Eliminating Position Bias of Language Models: A Mechanistic Approach

Ziqi Wang, Hanlin Zhang, Xiner Li +6

Position bias has proven to be a prevalent issue of modern language models (LMs), where the models prioritize content based on its position within the given context. This bias ofte…