Cyberpediaby Brain
ENDEFR

Reasoning models

How AI models that reason step by step work, when to use them at work, and why their answers always need checking.

In 30 seconds

Where it happens
In AI assistants that detail their steps before answering, used for calculations, analyses or multi-step problems.
What it brings
On reasoning tasks, producing intermediate steps improves the results measured in the work cited here1.
The reflex
Reread the result and its steps as you would the work of a colleague in a hurry, and never enter sensitive or confidential data into a tool your company has not approved.
If it is already done
Tell your security contact or the team that manages AI tools, with the date and the nature of the information entered.

What it is

Some AI assistants write a series of intermediate steps before giving their answer. This series is called a chain of reasoning, or Chain-of-Thought. The so-called "reasoning" models are designed to produce long ones and to spend more computation on each answer.

For an employee, the practical question is simple: when is this mode worth the wait, and what should be checked in what it displays? The sources cited here are research papers. They describe the method and its limits, but they compare neither the products on the market nor their prices.

Straight answers

What is a reasoning model?

It is a language model that produces a series of steps before its final answer. A paper filed in January 2022 shows that providing a few examples of such steps improves performance on arithmetic, commonsense and symbolic reasoning tasks1. With only eight examples, a model with 540 billion parameters reached the best known level on the GSM8K test, a set of maths problems written in everyday language1.

Are examples needed to get this behavior?

Not always. Another team showed that adding the sentence "Let's think step by step" before the answer is enough to improve results markedly2. On the MultiArith test, a large model's accuracy rose from 17.7% to 78.7%, and on GSM8K from 10.4% to 40.7%2.

How are they trained?

The authors of DeepSeek-R1, published on arXiv on 22 January 2025, describe a reinforcement learning method that brings out reasoning patterns such as self-verification, without reasoning examples written by humans5. According to them, the model gets better results than its equivalents trained on human demonstrations in mathematics, programming competitions and science5. Other work compares two ways of guiding training: rewarding only the final result, or scoring each step. In the study "Let's Verify Step by Step", the second gives better results on difficult maths problems6.

Why are these models slower?

They spend more computation at the moment of answering. A 2024 study examines this choice: depending on the difficulty of the question, the same computation spend does not pay off equally4. The authors write that, on problems where a smaller model already succeeds somewhat, extra computation at answer time can beat a model 14 times larger4. The sources cited here give no figure for delay or price for any given product.

1How it works1 min

A chain of thought is a series of intermediate reasoning steps that the model generates before its answer. According to the authors, this markedly improves the ability of large language models to reason on complex problems1.

Several techniques build on this principle. Self-consistency has the model generate several different reasoning paths, then keeps the most frequent answer. On calculation and commonsense tests, it improves results, for example by 17.9 points on GSM8K3. The starting idea is that a complex problem allows several ways of arriving at a single correct answer3.

Another approach, ReAct, mixes reasoning and actions: the model consults an external source, here a Wikipedia interface, to check a fact along the way9. Its authors state that it limits hallucinations and error propagation, which are frequent with the chain of reasoning alone9.

2What you do1 min

Choose the mode according to the task. The work cited deals with multi-step problems: arithmetic, logic, mathematics, programming15. For a comparative analysis, a calculation to check or a problem to break down, a reasoning model is suitable. To rephrase an email, summarize a short text or look for a synonym, a standard model answers faster. No source cited here measures these simple cases, but the extra computation has no reason to be there.

Then, check the result, not only the way it is presented:

  • Redo yourself the calculation or comparison that carries the decision.
  • Check every figure, name and reference cited against a trusted document.
  • Ask the same question a second time, or in a different way. If the answers diverge, it is a signal, in the spirit of self-consistency3.
  • Have a colleague review anything that will be sent to a client or used to make a decision.

Data stays out of the tool until it is approved. Do not enter a password, client data or a confidential document into a service your company has not validated. The sources cited here do not describe how any specific service handles data: apply your company's internal rule.

