Thank you, I am also very confused on this point. I hope someone else can clarify. As a guess, could it mean that you would run the model forward a few tokens

Thank you, I am also very confused on this point. I hope someone else can clarify. As a guess, could it mean that you would run the model forward a few tokens for each of its top predicted tokens, keep track of which branch is performing best against the training data, and then use that information somehow in training? But search is supposed to make things more efficient at inference time and this thought doesn't do that...