A family of large language models for materials research with insights into model adaptability in continued pretraining

Nature Machine Intelligence, Published online: 27 February 2026; doi:10.1038/s42256-026-01199-8

Dhruv et al. introduce LLaMat, a family of LLMs for materials science, showing how specialized training enables smaller models to surpass larger general LLMs. The study reveals that extensively pretrained models like LLaMA-3 adapt less effectively than LLaMA-2 to new scientific domains.