“Quiet-STaR, a method for token-level reasoning, improved zero-shot results on CommonsenseQA from 36.3% to 47.2% and on GSM8K from 5.9% to 10.9% for the Mistral 7B model.”