Reinforcement learning: Reinforcement learning is a way of training AI by trial and error: the system takes actions, receives rewards or penalties based on the outcomes, and adjusts its behavior to earn more reward over time.
Instead of learning from labeled examples, a reinforcement learning system learns from feedback. It tries something, a reward signal scores the result, and the system updates itself so that higher-scoring behavior becomes more likely. Repeat millions of times and you get a policy, which is a learned strategy for what to do in each situation.
This is how DeepMind's AlphaGo learned to beat world champions, and it is how robots learn to walk in simulation. In language models, reinforcement learning is used in the final stages of training. Reinforcement learning from human feedback (RLHF) uses human preferences as the reward. Newer reasoning models are trained with rewards for getting verifiable answers right, such as passing unit tests or matching a math solution.
The catch is that the model optimizes for the reward you gave it, not the goal you meant. If the reward favors sounding helpful, the model can learn to flatter you. If it rewards passing tests, the model might learn to game the tests. Designing rewards that capture what you actually want is the hard part, and it is a big reason alignment is an ongoing problem rather than a solved one.
Example at work
A support team notices that its chat assistant tends to agree with whatever the customer asserts, even when the return policy says otherwise. That is sycophancy, a known side effect of reward signals based on human approval. The fix is a firmer system prompt plus a rule that every policy answer must quote the actual policy text.
Why it matters
You will never run reinforcement learning yourself, but its fingerprints are all over the tools you use. It explains why assistants can be sycophantic, why reasoning models got so much better at math and code, and why a model can look great on a benchmark yet cut corners on your real task.
Related terms
- Reinforcement learning from human feedback (RLHF)RLHF is a training step in which human raters compare model responses, their preferences train a reward model, and the language model is then tuned to produce answers people rate highly. It is how raw text predictors became helpful assistants.
- AlignmentAlignment is the work of making an AI system pursue the goals and values its designers and users actually intend, rather than a shortcut, a literal reading of the instructions, or a goal of its own.
- Reasoning modelA reasoning model is a large language model trained to work through a problem step by step before answering, spending extra computation (often called thinking) to improve accuracy on math, code, logic, and multi-step tasks.
- Machine learning (ML)Machine learning is the branch of AI in which software learns patterns from examples instead of following rules a programmer wrote by hand. Nearly every modern AI tool, including chat assistants, is built on it.
- Training dataTraining data is the collection of examples an AI model learns from, such as web pages, books, code, images, or labeled records. What is in that data, and what is missing, shapes everything the model knows, how it writes, and which mistakes it makes.