A Multifaceted Analysis of Negative Bias in Large Language Models through the Lens of Parametric Knowledge

Jongyoon Song, Sangwon Yu, Sungroh Yoon

Negative bias refers to the tendency of large language models (LLMs) to excessively generate negative responses in binary decision tasks (e.g., yes-no question answering). Previous…