A training method where a model learns from its own predictions at the token level, providing fine-grained feedback.