Large Language Model Research Requires Accounting for Correlated Responses

Poster Number: P209
Presentation Time: 05:00 PM - 06:30 PM

Abstract Keywords: Large Language Models (LLMs), Evaluation, Standards
Primary Track: Foundations
Programmatic Theme: Clinical Research Informatics

Given randomness in large language model (LLM) text generation, analyses of repeated prompting of models are often necessary to understand variability in model responses. This study evaluates the effect on study results and conclusions when researchers ignore correlation of model responses with repeated LLM prompting. Using a recent research paper published in JAMA as a case study, this study demonstrates that ignoring correlation leads to misleading statements of statistical significance and study conclusions.

Speaker(s):
Robert Gallo, MD
VA Palo Alto Health Care System

Author(s):
Ethan Goh, MD, MS - Stanford University; Thomas Savage; Jonathan Chen - Stanford University Hospital;