How do large language models learn to 'reason'?

Large language models learn to reason by practicing with lots and lots of examples, just like a kid learns to count by counting blocks over and over.

Imagine you have a big box full of puzzles, and every time you solve one, you get a little smarter about how puzzles work. That's kind of what happens inside a language model. It looks at many sentences, stories, or even math problems, all the different ways people use words to think and explain things, and it tries to figure out the patterns.

Each time the model sees a sentence, like "If I have 3 apples and eat 1, how many are left?" it tries to answer. If it gets it right, it feels good. If it gets it wrong, it learns from its mistake. It's like playing a game of "What comes next?", the more you play, the better you get at guessing.

Once the model has practiced enough, it can take new problems and work through them step by step, just like you would when solving a puzzle or figuring out how many toys fit in a box. It uses all the patterns it’s learned to think, not magic, just clever practice!