Be interesting to plug this into a bayesian optimization like framework: find out regions of language space where the models maximally disagree and then target

Be interesting to plug this into a bayesian optimization like framework: find out regions of language space where the models maximally disagree and then target those areas for extra training