What is WebThinker
WebThinker is a deep research agent proposed by institutions including Renmin University of China, the Beijing Academy of Artificial Intelligence, and Huawei Poisson Lab. WebThinker empowers Large Reasoning Models (LRMs) to autonomously conduct web searches, web navigation, and report writing during the reasoning process. Based on a Deep Web Explorer and autonomous thinking, searching, and writing strategies, WebThinker enables LRMs to dynamically acquire information and generate high-quality research reports in real time. WebThinker's reinforcement learning-based training strategy further optimizes tool usage efficiency. WebThinker performs excellently in complex reasoning and report generation tasks, significantly improving the reliability and practicality of LRMs in knowledge-intensive tasks.
Main Features of WebThinker
- Autonomous Decision-Making: The LRM autonomously determines during the reasoning process when external knowledge is needed and when the report needs to be updated.
- Deep Exploration: Supports multi-step searches and page navigation to dig deep into information.
- Dynamic Writing: The model can write and revise report content in real time, equipped with a dedicated toolset (such as writing, checking, and editing) to ensure the coherence and completeness of the report.
- Tool Optimization: Optimizes the efficiency of the LRM's use of research tools.
Technical Principles of WebThinker
- Deep Web Explorer: Empowers the LRM with capabilities beyond traditional simple search, navigating between web pages based on interactive elements such as clicking links and buttons to dig deep into information. The model autonomously decides search queries, continuously explores until sufficient information is collected, and returns a refined summary.
- Reinforcement Learning-Based Training Strategy: Based on iterative online Direct Preference Optimization (DPO) training, it improves the LRM's efficiency in utilizing research tools (including search, navigation, and report writing tools). It constructs preference datasets, prioritizing reasoning paths that lead to correct answers, high-quality reports, and more efficient tool use.
- Operating Modes: The problem-solving mode equips the LRM with a Deep Web Explorer to deeply explore the web to solve complex problems. The report generation mode further grants the LRM writing, checking, and editing capabilities, iteratively writing comprehensive research reports while thinking and searching.
WebThinker Project Address
- Project Official Website：https://foremost-beechnut-8ed.notion.site/WebThinker
- GitHub Repository：https://github.com/RUC-NLPIR/WebThinker
- HuggingFace Model Library：https://huggingface.co/collections/lixiaoxi45/webthinker
- arXiv Technical Paper：https://arxiv.org/pdf/2504.21776
Application Scenarios of WebThinker
- Complex Question Answering: Quickly provides accurate answers for doctoral-level scientific questions or interdisciplinary problems.
- Research Report Generation: Autonomously searches and writes scientific research reports, ensuring comprehensive, accurate, and coherent content, and improving report generation efficiency.
- Deep Information Mining: Based on multi-step searches and page navigation, obtains in-depth information to support complex analysis and research.
- Educational Assistance: In the education field, helps students find learning materials and answer academic questions, generates teaching syllabi for teachers, and improves learning and teaching efficiency.
- Enterprise Decision Support: Provides enterprises with decision support such as market analysis and competitor analysis, helping management quickly obtain key information and make wiser decisions