Reasoning Models: Why Some AI Systems Think Before Answering
Understand why writing out steps makes models smarter, and what reasoning models are trained to do.
Try this in under a second: a bat and a ball cost €1.10 together, and the bat costs €1.00 more than the ball. How much is the ball? Most people blurt “10 cents”. Spend ten seconds with a pencil and it turns out to be 5.
LLMs have their own version of this problem, for a very mechanical reason. Every token gets the same fixed amount of computation: one pass through the network. If the model must state the answer as its very first token, it gets exactly one pass to work it out. For a hard, multi-step problem, that simply isn’t enough.
But every token the model writes becomes part of its input. So if it first writes intermediate steps (“let the ball cost x, so the bat costs x + 1…”), each step gets its own pass, and later steps can read the earlier ones. The text becomes a scratchpad. In 2022, researchers found that simply adding “Let’s think step by step” to a prompt made models markedly better at maths and logic. This is chain of thought.
Reasoning models, from OpenAI’s o1 (2024) and DeepSeek-R1 (early 2025) onwards, take this much further. They’re trained with reinforcement learning on problems whose answers can be checked, such as maths and code: try, check the final answer, reward what worked. Along the way they learn to think at length, to try another route when stuck, and to double-check. Many apps hide that thinking or show only a summary.
It’s a new way to scale: instead of only making models bigger, let them spend more compute per question. The catch is cost and patience. Thousands of thinking tokens are thousands of passes, so answers take longer and cost more. For “what’s the capital of France?” thinking is wasted. For a tricky proof or a stubborn bug, it can be the difference between wrong and right.
Key takeaways
- Each token gets one fixed-size pass of computation, so a blurted answer gets just one.
- Writing intermediate steps gives the model more passes and a scratchpad it can reread: chain of thought.
- Reasoning models are trained with reinforcement learning on checkable problems to think at length.
- Thinking costs time and money; it pays off on hard multi-step problems, not on simple lookups.
Check your understanding
Go deeper (more technical)
Chain-of-thought prompting was popularised by Wei and colleagues (2022); “Let’s think step by step” comes from Kojima and colleagues (2022). DeepSeek-R1’s report showed that long, self-correcting reasoning can emerge from reinforcement learning with simple correctness rewards.
Accuracy on hard benchmarks tends to rise as models are allowed more thinking tokens: test-time compute scaling. A caution: the visible reasoning isn’t always a faithful account of how the model reached its answer. Studies (for example Turpin and colleagues, 2023) found explanations that sounded sensible but left out what actually swayed the answer.