Learn AI /Large Language Models /Reasoning model
Reasoning model
- Intermediate
- Large Language Models
- Policy officers
- Technical readers
Also called: chain of thought, thinking model, test-time compute.
Back to the milestone you came from
In five seconds
What it is
A reasoning model is a language model designed for stronger performance on multi-step tasks, often using extra computation while answering, though visible steps may not reflect its internal process.
In plain words
Some chat systems try to answer hard questions by doing extra step-by-step work before they reply. People often call these reasoning or thinking models. This can help, but it can also make answers slower and it does not make them automatically correct.
An everyday example
You need to collect a parcel before the counter closes, refuel before a long drive and buy chilled food last. Instead of choosing the first route that comes to mind, you compare several orders, discard those that miss a constraint and check the remaining plan before setting out. More planning can improve the route, but it does not guarantee that the opening hours were remembered correctly. A written list of your checks is also not a complete record of everything you considered.
A student has human goals, background knowledge and conscious experience; a language model does not need those to produce step-like text. Also, a person can reason in ways that are not written down, while a model's behaviour comes from generating tokens according to learned patterns. So the analogy helps with the idea of intermediate work, but it should not be taken to mean the system reasons like a person.
Remember thisA reasoning model uses additional inference-time work for multi-step tasks. Its visible explanation may be useful, but it is not a faithful window into the model's internal process or proof that the answer is correct.
This term is contested
The terminology is unsettled. 'Reasoning model', 'thinking model', 'chain of thought' and 'test-time compute' overlap in practice, but different researchers and vendors use them with different scope and precision.
Understand it
Understand it
Why it matters
If you procure, regulate, use or evaluate AI systems, these terms affect practical choices. A system marketed as a reasoning or thinking model may perform better on harder tasks, but it may also be slower, costlier and still unreliable. The phrase chain of thought can refer to visible written steps, while test-time compute refers to extra computation more broadly, so vendors may sound similar while describing different features. This matters for benchmarking, safety review, procurement language, user expectations and policy discussions about transparency. A model that shows steps is not automatically more truthful, and a model that hides steps is not automatically less capable.
Short direct answer versus extra test-time compute
Show how the same kind of language model can answer either directly or through extra intermediate steps, and why that can improve hard tasks but increase delay.
01
Same input
Both paths start with the same prompt and context.
At the top centre, draw one input box labelled 'Prompt + context'. Split it into two arrows, one to the left column and one to the right column.
02
Direct path
A standard setup may generate an answer quickly with little or no deliberate intermediate work.
In the left column, draw two stacked boxes: 'Generate next tokens' then 'Final answer'. Add a small clock icon with label 'lower latency'.
03
Reasoning-oriented path
A reasoning-oriented setup spends more tokens on intermediate work before the final answer.
In the right column, draw three stacked boxes: 'Generate intermediate steps', 'Check or revise', then 'Final answer'. Add a larger token meter and a clock icon labelled 'more tokens, more latency'.
04
Possible benefit and risk
Extra work can improve performance on multi-step tasks, but the final answer can still be wrong.
At the bottom, draw a shared note box across both columns: 'Helpful for difficult tasks; not a guarantee of truth'. Use a green upward arrow near the right column labelled 'often better on hard tasks' and a warning icon near both columns labelled 'can still hallucinate or make logic errors'.
A side-by-side diagram compares two answer paths from the same prompt. The left path shows a short direct completion. The right path shows extra intermediate token generation, optional checking, and then a final answer. Read left to right to see how additional steps can help on difficult tasks while increasing latency.
Hovering or tapping each box reveals a short tooltip with examples: arithmetic, coding, planning, comparison. A toggle labelled 'show hidden steps as visible text' switches the right column's first box label between 'hidden intermediate work' and 'visible chain of thought'.
Describe this without the diagram
The diagram compares two ways a language model can answer the same prompt. On the left, the prompt goes to a short path: generate next tokens, then final answer. This is faster and uses fewer tokens. On the right, the prompt goes to a longer path: generate intermediate steps, check or revise, then final answer. This uses more tokens and usually takes longer. A note under both paths says extra intermediate work can help on difficult multi-step tasks, but neither path guarantees a correct answer. Another note explains that chain of thought refers to the intermediate written steps, while test-time compute refers more broadly to the extra computation used during answering.
With reduced motion: A static side-by-side comparison replaces animation. The left column shows the direct path with two boxes. The right column shows the reasoning-oriented path with three boxes and labels for higher token use and longer latency. Tooltips are replaced by small example captions under each box.
These related terms point to one idea, but they are not identical. A reasoning model is a language model tuned to do better on multi-step tasks such as maths, coding, planning or careful comparison. A thinking model usually means the same thing in product language. Test-time compute means giving the model more computation while it is answering, for example by letting it generate, check or refine intermediate text. Chain of thought usually means the intermediate written steps, whether shown to the user or kept hidden.
What goes in is still a prompt: your question, instructions and any context. What happens next may differ from a basic chatbot reply. Instead of producing a short answer immediately, the system may spend more tokens on intermediate work. A token is a small unit of text, such as part of a word. The model may draft steps, evaluate alternatives, or call tools in sequence before returning a final answer.
What comes out may be a better answer on tasks that need several steps, but there are trade-offs. More test-time compute often means higher latency, which means more waiting time, and sometimes higher cost. It also does not guarantee truth. A model can produce neat-looking steps and still make a mistake, or hide a weak guess behind confident wording.
These terms are often confused with next-token prediction. The base mechanism is still next-token prediction: the model generates text one token at a time. The difference is that a reasoning-oriented setup uses that same mechanism for longer or more structured intermediate work. It is also often confused with human understanding. The system may imitate step-by-step reasoning behaviour without having human-like awareness or reliable world knowledge.
