Augmented and Programmatically Optimized LLM Prompts Reduce Chemical Hallucinations.

Area of Science:

• Computational chemistry
• Artificial intelligence in scientific research

Background:

• Large Language Models (LLMs) offer potential for scientific data analysis but are prone to inaccuracies (hallucinations).
• Improving LLM accuracy and reliability is crucial for their adoption in research applications.

Purpose of the Study:

• To enhance the performance of LLMs in scientific tasks, specifically chemical research.
• To reduce prediction errors in LLM-driven molecular property calculations.

Main Methods:

• Combined augmented generation with machine learning-driven prompt optimization for LLM task enhancement.
• Applied the enhanced LLM approach to predict the Topological Polar Surface Area (TPSA) of molecules.

Main Results:

• Reduced the root-mean-squared error (RMSE) for TPSA prediction from 62.34 to 11.76.
• Demonstrated significant performance improvement over standard LLM function for this chemical research task.

Conclusions:

• Augmented generation and ML-optimized prompts effectively improve LLM accuracy in scientific applications.
• This approach allows for more reliable LLM use in research without custom model training.