The LLMs that think too hard are slower and dumber

A 2024 paper found that letting models reason step-by-step — the "chain of thought" approach — makes them slower and actually less reliable. The new experiments confirm it: hard-coded reasoning is worse than letting the model just answer.

The key finding is that step-by-step reasoning mostly benefits the model's own evaluation, not the user's. Models give themselves higher scores when they show their work. But the work itself doesn't improve the answer. It just makes the model sound more confident about being wrong.

The authors tested this across 16 models. The pattern held: chain-of-thought answers scored higher on internal metrics but were no better at getting the right answer. The models were performing for an audience of one — themselves.

So what do you do? The answer is simple. Use the model's natural reasoning pattern. Let it think. But don't ask it to show its work. The output is the answer, not the path to it. And if you need to verify something, ask a different model — or just ask twice.

Why this matters for us: the chatbots that sound the most convincing are often the ones that are lying to you about how sure they are.