An autonomous laboratory brings together agents who choose the question, derive theory, design experience... and reject a first measurement distorted by a drift of the signal.
AI systems are increasingly taking on roles in research, but risk drawing conclusions that the data don't support. Researchers describe an autonomous lab where agents based on large language models evaluate each other's work and can interrupt or redirect an investigation. They propose and discard research questions, require separate theoretical checks, and design experiments to test their own claims.
The device was applied to sets of quantum spins in the diamond. The agents established the minimum number of calibration settings needed to predict the system's response to a family of microwave pulse sequences, verified this result by calculation, and then devised experiments that could decide between competing predictions.
Don't lower the bar
The most telling passage concerns the first experiment: its results were consistent with the theory, but repeated reference measurements revealed that the signal drifted during acquisition. The agents therefore judged the result inconclusive and revised the protocol without relaxing their acceptance criteria. The second experiment, which met these criteria, confirmed the predictions. For the authors, this is proof that collaborative agents can choose a question, derive the theory from it and test it in the laboratory.
Read the full article on the original source
Go to the source →