| Abstract Scope |
Predicting alloy properties from chemical composition remains a central challenge in materials science, especially in small-data contexts where conventional machine-learning models often show limited performance. Recent advances in Large Language Models (LLMs) have revealed an unexpected ability to handle tabular regression tasks, even though these models were originally designed for natural language processing. Building on these observations, studies have explored the use of LLMs for materials prediction problems, showing that they can match or sometimes surpass classical regressors under specific conditions. In this work, we present a focused, non-exhaustive review of LLM-based regression approaches applied to alloy property prediction. We cover methods where LLMs support feature engineering, generate composition-informed embeddings (e.g., BERT), or perform end-to-end regression via in-context learning. These techniques are compared with traditional models (e.g., XGBoost) to identify scenarios where LLMs may provide advantages. Our goal is to clarify their potential, limitations, and emerging role in materials informatics. |