A low-dimensional region within a model's internal representations that captures reasoning logic independent of language form.