Chain of thought and reasoning models

What it is

Chain of thought is a way to get a language model to show its work. Instead of moving straight from a question to an answer, the model writes out the intermediate steps on the way there. A reasoning model takes this idea further: it is trained to produce a long stretch of internal reasoning before giving the reply a user sees. Both approaches rest on the same practical observation: on problems with several steps, working through the pieces one at a time tends to help.

How it works

Imagine a short word problem about a shop that starts with some apples, sells a few, and then receives a new delivery. A model that answers directly has to squeeze all of that into a single guess. With chain of thought, the prompt includes a worked example in which the author writes down each step before giving the result, and the model imitates that pattern on the new question. It restates what is known, handles one operation at a time, and only then states the answer. Each written step becomes context for the next one, so the reasoning on the page is part of the computation itself.

A reasoning model changes where this comes from. Rather than depending on examples written into the prompt, it is trained, usually with reinforcement learning, to generate its own extended reasoning and to spend more effort before it answers. This extra work done at the moment of answering is often called test-time compute. ^[2]

Where it came from

The prompting version of the idea was studied by Wei et al. at Google in January 2022. Their finding was that adding a few worked examples with step-by-step reasoning to the prompt greatly improved large models on math word problems. With eight such examples, the 540-billion-parameter PaLM reached state-of-the-art accuracy on the GSM8K benchmark. ^[1]

Trained reasoning came later. OpenAI's o1, released in September 2024, was the first widely released model of this type. On the AIME 2024 math exam it averaged 74% with one sample per problem, compared with 12% for GPT-4o. ^[2]

In January 2025, DeepSeek-R1 made its training recipe public. The authors described large-scale reinforcement learning with rule-based rewards on problems whose answers can be checked. The weights were released under the MIT license, so others could download and study the model. ^[3]

Where it is used

Reasoning models are strongest on math, coding and other tasks where answers can be checked. A checkable answer gives the training process a clear signal about what counts as success. ^[3] The approach is related to reinforcement learning from human feedback, the method used to shape chat assistants, but here the reward often comes from checking the result rather than from human ratings alone.

Limits

Reasoning is slower and costs more per answer, because the model generates far more text before the reply appears. That trade-off is the main reason to match the tool to the task: a quick factual question rarely needs a long internal chain. The benefit is also less clear on open-ended tasks where no answer can be checked. Finally, the written chain is itself generated text. It can contain mistakes even when the final answer looks plausible.