Reinforcement learning (RL) is a training paradigm where an agent learns by interacting with an environment, receiving scalar reward signals for actions, and iteratively updating its policy to maximise long-run cumulative reward. The agent has no direct access to ground-truth labels; it must infer good behaviour purely from the reward feedback loop, balancing exploration of new strategies against exploitation of known ones.

Think of it like a child learning to ride a bicycle: no one describes the exact muscle movements needed. The child tries, falls, adjusts, and over many attempts builds an intuitive policy for staying upright-guided entirely by the signal of "fell over" versus "kept going".

In the context of language model alignment, RL is applied after supervised pretraining. A policy model generates candidate responses; a Reward Model scores them; and a policy-gradient algorithm such as PPO nudges the model toward higher-scoring outputs. This loop-used in RLHF-is how models learn to be helpful, harmless, and honest rather than just statistically fluent.

RL is the foundation of RLHF and connects directly to Reward Model training, Fine-tuning, and Loss Function design; phenomena like reward hacking are pathologies that emerge when the reward signal is imperfectly specified. Gradient Descent optimisers handle the weight updates during policy learning.

You will encounter reinforcement learning any time you read about game AI, robotics, or the alignment techniques used to shape the behaviour of modern LLMs and chat assistants.