Reinforcement Learning

Most machine learning learns from labeled examples: here is an input, here is the right answer. Reinforcement learning works differently. An agent takes actions, receives a reward signal indicating how well it did, and adjusts to earn more reward. Nobody supplies the right answer; the system discovers it.

That makes it the natural fit for problems where good behavior is easy to score but hard to demonstrate — games, robotic control, and any task with a checkable outcome. It is also considerably harder to get working than supervised learning, because the system must explore enough to find good strategies without wandering indefinitely.

It is now central to language models rather than a separate branch. RLHF uses it to turn a raw text predictor into an assistant, and reasoning models are trained largely through reinforcement learning on problems whose answers can be verified automatically.

Reinforcement learning is where a model's behavior is shaped rather than its knowledge, and where the well-known failure mode lives: a system optimizes exactly what you measured, not what you meant. Reward specification is genuinely difficult, and most surprising machine behavior traces back to a reward that was easier to satisfy in an unintended way.

Rarely something you run yourself, but the concept explains behavior you will meet constantly. When a model games an instruction — technically satisfying it while missing the point — that is reward optimization working as designed on an imperfect target.