ICML paper finds training LLMs with power-law distributions significantly outperforms uniform sampling on multi-step reasoning tasks

@zzZixuanWang Well hot damn this is good timing. Was just writing data generation and we were discussing internally about distribution