A statistical framework for evaluating the repeatability and reproducibility of large language models.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Linguistics
• Medical Informatics

Background:

• Large language models (LLMs) offer potential in medicine but their reliability is a major concern due to stochastic text generation.
• Variability in LLM outputs can impact medical applications, yet standardized metrics for assessing this are lacking.

Purpose of the Study:

• To propose and validate a statistical framework for systematically quantifying the reliability of LLMs in medical applications.
• To introduce metrics for repeatability and reproducibility, assessing semantic consistency and internal stability of LLM responses.

Main Methods:

• Developed a framework measuring LLM repeatability (identical conditions) and reproducibility (different conditions).
• Evaluated semantic consistency and internal stability of LLM responses.
• Applied the framework to medical reasoning tasks using United States Medical Licensing Examination (USMLE) questions and Undiagnosed Diseases Network (UDN) cases.

Main Results:

• LLM responses showed less variability for complex UDN rare disease cases compared to standardized USMLE questions.
• Repeatability and reproducibility metrics did not correlate with diagnostic accuracy.
• The study highlights that consistent LLM output is not equivalent to accurate output.

Conclusions:

• The proposed framework provides a systematic approach to quantify LLM reliability in medicine.
• This quantification is crucial for the safe and effective integration of LLMs in clinical practice and biomedical research.
• Ensuring LLM reliability is paramount for advancing AI-driven healthcare solutions.