AI Agents for Science: From Emerging Capabilities to Verified Autonomy

=======
Abstract
=======
Recent advances in AI are opening new possibilities for scientific discovery, with AI agents increasingly capable of solving complex problems, using scientific tools, and carrying out multi-step research workflows. These developments raise an important question: how close are we to truly autonomous scientific research? In this talk, I will first review recent progress in AI agents for science, highlighting emerging capabilities and the challenges that remain. I will then focus on quantum engineering, where autonomous agents must go beyond scientific reasoning to design experiments, interact with quantum systems, interpret observations, and adapt their strategies. I will introduce our recent work on Quantum-Harbor, a virtual laboratory designed to evaluate verified autonomy in quantum engineering. Through 49 expert-designed tasks and evaluations of 17 frontier agentic systems, we investigate whether AI agents can not only complete scientific tasks but also support their conclusions with verifiable evidence. Our findings reveal substantial differences in performance and highlight the challenges of achieving reliable scientific autonomy. Finally, I will discuss the opportunities and open questions toward more autonomous scientific research.

===============
About the Speaker
===============
Naixu Guo is a PhD researcher at the Centre for Quantum Technologies, National University of Singapore, advised by Patrick Rebentrost. His research focuses on quantum algorithms and quantum learning theory. More recently, he has also been exploring AI for science, particularly the development and evaluation of autonomous scientific agents.