This method, introduced by Wei et al. (2022), has been proven to significantly improve the performance of large language models on various reasoning tasks.