Review for NeurIPS paper: A Simple Language Model for Task-Oriented Dialogue

Summary and Contributions: The authors propose SimpleTOD, which can replace modular task-oriented dialogue models to unified causal language model in an end-2-end manner. There are three sub-tasks in the task-oriented dialogue. They are dialogue state tracking, action prediction, and response generation. SimpleTOD treats all three sub-tasks as sequence generation. Whole up to dialogue context C_t is used as the first input to the model, and the model generates dialogue state B_t at turn t.