“The Chain of Hindsight (CoH) technique fine-tunes a language model to improve its own outputs by presenting it with a sequence of past outputs annotated with feedback.”