Zhejiang University unveils Qiushi Engine in bid to extend AI beyond research assistant role

Researchers at Zhejiang University in Hangzhou, Zhejiang province unveiled an AI scientific discovery system on July 15 that can independently organize and continuously advance scientific research, making it the latest effort to move beyond AI's traditional role as a research assistant.

Named the Qiushi Engine, the scientific discovery system was developed by the College of Information Science and Electronic Engineering at the university. It is designed to tackle open-ended scientific problems by autonomously planning, executing, and refining research tasks throughout the research process.

"Previously, AI mainly served as a research assistant, helping scientists search literature, write code, and process data," said Yang Yihao, the project's lead researcher. "But scientific research is far more than completing isolated tasks. It involves continuously raising questions, testing hypotheses, revising approaches, and moving a research project forward toward a common goal."

Unlike conventional AI tools that execute individual instructions, the Qiushi Engine continuously evaluates research progress and adjusts its strategy as it goes. It decides what to do next, generates follow-up tasks, and can sustain reasoning across thousands of steps.

Yang noted that the Qiushi Engine's approach closely aligns with the concept of "Loop Engineering", a scientific discovery framework proposed by OpenAI last month.

In practice, researchers simply provide an open-ended research objective, and the system carries out literature review, theoretical analysis, experimental design, programming, data analysis, and result evaluation itself, adjusting its plans based on experimental feedback in an iterative cycle.

The system has already demonstrated its capabilities on a real-world optical research platform. Given only an open objective, the Qiushi Engine ran a 206-step research process lasting about 21.5 hours and proposed four candidate directions. Through multiple rounds of trial and error against real experimental feedback, it identified the causes of failures, revised its plans, and ultimately proposed and experimentally validated an optical bilinear interaction mechanism for Transformer-like relational computation. In a separate task, the team said it completed in about six hours what would typically require researchers several weeks of work.

The team emphasized that while AI can take on substantial, sustained parts of the research process, human researchers remain responsible for setting objectives, reviewing evidence, and validating conclusions through appropriate scientific methods, including real-world experiments where applicable.

Pan Yunhe, an academician of the Chinese Academy of Engineering and an artificial intelligence expert, said global technological competition is increasingly extending beyond the development of more powerful large models to their application in solving complex real-world problems, with scientific research emerging as one of the most important areas.

Pan believes that as AI for Science continues to evolve, artificial intelligence is shifting from a research tool into a research collaborator: humans will focus on identifying scientific questions and setting directions, while AI will take on continuous exploration, analysis, and optimization.

Addressing concerns over whether AI could replace researchers, Chen Hongsheng, dean of the College of Information Science and Electronic Engineering and head of the research team, said future education will place greater emphasis on cultivating students' higher-order abilities, such as asking good questions, critically evaluating evidence, and understanding the limitations of methodologies.

Looking ahead, the Qiushi Engine is expected to integrate with a wider range of research tools and experimental platforms to support fields including materials science, quantum science, high-pressure science, and the humanities and social sciences, and to expand from internal research validation to broader real-world scientific applications.