Reinforcement learning is a training approach in which a model learns by taking actions in an environment and receiving rewards or penalties based on the outcomes, gradually adjusting its behavior to maximize the reward it accumulates over time. Rather than being shown a single correct answer for each input the way supervised learning works, it explores different strategies and learns from the consequences of its own choices.
A widely used variant in modern AI is Reinforcement Learning from Human Feedback (RLHF), where human raters compare and rank a model's candidate responses, and those preferences are used to train a reward signal that shapes the model's future outputs. This is a major part of how large language models like ChatGPT are tuned after initial training, since it lets the model learn to prefer responses humans actually find helpful, honest, and safe, rather than just responses that are statistically likely given the training data.
