LLM Epistemic Sensitivity Varies by Phrasing and Verb Choice

Evaluating 10 LLMs shows that model performance on navigating user beliefs versus facts is highly sensitive to linguistic phrasing. Accuracy gaps vary significantly depending on the epistemic verb used, such as 'vaguely remember' versus 'seriously doubt,' due to model task confusion.