**Xiaoxi Li$^{1*}$, Jiajie Jin$^{1*}$, Guanting Dong$^{1*}$, Hongjin Qian$^{2}$, Yutao Zhu$^{1}$, Yongkang Wu$^{3}$, Ji-Rong Wen$^{1}$, Zhicheng Dou$^{1\dag}$**

$^{1}$Renmin University of China, $^{2}$BAAI, $^{3}$Huawei Poisson Lab

Github: https://github.com/RUC-NLPIR/WebThinker

— Mar 31, 2025

<aside> 💡

WebThinker is a reasoning agent designed to autonomously search, deeply explore web pages, and draft research reports, all within its thinking process. Moving away from traditional agents that follow a predefined workflow, WebThinker enables the large reasoning model itself to perform actions on its own during thinking, achieving end-to-end task execution in a single generation.

</aside>

<aside> 📢

Our code and demo have been open-sourced in https://github.com/RUC-NLPIR/WebThinker, please check them out! For more details, please look forward to our technical paper, which will be released as soon as possible !!!

</aside>

Performance Overview:

Figure 1: Overall performance comparison on complex real-world problem solving and scientific research report generation tasks.

System Demo-1 (English): What are the models of OpenAI and what are the differences?

Demo4-Openaimodel_cut_compressed.mp4

System Demo-2 (中文): 我想投稿NeurIPS 2025，请告诉我这个会议的详细信息？

Demo6-NIPS-zh_cut_compressed.mp4

Deep Research Reports:

Example-1

🌟 Question: Please provide all the model information of OpenAI.

💡 Generated Research Report: Research Report on OpenAI Models: Evolution, Features, and Applications