Next-Token Prediction Should be Ambiguity-Sensitive: A Meta-Learning Perspective