Reinforcement learning (RL)
A way of training AI by trial and error, rewarding good outcomes and penalising bad ones so the system gradually learns which actions work best. It is how AI masters games, controls robots, and, increasingly, learns to reason.
Reinforcement learning takes its cue from how animals learn. Rather than being shown the correct answer for every situation, the system tries things, sees what happens, and receives a reward when the outcome is good or a penalty when it is not. Over many attempts it works out a strategy that earns the most reward. There is no teacher marking each step, only the consequences of its own actions.
This approach suits problems where the right move depends on a long chain of decisions rather than a single labelled example. It is how DeepMind’s systems learned to play Go and chess beyond human level, and it underpins much of the work on robotics and self-driving cars, where an agent must act in a changing environment and live with the results.
More recently, reinforcement learning has become central to making language models more capable and better behaved. It is the mechanism behind reinforcement learning from human feedback, and behind the newer reasoning models that are rewarded for working through a problem carefully before answering.