Reasoning Models

What are reasoning models?

Most chatbots answer like a colleague who blurts out the first thing that comes to mind. A reasoning model is more like the careful type who thinks before speaking: it gets the question, works through a chain of thought internally, then answers. That "thinking" burns extra compute, trading time for accuracy.

How are they different from ordinary models?

Speed vs. accuracy
Ordinary models are fast but stumble on math, logic and multi-step problems. Reasoning models are slower, but far steadier on the hard stuff.
You can watch the work
Many reasoning models show their thinking out in the open, so you can see each step — and exactly where it went wrong.

How are they trained?

Reasoning models are typically trained with reinforcement learning that rewards "thinking it through correctly and completely". Over time, the model learns to draft its work before committing to an answer — trained to be more careful.

What's wrong with them?

The classic failure is "overthinking": circling a simple question for way too long, burning time and tokens. Some models also second-guess themselves, rewriting a correct answer into a wrong one. So researchers are now trying to teach them when to think — and when to just answer.

Bottom line: reasoning models "think before they answer" — more compute buys more accuracy, at the cost of sometimes thinking too much.