Medical AI models like MedGemma show dangerous sensitivity to prompt formatting, with chain-of-thought prompting paradoxically reducing accuracy by 5.7%

MedGemma (4B and 27B parameters) tested on 4,183 MedMCQA and 1,000 PubMedQA questions reveal critical robustness issues in medical settings

Chain-of-Thought prompting unexpectedly decreased accuracy by 5.7% compared to direct answering, contradicting common AI practices

Few-shot examples degraded performance by 11.9% while increasing position bias from 0.14 to 0.47, showing models favor certain answer positions

Answer shuffling caused the model to change predictions 59.1% of the time with accuracy drops up to 27.4 percentage points

Front-truncated context caused severe accuracy collapse below baseline, while back-truncation preserved 97% of full-context accuracy