A Systematic Review of Prompt Engineering Paradigms in Organic Chemistry: Mining, Prediction, and Model Architectures

Large language models (LLMs) have emerged as transformative tools in scientific research, offering a powerful alternative to traditional, resource-intensive machine learning methods. By leveraging the vast knowledge encoded during pre-training, prompt engineering—the systematic design and optimization of input instructions—enables researchers to guide LLMs toward accurate and domain-specific outputs without updating model parameters. This review presents the first systematic examination of prompt engineering techniques within organic chemistry, focusing on two critical application areas: text mining and predictive tasks. We analyze the core paradigms of prompt engineering, including prompt design, prompt learning, and prompt tuning, and clarify terminological inconsistencies in the literature. The discussion is contextualized within the three principal LLM architectures (encoder-only, decoder-only, and encoder-decoder), with an evaluation of their respective performances on chemistry-related tasks. Furthermore, we explore practical workflows for extracting structured chemical data from texts and knowledge graphs, as well as advanced prompt strategies for reaction condition prediction, reaction optimization, and catalytic performance prediction. This review highlights the significant potential of LLM-driven prompt engineering to accelerate discovery in organic chemistry, from synthetic pathway optimization to automated literature analysis, while also addressing persistent challenges such as the limitations associated with various prompt engineering techniques and the constraints related to each related sub-task. We conclude by outlining future research directions aimed at deepening the integration of chemical knowledge with evolving AI methodologies.