Stress-testing large language model agents in a robotic chemistry laboratory

作者：Lulu Guo, Yingkai Sun, Xiaobo Li, Luyao Ge, Ziming Wang, Haitao Zheng, Jingyu Li, Huijuan Zhang, Bingxu Chen, Daobin Liu, Yuebo Liu, Jie Li, Xiaohui Li, Linjiang Chen, Yi Luo, Jun Jiang · 单位：Stress-testing large language model agents in a robotic chemistry laboratory, State Key Laboratory of Precision and Intelligent Chemistry, Hefei National Research Center for Physical Sciences, at the Microscale, School of Chemistry and Materials Science, University of Science and Technology of China, Center for Scientific Intelligence Innovation, Hefei, China, School of Chemistry, School of Computer Science, University of Birmingham, Birmingham, UK, physical action and adaptation to evidence. Here, we use a robotic chemistry laboratory as a physical-world, laboratory constraints, laboratories. This capability extends beyond generating plausible hypotheses and experimental, laboratories have enabled automated and closed-loop physical experimentation (6–13), while, convert scientific intent into executable action in a robotic laboratory (19, 20). This challenge is, particularly pronounced in modular robotic laboratories that integrate heterogeneous capabilities, powered by large language models in a modular robotic chemistry laboratory, evaluating their · 会议/期刊：arXiv preprint · 方向：cs.AI · 发布日期：2026-07-28