Training
What Is Reinforcement Learning? The Basics Behind RLHF
Reinforcement learning is how a model learns from feedback on its own actions rather than from labeled examples — the foundation of the training stage that turns a base model into a helpful assistant.
Learning from consequences, not labels
Most training so far in a model's life is supervised — given an input, learn to match a known correct output. Reinforcement learning is different: an agent takes actions, receives a reward signal indicating how good that action was, and gradually adjusts its behavior to earn higher rewards over time — with no single "correct answer" provided for each step.
Why it's the right fit for shaping model behavior
There's no single "correct" reply to "explain quantum computing to a 10-year-old" the way there's a correct answer to a math problem. Reinforcement learning lets a model be trained toward replies that are rated as better by some judge — a human, or a trained reward model — rather than needing an exact target output for every possible input.
Where it shows up in LLM training
This is exactly the mechanism behind RLHF — reinforcement learning from human feedback — the training stage that takes a raw, freshly pretrained model and shapes it into something that reliably behaves like a helpful assistant rather than a raw text-completion engine.
Frequently Asked Questions
What is reinforcement learning?
A training approach where an agent learns by taking actions and receiving reward signals for how good those actions were, rather than being given a single correct output for each input.
How is reinforcement learning used in training LLMs?
It's the mechanism behind RLHF, the training stage that shapes a raw pretrained model into a helpful, well-behaved assistant based on feedback about which responses are better.