Reinforcement learning
Learning by trial and reward rather than from fixed examples. Used to tune chat models and to train reasoning models.
Also: RL, RLVR, reinforcement learning from verifiable rewards
In reinforcement learning, a model tries an action, receives a reward, and adjusts to make well-rewarded actions more likely. For language models the actions are whole answers. RLHF uses a reward learned from people. Since 2024, labs have also used rewards that can be checked automatically, such as whether a maths answer is correct or code passes its tests. Training on such verifiable rewards at scale produced reasoning models like OpenAI o1 and DeepSeek-R1, which learned to write long chains of thought.