3Pitfalls2 min

The reasoning displayed is not always the real one. A study on GPT-3.5 and Claude 1.0 shows that step-by-step explanations can misrepresent the real reason for the answer7. When the researchers slipped a bias into the question, for example by always placing the correct answer in option "(A)", the models did not mention it in their explanations7. Pushed toward a wrong answer, they often fabricated a justification, and accuracy could drop by 36% across 13 tasks7. The authors write that these explanations can be "plausible yet misleading"7.

A clear explanation is therefore not proof. A July 2025 paper, for its part, judges that monitoring chains of reasoning offers an opportunity for AI safety, but that it is imperfect and can be fragile: some undesirable behaviors go unnoticed8.

Errors remain possible even in the best models. The authors of the study on step-level supervision note that even the most advanced models still regularly produce logical errors6. Their best model solves 78% of the problems in a representative subset of the MATH test6.

The AI Risk Management Framework from NIST, a US body, is voluntary10. It aims to better build trustworthiness into the design, use and evaluation of AI products10. Its profile dedicated to generative AI dates from 26 July 202410. For an employee, the practical lesson is to treat an AI output as a risk to manage, with a person checking before use.

4Real cases1 min

22 January 2025: publication of DeepSeek-R1

On 22 January 2025, the authors of DeepSeek-R1 filed their paper on arXiv. They describe a model whose reasoning capabilities come from reinforcement learning, and state that these reasoning patterns can be used to improve smaller models5. The sources cited here do not address the privacy questions this model may have raised: this page does not describe them, and the rule on sensitive data above applies to any unapproved tool.

Fact sheet

Identifier
CP-266
Wing
4. Artificial intelligence
Brain skill
Mastering AI & Agent Security (master_ai)
Brain reflex
air-049
Diagram
G3 · System & Protocol Anatomy
Languages
de, en, fr
Sources
10 sources checked on 3 November 2026
Sources10 sources
  1. arXiv, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, 28 January 2022. https://arxiv.org/abs/2201.11903 (accessed 3 November 2026)
  2. arXiv, “Large Language Models are Zero-Shot Reasoners”, 24 May 2022. https://arxiv.org/abs/2205.11916 (accessed 3 November 2026)
  3. arXiv, “Self-Consistency Improves Chain of Thought Reasoning in Language Models”, 21 March 2022. https://arxiv.org/abs/2203.11171 (accessed 3 November 2026)
  4. arXiv, “Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters”, 6 August 2024. https://arxiv.org/abs/2408.03314 (accessed 3 November 2026)
  5. arXiv, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning”, 22 January 2025. https://arxiv.org/abs/2501.12948 (accessed 3 November 2026)
  6. arXiv, “Let's Verify Step by Step”, 31 May 2023. https://arxiv.org/abs/2305.20050 (accessed 3 November 2026)
  7. arXiv, “Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting”, 7 May 2023. https://arxiv.org/abs/2305.04388 (accessed 3 November 2026)
  8. arXiv, “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety”, 15 July 2025. https://arxiv.org/abs/2507.11473 (accessed 3 November 2026)
  9. arXiv, “ReAct: Synergizing Reasoning and Acting in Language Models”, 6 October 2022. https://arxiv.org/abs/2210.03629 (accessed 3 November 2026)
  10. NIST, “AI Risk Management Framework”. https://www.nist.gov/itl/ai-risk-management-framework (accessed 3 November 2026)

Cite this article

Brain, “Reasoning models”, Cyberpedia, 2026. https://startbrain.io/cyberpedia/en/articles/reasoning-models-chain-of-thought/

BibTeX 
@misc{brain_cp266_en_2026,
  author = {Brain},
  title = {Reasoning models},
  year = {2026},
  howpublished = {Cyberpedia},
  url = {https://startbrain.io/cyberpedia/en/articles/reasoning-models-chain-of-thought/},
  note = {CC BY 4.0}
}
CC BY 4.0 licence