Dimitris Papailiopoulos

We show that transformers achieve length generalization when training on shorter main task and longer auxiliary tasks together.