Speaker's Brief Introduction：蔡家麒，麻省理工学院（MIT）Pappalardo Fellow。2019年毕业于华中科技大学物理学院，2024年于华盛顿大学取得博士学位后，在MIT任Pappalardo Fellow至今。将于今年年末加入合肥国家实验室。他专注于凝聚态物理实验研究，利用光学与电学输运手段，探索二维量子材料中的新奇物相，尤其关注其在量子器件、量子计算与量子传感等领域的潜在应用。他因博士期间在分数量子反常霍尔效应上的开创性工作，获得2025年William McMillan Award。
Abstract: Modern physics condensed matter experiment increasingly operate in high-dimensional parameter spaces. In such settings, the central bottleneck is no longer only how to perform a measurement, but how to define, evaluate, and guide the search for meaningful physical signals. This makes condensed matter experiment a natural frontier for agentic AI.
In this talk, I will discuss the recent rise of agentic AI from the perspective of what is called the “second half” of AI: not only building more capable models, but also defining the right problems, constructing faithful evaluations, and designing harnesses that allow agents to act safely and productively in real scientific workflows. I will first introduce QDevBench, an instrument-centered benchmark for condensed-matter experiments. We develop QDevBot through a co-evolution loop between benchmark tasks, expert feedback, and deployed experimental workflows, aiming toward an agentic copilot for experimentalists. I will then present how the same co-evolutionary philosophy can extend beyond measurement. In KLayoutClaw, we combine agentic workflows with traditional computer-vision and layout-analysis tools, producing a system that substantially outperforms generic coding agents on device-layout tasks by grounding actions in native geometric and visual representations. Finally, I will describe recent work on PatchOptic: an approach to optimizing agentic workflows through bidirectional data accessors, whose mathematical foundation is profunctor optic categories. This framework suggests a general way to control what agents can see, how they can modify structured states, and how local patches compose into reliable global workflows.