Inspired by temporal-difference learning, we develop a new loss function that improves accuracy and data efficiency in incremental sequence classification.