Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Reasoning Model

nounLarge Language Modelsalso reasoning LLM, large reasoning model, thinking model

In one line

A reasoning model is an artificial intelligence model trained to work through a problem in intermediate steps, often spending extra computation at answer time, before it gives a final response.

A reasoning model is an artificial intelligence model that works through a problem in intermediate steps, often using extra computation at the moment of answering, before it produces a final response. Where a standard large language model tends to answer in one pass, a reasoning model first generates a long internal chain of thought in which it plans, tries approaches, checks results and corrects mistakes. This tends to improve accuracy on mathematics, programming, science and multi-step logic, at the price of more latency and cost.

Background

Early language models were good at fluent text but weak at problems that need several logical steps. In 2022, researchers showed that asking a model to write out its steps, a technique known as chain-of-thought prompting, raised accuracy sharply on arithmetic and commonsense benchmarks. Later work trained models to produce such steps on their own rather than relying on a prompt. In 2024, OpenAI introduced its o1 series, trained with reinforcement learning to think at length before replying. In 2025, DeepSeek-R1 and other open models showed that reasoning behavior could be developed with large-scale reinforcement learning and rule-based rewards, and most major laboratories now offer reasoning models or hybrid models with an adjustable thinking mode.

How reasoning models work

The central idea is to let the model spend more computation on a hard question. This is known as test-time compute or inference-time scaling. The model generates reasoning tokens, a scratchpad that may be hidden or summarized for the user, then writes the answer. Longer or repeated reasoning usually helps up to a point.

  • Chain of thought: a sequence of intermediate steps leading to the answer.
  • Self-checking: the model reviews its work, notices an error and backtracks.
  • Search and sampling: some systems sample several reasoning paths and choose the most consistent answer, or use a verifier to rank them.
  • Tool use: reasoning can be combined with calculators, code execution or search.

Training methods

Reasoning models start from a pretrained transformer language model and are then shaped by additional training.

  1. Supervised fine-tuning on examples of step-by-step solutions, sometimes written by people and sometimes generated by other models.
  2. Reinforcement learning with rewards for correct final answers on problems where the answer can be checked automatically, such as mathematics with a known result or code that passes tests.
  3. Process supervision, in which a reward model scores each step rather than only the final result.
  4. Distillation, in which a smaller model learns from the reasoning traces of a larger one. See model distillation.

The training signal works best in domains with verifiable answers, which is one reason progress has been fastest in mathematics and programming.

Capabilities

Reasoning models score much higher than earlier models on competition mathematics, graduate-level science questions, coding challenges and planning puzzles. They are also better at following complex constraints, auditing their own work and decomposing a task. In AI agents, they are often used as the planner that decides which steps to take.

Limitations

  • Cost and delay: long reasoning uses many tokens, so answers are slower and more expensive.
  • Overthinking: simple questions can receive needlessly long reasoning.
  • Faithfulness: the written reasoning may not reflect the real process that produced the answer, so it should not be treated as a perfect explanation.
  • Errors persist: a hallucination can still appear, and a confident chain of logic can lead to a wrong result.
  • Uneven gains: improvements are smaller on open-ended tasks without a clear correct answer.

Evaluation

Reasoning is measured with benchmarks such as GSM8K and MATH for mathematics, GPQA for science questions, competitive programming sets and software engineering tasks. Because benchmarks saturate quickly and training data can leak, researchers also build new, held-out problem sets and test whether performance holds when a question is rephrased.

Applications

Typical uses include debugging and writing code, solving technical problems, analyzing contracts and financial models, planning multi-step tasks, tutoring in mathematics and science and assisting researchers with derivations. Users often choose a reasoning model for hard tasks and a faster standard model for routine ones.

A worked example

Suppose a user asks a reasoning model to schedule five meetings for four people with conflicting calendars. The model writes out the constraints, tries an arrangement, finds that one person is double-booked, backtracks, tries a second arrangement and verifies each constraint before reporting the schedule. A standard model might have produced a plausible schedule on the first try, with one conflict left unnoticed. The reasoning model took longer and used more tokens, but its answer was easier to trust because it had been checked.

Test-time compute and scaling

A key finding behind reasoning models is that accuracy can improve not only by training larger models on more data but also by letting a model compute for longer on each question. Studies show that performance on hard problems often rises steadily as the number of reasoning tokens grows, although the gains flatten and then stop. This gives developers a new control: a budget for thinking. Many services let users set a low, medium or high effort level, trading speed and cost against accuracy. It also changes how systems are priced, because reasoning tokens are usually billed even when they are not displayed.

Safety and oversight

Visible reasoning offers a possible window into how a model reaches its answers, which researchers use to monitor for misbehavior such as attempts to exploit a flawed reward. Yet the trace is not guaranteed to be faithful, and training the model to produce reasoning that looks good can reduce its usefulness for oversight. Reasoning models also raise questions about AI safety because stronger planning ability can make errors and misuse more consequential, so developers test them with extra care.

Choosing between model types

A practical rule is to use a reasoning model when a task needs several dependent steps, when an error is costly or when the answer can be verified, and to use a standard model for drafting, summarizing and quick lookups. Hybrid models allow one system to handle both by switching thinking on or off.

FAQs

Do reasoning models actually think? They produce intermediate computation that functions like reasoning, but whether it matches human thinking is debated. It is best described as a learned procedure for working through problems.

Are reasoning models always better? No. They help most on hard, multi-step problems and can be slower and costlier on simple requests.

Can I see the reasoning? Some providers show a summary or the full trace, and others keep it hidden.

Key facts

  • Chain-of-thought prompting, described by Wei and colleagues in 2022, showed that asking a language model to produce intermediate steps improves results on arithmetic and logic tasks.1
  • OpenAI released its o1 reasoning models in 2024, trained with reinforcement learning to produce long internal chains of thought before answering.2
  • DeepSeek-R1, published in 2025, reported that reasoning behavior can emerge from large-scale reinforcement learning with rule-based rewards.3

Related concepts

Quick checkWhat makes a reasoning model different from a standard language model?Show answer

It is trained to produce and use a long chain of intermediate reasoning steps, often at inference time, before giving a final answer, which improves accuracy on hard problems.

Sources

  1. Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, https://arxiv.org/abs/2201.11903
  2. OpenAI, Learning to Reason with LLMs, https://openai.com/index/learning-to-reason-with-llms/
  3. DeepSeek-AI, DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, https://arxiv.org/abs/2501.12948

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue