Causal representation learning for downstream tasks (formalized as reward), with algorithms and regret bounds.