star

Definition

Capitalized as STaR, it stands for Self-Taught Reasoner: a bootstrapping loop in which a model generates its own chains of reasoning, keeps only those that reach the correct answer, and fine-tunes on them, then repeats. The point is that reasoning ability can be improved without hand-written rationales, using answer correctness as the only supervision. It is a common ancestor cited by later self-improvement and rejection-sampling work.