“In the context of language models, a true reinforcement learning environment does not exist because the state (a prompt) is typically discarded after a single completion, rather than being part of an ongoing interaction.”