How “Planning several errands with competing constraints” maps onto the real thing
| In the story | In the AI system |
|---|---|
| rough paper with intermediate steps | chain-of-thought-style intermediate text or hidden working |
| extra time for checking | more test-time compute |
| final answer sheet | the model's visible response |
| long maths problem | a multi-step task such as planning, coding or comparison |
A worked example
A workshop coordinator must order five agenda items. Registration must come first, the safety briefing must immediately follow registration, the practical exercise must follow the equipment check, and closing must come last.
Starting from: Give a valid agenda order and check every constraint before answering.
- The model extracts the ordering and adjacency constraints from the request.
- It may generate several candidate sequences, rather than committing to the first plausible one.
- It checks whether registration is first, the safety briefing follows it immediately, the equipment check precedes the practical exercise, and closing is last.
- It returns a valid order: registration, safety briefing, equipment check, practical exercise, closing.
Result: Extra inference-time computation can help with interacting constraints. It does not guarantee correctness, and a polished step-by-step explanation may not reveal which checks the system actually performed internally.
In a United Nations workplace
A gender adviser in a liaison office must clear a briefing note for the front office by close of business, ahead of informals the next morning. Several public statements are available only in provisional translation because the translation queue is backed up.
A reasoning model may spend extra computation grouping positions, testing alternative theme structures, tracing draft points to the supplied statements and flagging apparent inconsistencies. That can produce a better organised first draft, but any displayed rationale is generated text, not a reliable record of the model's internal processing.
Before you try thisDo not upload non-public statements, an uncleared briefing note or personal data to a tool that has not been approved for them. Provisional translations must remain clearly marked and checked by a qualified reviewer. Extra computation does not make mistranslations, omitted qualifications or invented consensus less consequential.
Where you may hear it
- A vendor demo to a procurement unit
- An ICT governance board reviewing model options
- A chief of section asking why one model takes longer to answer
- An inter-agency working group comparing outputs on a multi-step task
Avoid confusion
Avoid confusion
Often confused with
Next-token prediction
A language model produces text by repeatedly estimating probabilities for small pieces of text that could come next in the current sequence.
"Understanding" (what models do and do not do)
In language models, apparent comprehension usually means producing text that fits patterns in data and context, not reliably building a grounded grasp of the world like a person does.
Completion
The text a language model produces after a prompt, built token by token from patterns learned during training.
Hallucination
A generated answer that sounds plausible but includes false, unsupported or invented details.
Common misconceptions
Common misconception: If a model shows a long chain of thought, it must have reasoned correctly.
Correction: Not necessarily. Intermediate steps can help on some tasks, but they can also contain errors, post-hoc justifications or copied patterns that only look logical. The quality of the final answer still needs evaluation against evidence or calculation.
Common misconception: Reasoning models use a completely different mechanism from ordinary language models.
Correction: Usually the core mechanism is still token-by-token generation in a language model. The difference is often in training, prompting, tool use or allowing more test-time compute for intermediate work, not a complete break from next-token prediction.
Common misconception: Test-time compute and chain of thought are the same term.
Correction: They overlap but are not identical. Chain of thought usually refers to intermediate written steps. Test-time compute is broader: it includes any extra computation during answering, whether the steps are shown, hidden, checked, branched or combined with tools.
Common misconception: A thinking model understands problems the way a person does.
Correction: The label is product shorthand, not proof of human-like understanding. Such systems can produce useful multi-step outputs, but that does not mean they have awareness, stable world models or human judgement.
Quick check
Which statement best captures the relationship among these terms?
Option 2 is the best answer. In common use, reasoning model and thinking model often refer to systems designed for stronger multi-step performance. Chain of thought usually points to intermediate text, while test-time compute covers a wider set of extra-answering procedures, including hidden ones.
Go deeper
Go deeper
The simple version
Some chat systems try to answer hard questions by doing extra step-by-step work before they reply. People often call these reasoning or thinking models. This can help, but it can also make answers slower and it does not make them automatically correct.
The technical version
In current usage, a reasoning model is usually an LLM configured or trained to allocate more computation at inference time to multi-step problem solving. The base model still generates tokens autoregressively, but the system may produce intermediate scratchpad text, sample multiple candidate continuations, verify sub-results, use tools, or run revise-and-rerank loops before emitting the final completion. Chain of thought refers most narrowly to the intermediate textual steps. Those steps may be exposed to the user, summarised, or kept hidden. Test-time compute is the broader resource concept: additional forward passes, longer token budgets, branching, self-consistency-style sampling, or tool-mediated loops used during answering. Thinking model is informal product language and can blur these distinctions. Performance gains are often strongest on tasks with decomposable structure, but they come with latency and throughput costs and do not remove risks such as hallucination, arithmetic error, brittle planning or misleading rationales.
Reviewed external resources
EncyclopaediaReasoning model (opens in a new tab)Wikipedia
This link gives a broad overview of the term and related ideas, suitable for readers who want a quick starting point.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This course offers practical explanations of how language models work and suits learners who want hands-on technical grounding.
Deeperlink checked 2026-07-31
VideoLarge language models, explained visually (opens in a new tab)3Blue1Brown
This video gives an intuitive visual explanation of large language models and suits visual learners new to the topic.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Reasoning model (opens in a new tab)YouTube search
This search results page helps readers find multiple video explanations of reasoning models and suits people who want to compare teaching styles rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Reasoning model (opens in a new tab)Reddit search
This search results page helps readers see community questions and disagreements about the term and suits readers who want informal discussion with caution about accuracy.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A reasoning model uses additional inference-time work for multi-step tasks. Its visible explanation may be useful, but it is not a faithful window into the model's internal process or proof that the answer is correct.