What it means
Asking a model to reason step by step, rather than answer immediately, substantially improves accuracy on multi-step problems. The reason is mechanical: a model produces one token at a time with a fixed amount of computation per token, so a question requiring several logical steps has nowhere to put that work if the answer must come out immediately. Generating intermediate steps creates room for it.
This started as a prompting trick — literally appending "think step by step." It has since been absorbed into the models themselves: reasoning models are trained to do this automatically and at length.
An important caveat: the written reasoning is not a reliable account of how the model reached its answer. Studies have shown models producing correct answers with flawed stated reasoning, and reasoning that does not mention factors demonstrably influencing the result. Treat it as useful output, not as an audit trail.
Why it matters
It is the cheapest quality improvement available on hard tasks, and it made reasoning models a distinct product category. The interpretability caveat matters for anyone tempted to use the visible reasoning as an explanation for a regulated decision — it doesn't qualify.
In practice
For multi-step work, ask for the reasoning explicitly or use a reasoning model. Read the steps to catch errors, but do not present them as an explanation of the model's actual process.