Experimental reasoning in large language models

Abstract

Scientific reasoning requires interaction with the physical world: designing experiments, interpreting outcomes, and deciding what experiment should come next. Here we show that general-purpose large language models (LLMs) already possess substantial capabilities for experimental reasoning. We connected LLMs to a liquid-handling robot and camera through PEARL, which provides perception, action, and persistent state. Given objectives rather than prescribed workflows, the system designed and successfully executed experiments, interpreted observations, and revised its approach as evidence accumulated. When information was insufficient, PEARL autonomously designed and performed intermediate experiments to acquire it. In quantitative enzyme characterization, it judged whether measurements were sufficient, new data were required, or conditions needed to change. Across eight scenarios, these behaviors were reliable with limited agent scaffolding and no supplied scenario-specific recovery recipes. More broadly, general-purpose LLMs may provide a reasoning layer between scientific objectives and physical experimentation, shifting laboratory automation from executing predefined procedures toward AI systems that participate directly in the experimental process.