Training technique where the model is fed ground‑truth tokens for the next step instead of its own predictions, speeding convergence but risking exposure bias.