I reviewed Nathan Lambert’s slide contrasting classic reinforcement‑learning (RL) with the language‑model‑based RL‑from‑Human‑Feedback (RLHF) workflow, noting how the reward model serves as the target environment and the language model itself acts as the prompting agent.
I treat the reward model as the target environment in the LM‑based RLHF workflow. reward model is acting like a target environment
I consider the language model the agent that performs prompting actions. agent is language model which does an action of prompting