Reinforcement learning trains a model, often called an agent, by letting it interact with an environment and receive rewards or penalties based on the outcome of its actions. There's no dataset of correct answers up front — the agent has to try things, see what works, and adjust its strategy to earn more reward over time.
This is the approach behind systems that mastered complex games like Go and StarCraft, where the "right move" in any given position isn't something you could hand-label in a training set. It's also used to fine-tune language models, where human feedback about which responses are more helpful acts as the reward signal.
The catch is that reinforcement learning can find unexpected shortcuts to maximize reward that technically satisfy the goal you specified but miss what you actually wanted — a well-known pitfall researchers call reward hacking, and a big reason careful reward design matters as much as the algorithm itself.