Across Generations, Sizes, and Types, Large Language Models Poorly Report Self-Confidence in Gastroenterology Clinical Reasoning Tasks

Abstract

<title>Abstract</title> This study evaluated confidence calibration across 48 large language models (LLM) using 300 gastroenterology board exam style questions. Regardless of response accuracy, all models demonstrated poor certainty estimation. Even the best-calibrated systems (o1 preview, GPT-4o, Claude-3.5-Sonnet) showed substantial overconfidence (Brier scores 0.15-0.2, AUROC ~0.6). Most concerning, models maintained high certainty regardless of question difficulty or their actual knowledge limitations. This metacognitive deficiency poses significant challenges for safe clinical implementation of current LLMs in gastroenterology.

Ähnliche Arbeiten