Nature Machine Intelligence reported a study on applying language models to select the next candidate in experimental tasks. The description states that training with Bayesian objective functions allows using such models as natural-language-guided optimizers.
The authors consider tasks ranging from reaction optimization to molecular design. The source emphasizes that Bayesian optimization provides a fundamental way to choose the next experiment but typically depends on subject-matter expertise, which does not transfer well between fields.
The practical value of the work is related to an attempt to obtain calibrated estimates of uncertainty from language models, not only scientific knowledge. However the material provided contains only editorial description and does not reveal experimental results, comparisons with baseline methods, or limitations of the approach.
editorial commentary
Why it matters
If the results are confirmed in the full text and in practical evaluations, the next observable signal will be the publication of comparative data on the quality of experiment selection and calibration of uncertainty. Substantial uncertainty remains: the original material does not report metrics, the scale of testing, or real laboratory effectiveness.