authority-bias

This paper investigates authority bias in LLMs using a controlled medical QA setting, revealing that models override correct answers in a graded manner proportional to perceived authority. The effect is localized to a critical late layer where correct answer representations are actively erased.