Retraining with Predicted Hard Labels Provably Increases Model Accuracy