Beyond human data: Scaling self-training for problem-solving with language models

TL;DR

Finetune a language model with data it generated itself helps.