It has been observed that professionals across finance, consulting, healthcare, policy research, legal advice, scientific/academic research among other fields routinely produce in-depth, specialized research reports. Yet creating such reports requires labor-intensive, time-consuming tasks—retrieving, extracting, gathering, verifying, and summarizing information.
To this end, our mission is to liberate these knowledge-intensive workers from low-value, time-intensive work. Deep Research helps to deliver reports with articulate reasoning chains and clearly sourced information, enabling effortless validation and accessibility with much time saved.
Moreover, through user trials, we have discovered to our delight that beyond providing high-quality, accessible information, Deep Research provides perspectives and analytical frameworks that inspire users and broaden their mindsets. This is why we envision Deep Research as the next paradigm of collaboration between human intelligence and computational power.
StepFun Deep Research intelligently bridges your research objectives with actionable, trustworthy insights. Leveraging an end-to-end Multi-Agent architecture, it autonomously executes complex research workflows—searching, intelligently filtering hundreds of web sources, performing computations and analyses via code execution, and delivering visually rich presentations. High-quality, up-to-date reports are generated within minutes.
Performance:
xBench-DeepSearch (Sequoia China): Designed for Agents in Chinese internet environments, focusing on complex information retrieval, StepFun Deep Research achieved a 70% pass rate—outperforming all other products and models and demonstrating industry-leading capabilities in Chinese-language deep search and synthesis.
BrowseComp (OpenAI): A cutting-edge benchmark for real-world web browsing and multi-hop reasoning, StepFun Deep Research achieved a 23% pass rate—ranking among the highest known results. This further validates its exceptional strength in navigating intricate, interconnected online information.
To better fulfill interdisciplinary research tasks, domain-specific data frameworks have been structured and developed, covering business survey, research consulting, workplace operations, consumption, and personal productivity. Leveraging reinforcement learning, our model aims to address practical solutions through:
Here are some specific cases that provide a clear view of how in-depth research addresses complex research questions.
Lithium Battery Equipment Industry Research Report
Cutting-Edge Technology in UAV Bridge Inspection
Guide to English Enlightenment for Four-Year-Old Girls
Deep Research supports data coming in diverse types and formats including:
Deep Research stands out for its autonomous access to and interaction with external tools and environment. Its research efforts are adjusted timely according to dynamic feedback. Reports are thus produced based on real-time feedback and multi-sourced data gathered. Beyond searching and browsing capabilities, Deep Research offers other integrated features including:
Deep Research learns from user interactions all the time. Through research plan refinement and prompt clarification, users can personalize insight generation workflows, streamline research methodologies and enhance domain-specific alignment. In this way, the system progressively adapts to users’ professional expertise and cognitive patterns, delivering increasingly tailored efficiency.
Click on the In-depth Research mode in the StepFun AI chat interface, then enter your research question in the chat box. In addition to typing text and providing links, you can also upload tables and documents to enrich the background information of your requirements.
Before the research begins, the Deep Research agent may ask some questions to clarify the research objectives and scope. We have found that more and more accurate demand information can significantly improve the quality of the report. For some time-sensitive questions, the Deep Research agent will call the search tool to clarify the research objectives and ask questions accurately.
Next, the in-depth research intelligent agent will develop a research plan tailored to the target problem. If you are not satisfied with the plan, you can modify it through dialogue. Once the plan is confirmed, the intelligent agent will search through over 130 web pages, browse key websites, and if necessary, execute code for calculations and analysis. It will also utilize visualization tools to create concise and clear charts. The entire process trajectory will be displayed in the question-and-answer flow, allowing you to review it at any time.
Since in-depth research requires accessing real-time, extensive, and broad information as the basis for report writing, the entire task execution typically takes tens of minutes or longer. Therefore, the task will be automatically executed in the cloud, allowing you to leave the current task interface at any time.
To facilitate quick access to key data information in the report, the in-depth research intelligent agent will use rich tables and interactive charts to present complex data, enhancing your reading efficiency.
We hope that Deep Research will be part of your daily workflow to enhance your work efficiency and experience. It is also expected that valuable and inspirational information and advice will be offered and easily accessed by this AI agent.