Reinforcement Learning

Reinforcement learning (RL) is a broad training paradigm in which a model learns by trial and feedback: it takes an action, receives a signal about how good that action was, and adjusts its behavior accordingly — much like how a person learns from the consequences of their choices. In the context of language models, the “actions” are generated tokens and the “reward” usually comes from a reward model trained on human preferences. RL can produce dramatic behavioral improvements but is also tricky to get right; the model can learn unexpected shortcuts if the reward signal doesn’t perfectly capture what you actually want. Understanding the basics of RL helps behavior architects reason about why models develop certain tendencies during training.