How do large language models develop new reasoning abilities?

Large language models learn new reasoning abilities by practicing with lots of examples, just like kids learning to solve puzzles.

Imagine you have a big box full of jigsaw puzzle pieces. Each piece has a picture on it, and the whole box is filled with different kinds of puzzles, some easy, some hard. Every time you complete a puzzle, you get better at solving them. That's how large language models work: they look at many examples of problems being solved, like math questions or riddles, and slowly figure out patterns and rules.

Think about learning to count, you start with small numbers, then bigger ones. A language model starts by learning simple sentences and gradually moves on to more complex ideas. Every time it sees a new problem, it tries different ways to solve it, just like you try counting on your fingers or using blocks.

The more examples the model practices with, the smarter it gets. It's like having a teacher who gives you lots of exercises, every puzzle you finish helps you become better at solving new ones. Over time, the model can even come up with its own ways to solve problems, just like you learn tricks to make counting faster!