairs management, and emergency response with data-driven insights.