Researchers investigated whether large language models (LLMs) can be used as a standalone acquisition policy for materials optimization. They tested five LLMs on four tasks and compared their performance to random selection and Gaussian-process methods. The results show that LLMs generally outperform random selection, but their performance varies depending on the task, model, and presentation of candidates. While they have potential as acquisition policies, their reliability remains sensitive to these factors. --- Why it matters: This matters because materials optimization is a complex and costly process, and finding more efficient methods can accelerate scientific discovery. LLMs' ability to serve as standalone acquisition policies could simplify this process and make it more accessible to researchers without extensive computational resources. Source: https://arxiv.org/abs/2608.19790