Back to glossary

Reasoning model

Akıl yürütme modeli

D2 · Generative AI, capabilities and limitations

A reasoning model is a language model trained or configured to work through intermediate steps before answering complex tasks.

Review status: 2026-12-05

Technical explanation

Training can encourage checking and revising a solution; inference settings can allocate effort before the final response.

Conceptual boundaries

The label describes model behavior and training or inference choices, not human consciousness or a guarantee of correct reasoning.

Provider-neutral example

Illustrative: a model checks candidate solutions to a scheduling problem before presenting one for review.

Limitations

A longer generated explanation does not establish that it faithfully reports how the answer was reached.

Related concepts

Atomic claims and evidence

  1. 1.1DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
    Source
    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
    Source role
    Authoritative source
    Exact locator
    Abstract: reasoning abilities, reinforcement learning and verification
    Supported claim
    DeepSeek-R1 reports that reinforcement learning can encourage verification and other reasoning behaviors.
    Last verification
    Review due
    Scope limitation
    This is a result for the studied training approach, not a universal recipe.
  2. 2.1Google: Gemini thinking
    Source
    Google: Gemini thinking
    Source role
    Authoritative source
    Exact locator
    Thinking models; Best practices: thinking effort and task complexity
    Supported claim
    Gemini documents configurable thinking effort for its thinking models.
    Last verification
    Review due
    Scope limitation
    This describes a provider feature, not a universal interface for reasoning models.
  3. 3.1Anthropic: Reasoning models do not always say what they think
    Source
    Anthropic: Reasoning models do not always say what they think
    Source role
    Authoritative source
    Exact locator
    Results: models can use hints without acknowledging them in their chain of thought
    Supported claim
    Anthropic finds that generated reasoning can omit influences on an answer.
    Last verification
    Review due
    Scope limitation
    This is an experimental finding about the studied models, not every reasoning trace.