What is Reinforcement Learning?

Unlike supervised learning, which trains a model on fixed examples of correct answers, RL trains through trial and interaction: an agent takes an action, the environment returns a reward signal (good, bad, or somewhere in between), and the agent updates its strategy — its policy — to favor actions that led to higher reward over time. This makes RL well suited to problems where the "correct" behavior isn't a single labeled answer but a sequence of decisions, like playing a game, controlling a robot, or optimizing a multi-step process. In modern language models, RL is central to alignment: after a base model is pretrained on text, it's further refined using RL against a reward signal that reflects human preferences (RLHF) or against verifiable rewards for tasks with a checkable correct answer, like math or code, which is part of how reasoning models are trained to "think" more effectively rather than just predict the next likely token. The technique's core challenge is reward design: an agent optimizes exactly what it's rewarded for, so a poorly specified reward function can produce technically-high-scoring but undesirable behavior — a classic and well-documented failure mode.