A novel training approach called TaS that enables large language models to first generate reasonable thoughts and then express corresponding responses, mimicking the human cognitive process.
