Chain-of-thought reasoning

Chain-of-thought reasoning is a prompting technique in which a LLM generates explicit intermediate reasoning steps before producing a final answer, rather than jumping directly from question to conclusion. The technique was introduced by Wei et al. (2022) and demonstrated that simply appending the phrase "let's think step by step" to a prompt could dramatically improve performance on mathematical, logical, and commonsense reasoning tasks. The effect is not merely cosmetic: when the model is forced to externalize its reasoning process, the resulting text itself becomes part of the context, creating a feedback loop in which each reasoning step constrains the next through the attention mechanism.

From a systems perspective, chain-of-thought reasoning is a form of iterative refinement in which the system uses its own outputs as inputs to a subsequent computation. The context window functions as a working memory, and the reasoning chain is a trajectory through a high-dimensional state space in which each token is a step toward an attractor basin representing the correct answer. The technique reveals that the model's reasoning capability is not uniformly distributed across all prompts but is activated by specific structural conditions — namely, the presence of intermediate steps that bridge the gap between problem and solution.

Chain-of-thought reasoning is not a technique that makes LLMs smarter. It is a technique that reveals they were already capable of multi-step reasoning, but only when the prompt structure creates a path through state space that the model can follow. The implication is unsettling: the intelligence is in the trajectory, not the model. The model is a landscape; the prompt is the hiker. We have been crediting the mountain for the climb.