Paper proposes ReCast framework for fixing sparse-hit learning in generative recommendation using RL.