Improving Next Tokens via Second-Last Predictions with Generate and Refine