It's all a next token predictor, which is trained at first to predict the next token. RLVR etc. are also training the next token predictor.