WebWeaver - A dual-agent framework open-sourced by Alibaba.

What is WebWeaver?

WebWeaver, a novel dual-agent framework developed by Alibaba's Tongyi team and belonging to the Tongyi DeepResearch family, is used for open-ended deep research. WebWeaver simulates the human research process, dividing tasks into two agents: planning (exploration and outline generation) and writing (content synthesis). WebWeaver's core innovation lies in dynamic outline optimization, treating the research outline as a "living document" that cycles between searching and refining, allowing the research direction to dynamically evolve with new discoveries. WebWeaver utilizes a memory-based hierarchical synthesis method, building reports section by section to ensure coherence, accuracy, and deep source reliance. WebWeaver has achieved new peak performance in open-ended deep research benchmarks, creating the WebWeaver-3k dataset, enabling small models to achieve expert-level research capabilities.

WebWeaver's main functions

• Efficient information retrieval and integrationWebWeaver can quickly retrieve information related to a research topic from a massive amount of web pages and integrate it into a logical structure, providing users with a comprehensive and accurate knowledge system.
• Dynamic outline optimizationTreat the research outline as a "living document," continuously adjusting and optimizing it based on new discoveries during the research process. This allows the research direction to flexibly adapt to new situations and avoids falling into fixed thinking.
• Layered Synthesis ReportWebWeaver builds reports using a layered composition approach, generating content section by section. Each part retrieves the most relevant evidence from the report memory, ensuring the report's coherence and accuracy and avoiding information loss issues in long text generation.
• Enhance small model capabilitiesBy using the WebWeaver-3k dataset, WebWeaver teaches complex skills to small models, enabling them to have expert-level research capabilities and lowering the barrier to high-quality research.

WebWeaver's technical principles

• Dual-agent frameworkWebWeaver consists of two agents: a planning agent responsible for exploring and generating research outlines, and a writing agent responsible for synthesizing report content based on the outlines. This division of labor simulates the human research process, improving research efficiency and quality.
• Dynamic loop mechanismThe planning agent continuously conducts web searches during the research process, comparing and optimizing the search results with the research outline, forming a dynamic cycle. This allows WebWeaver to promptly discover new information and integrate it into the research, avoiding the problem of traditional methods where the outline is difficult to change once it is finalized.
• Applications of memory banksWhen generating a report, the writing agent retrieves the most relevant evidence from a carefully curated memory. This memory stores information retrieved from the network and preprocesses and labels it according to certain rules.
• Reinforcement learning and optimizationWebWeaver continuously optimizes its search and generation strategies through reinforcement learning. During training, it adjusts its behavior based on feedback signals, improving search efficiency and report quality. This adaptive optimization mechanism enables WebWeaver to better handle diverse research topics and task requirements.

WebWeaver project address

• GitHub repositoryhttps://github.com/Alibaba-NLP/DeepResearch/tree/main/WebAgent/WebWeaver
• arXiv technical paper: https://arxiv.org/pdf/2509.13312

WebWeaver application scenarios

• academic researchIt can quickly integrate literature data, generate literature reviews and paper drafts, and help researchers conduct research efficiently.
• Enterprise Decision SupportEnterprises collect market and industry information to provide data support for market research, strategic planning, and investment decisions.
• EducationTeachers can access teaching resources to assist in course design, and students can use them for learning aids and graduation projects, thereby improving teaching and learning effectiveness.
• Media and News IndustryReporters and media personnel can quickly gather news background and expert opinions to improve the quality of news reports and special feature planning.
• Government and Public PolicyGovernment departments collect socio-economic information to provide a scientific basis for policy making, public affairs management, and emergency management.