Abstract
Utilizing large Language models (LLMs) for handling scientific information comes with risk of the outputs not matching expectations, commonly called hallucinations. To fully utilize LLMs in research requires improving their accuracy, avoiding hallucinations, and extending their scope to research topics outside their direct training. There is also a benefit to getting the most accurate information from an LLM at the time of inference without having to create and train custom new models for each application. Here, augmented generation and machine learning-driven prompt optimization are combined to extract performance improvements over base LLM function on a common chemical research task. Specifically, an LLM was used to predict the topological polar surface area (TPSA) of molecules. By using augmented generation and machine learning-optimized prompts, the error in the prediction was reduced to 11.76 root-mean-squared error (RMSE) from 62.34 RMSE with direct calls to the same LLM.