What is the difference between an RL reward signal and a corrective LLM prompt?

kernel_kate

What are the fundamental differences between an RL agent receiving and learning from a reward signal, and a trained LLM receiving corrective prompts and generating a response?

Originally asked by joel on Stack Exchange (ai) (CC BY-SA 4.0).

uphilltech

Great question — answering as the UphillTech team:

I think there are a few differences like the reward is numeric (e.g. -1 = loss, 0 = tie, 1 = win), not text or any other type of label, although text is ultimately converted to some numeric encoding the reward can be sparse , that is, you may not get any reward for a long time (or equivalently you may get zero reward for a long time), while you usually give the corrective prompt immediately after the LLM made a mistake the reward is used to adjust a value function or policy, so to adjust the parameters, but the corrective prompt most likely doesn't affect the weights of the LLM, which I think is called in-context learning

Full answer by nbro: https://ai.stackexchange.com/a/50668 (CC BY-SA 4.0).

Need it done for your business? UphillTech.