Sasha Rush and Yoav Goldberg argue prompt conditioning decouples machine learning models from policies

Goldberg suggests reframing methods like GRPO in ML terms