authority-bias

This paper investigates authority bias in LLMs used as conversational search engines for academic paper recommendation, finding that LLMs show significant preference for papers based on authority signals like author prestige and citations, with debiasing only partially effective.

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.