AI Trend Research

← AI glossary

AI glossary · Basic · Language models

Reasoning model

A reasoning model is a large language model (LLM) trained not to answer right away, but to first go through a long 'thinking process' that works through the problem step by step before giving a final answer.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-08
In plain words It is like the difference between a student who blurts out an answer from mental math and a student who writes out the working on scratch paper and double-checks it before answering. However, the model's 'thinking' is not the same as human thought; it is a process of generating more text (tokens) before producing the answer.

How it works

A regular LLM starts writing the answer as soon as it receives a question. A reasoning model, before writing the answer, first generates text in which it breaks down the problem, forms hypotheses, checks intermediate results, and fixes mistakes. This ability is mainly developed through reinforcement learning. The model is given tasks whose correctness can be checked, such as math problems or code, and is rewarded for solution approaches that get the right answer, so it learns a 'habit of thinking long and carefully'. This is why it is closely tied to 'test-time compute', which spends more computation at answer time.

Regular LLMs vs. reasoning modelsQuestion → Regular LLM; Regular LLM → Answers right away; Question → Reasoning model; Reasoning model → Thinking process; Thinking process → Final answerRegular LLMReasoning modelQuestionRegular LLMAnswers right awayQuestionReasoning modelThinking processBreak down · check ·fixFinal answer
Regular LLMs vs. reasoning models The top row is a regular LLM that answers the question right away; the bottom row is a reasoning model that goes through a step-by-step thinking process before answering.

Why it matters

It showed that, besides making models bigger or adding more data, performance can be raised by having the model think longer when answering. The effect is especially large in math, science, programming, and planning problems that require multiple steps. It also underpins services that must carry out multiple steps on their own, such as coding agents and AI agents. This trend is also why many services now offer separate 'fast mode' and 'deep thinking mode' options.

Things to keep in mind

Because the thinking process is longer, answers come out more slowly, and more tokens are processed, which raises costs. For simple questions, the model may overthink and become inefficient. You cannot assume that the thinking process shown on screen faithfully reflects the computation that actually happened inside the model. And thinking longer does not make hallucinations go away; the model can still reach a wrong answer at the end of a plausible-looking solution. Note that the Korean word for 'reasoning' in 'reasoning model' (추론) is the same word often used for 'inference', meaning running a model, but the two are different concepts.

Example

Several companies have released reasoning models or reasoning modes, such as OpenAI's o series, DeepSeek-R1, and the extended thinking feature of Anthropic's Claude. In chat services, users who turn on a 'thinking' option can see the model take a moment to deliberate before answering, and developers can adjust how much the model spends on thinking through the API to balance speed and cost against accuracy.

Related terms

CoT Test-time compute RLVR Reinforcement learning LLM Inference