Reinforcement learning
Training by trial and reward: the model tries something, gets a score, and shifts towards whatever scored well.
Most training shows a model the right answer. Reinforcement learning doesn't: the model has a go, a score comes back, and training makes high-scoring behaviour more likely. Over many rounds it discovers strategies nobody spelled out.
That's how today's reasoning models were made. Trained with rewards for correct answers on checkable problems, models started writing longer working, testing their own steps and backtracking from mistakes. In a paper in Nature, DeepSeek reported that these habits emerged without people writing examples of good reasoning. The catch is that the model optimises the score, so a badly designed score teaches the wrong lesson.
For example
Give a model thousands of maths problems with known answers, reward it only when the final answer is right, and it gradually learns to check its own working along the way.
Put it to use
Where you'll meet reinforcement learning in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.