Transformer-based models rely on pre-training, fine-tuning, and in-context learning, with domain-specific adaptation for process safety corpora. Multimodal approaches combine text, image, and structured data, while trade-offs between open-source and proprietary models impact industrial deployment decisions.