Reasoning Models
LLMs trained with reinforcement learning to think through a long internal chain of reasoning before answering, trading extra inference time for accuracy on hard problems.
Also test-time compute · thinking models · extended thinking · inference-time scaling · o1
Instead of blurting out the first answer, the model is taught to quietly work through the problem, check itself and try again before replying. Harder problems get more thinking time.
Where it fits
+All relationships as a list
Is a kind of / part of
Uses / built on
Trained with
How it works
Chain-of-thought showed that writing out steps helps. Reasoning models take this much further: they are trained to reason at length, often thousands of tokens, before giving a final answer.
The training recipe, in broad strokes:
- Start from a strong pretrained LLM.
- Give it problems with checkable answers: math with known results, code with unit tests.
- Reward correct final answers with reinforcement learning (policy-gradient methods).
Without being told how, models learn useful habits: breaking problems down, trying another approach when stuck, double-checking. DeepSeek-R1 documented this openly in 2025.
This introduced a second way to scale. Beyond spending more compute on training (see scaling laws), you can spend more compute at test time: let the model think longer and accuracy keeps improving on hard problems.
Trade-offs: answers take longer and cost more, and the visible reasoning is not necessarily a faithful explanation of the answer. Many assistants now let you switch reasoning on or off, or set a thinking budget.
Everyday use cases
- Case 01
Competition math and science questions
Problems where checking several approaches pays off.
- Case 02
Complex coding tasks
Planning a multi-file change and verifying it against tests.
Famous examples
Strengths, weaknesses and when to use it
Strengths
- Large gains on math, coding and multi-step problems; can trade time for quality.
Weaknesses
- Slower and more expensive per answer; can overthink simple questions.
When to use itHard, verifiable, multi-step problems. Use a fast standard model for simple chat.
Compare with
Check yourself
- Q01
What is "test-time compute"?
Show answer
A: Spending more computation while answering, for example by reasoning longer
- Q02
How are reasoning models typically trained to reason?
Show answer
A: Reinforcement learning on problems whose answers can be checked