WebResearcher - An open-source iterative deep research agent from Alibaba.

What is WebResearcher?

WebResearcher, developed by Alibaba's Tongyi Labs, is an iterative deep research agent belonging to the Tongyi DeepResearch family. Based on an innovative iterative deep research paradigm, it simulates the cognitive workflow of human experts, autonomously decomposing complex problems, coordinating tool usage, and integrating findings into a coherent and well-reasoned narrative. Compared to traditional research agents, WebResearcher avoids information overload and noise accumulation by processing the research process in stages, ensuring continuous deep reasoning capabilities. WebResearcher is equipped with a scalable data synthesis engine and a dedicated multi-stage training process, including rejection-based fine-tuning and reinforcement learning with verifiable rewards, demonstrating superior performance in complex reasoning tasks.

Main functions of WebResearcher

• Autonomous decomposition of complex problemsBreak down complex research tasks into multiple manageable sub-tasks.
• Coordination tool usage: Use various tools as needed, such as search engines and academic databases.
• Integration and DiscoveryIt integrates the retrieved information and tool outputs into a coherent and well-reasoned narrative.
• Continuous Deep ReasoningThrough iterative processes, we continuously perform deep reasoning to avoid information overload and noise accumulation.

WebResearcher's technical principles

• Iterative research processThe research process is broken down into multiple discrete rounds, each consisting of three parts: "Think," "Report," and "Action." The "Report" from each round serves as a central memory, integrating new findings into a coherent, high-density summary that is then passed on to the next round. This cyclical process of synthesis and reconstruction prevents cognitive space overload and noise pollution, allowing for sustained deep reasoning.
• Scalable data composition engineUsing a multi-agent framework, a three-stage workflow is employed to automatically generate large-scale, high-quality, and complex inference task data. This includes initial data generation, iterative complexity enhancement, and rigorous quality control.
• Training and reasoning:
  • Rejection-based Fine-Tuning (RFT)Fine-tuning is performed on high-quality trajectories to ensure that the final answer matches the true value perfectly, cultivating robust tool usage skills and knowledge-based reasoning abilities.
  • Reinforcement Learning (RL)Further enhance the agent's multi-step logical reasoning ability through reinforcement learning with verifiable rewards (RLVR).
  • Test-Time Scaling (TTS)In the reasoning process, performance is improved by running multiple parallel reasoning paths and using a specialized fusion agent to synthesize the final answer from the last few steps of each path.

WebResearcher project address

• GitHub repositoryhttps://github.com/Alibaba-NLP/DeepResearch/tree/main/WebAgent/WebResearcher
• arXiv technical paper: https://arxiv.org/pdf/2509.13309

Application scenarios of WebResearcher

• academic researchIt helps researchers quickly sort through literature, uncover key information, assist in conducting complex academic research, and improve research efficiency and quality.
• Market AnalysisWebResearcher can collect and analyze market data, uncover industry trends and consumer needs, provide businesses with accurate market insights, and assist in decision-making.
• Technology DevelopmentIn the technology field, it is used for technology trend research, competitor analysis, etc., to help developers grasp the forefront of technology and accelerate technology iteration.
• Educational guidanceIt provides students and educators with integrated learning resources and knowledge explanations to support the teaching and learning process.
• HealthcareIt assists medical personnel in disease research, drug development information collection, and other tasks, providing data support and knowledge background for medical decision-making.