An arXiv preprint introduces Acmite, a method for reducing gender bias in large language models. The abstract argues that LLMs reproduce social stereotypes from their training data, and that existing debiasing approaches often depend on explicit biased examples or predefined group-term lists. Acmite is instead guided by concept-level mutual information, though the abstract does not specify the full mechanism.
Mutual information is a measure of statistical dependence; in this context, the method appears to steer model representations away from gender-stereotypical associations without requiring hand-labelled bias examples. The authors position this as a departure from common debiasing pipelines.
Because the available text is only the abstract, the preprint's experimental setup, baselines, and quantitative results are not described here. The paper is listed as arXiv:2610.01696v1.