Paper recorded by Signals 4 on 2026-10-06 in cs.CL. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-06 on arXiv · recorded by Signals 4 on 2026-10-07
Category: cs.CL · 自然语言处理 · first seen 2026-10-07
A common approach to measuring bias in Large Language Models is to compare the log-likelihoods of two contrastive stereotype sentences. We argue that such single-pair comparisons are often unreliable: simply rewriting the same stereotype with an alternative attribute can yield logically inconsistent preferences. To address this, we propose a dual minimal pair setup that introduces two axes of comp