June M. Liu

Multi-format training can drastically enhance language model consistency across different answer formats, with just 30% of training data needing augmentation to achieve significant gains.