Large language models learn by comparing predictions against expected answers, adjusting internal parameters, and repeating that process across many examples.