Chain of Thought
Chain of thought is a prompting technique that elicits step-by-step reasoning from large language models by providing intermediate reasoning steps in the prompt.
Chain of thought (CoT) prompting was introduced by Wei et al. at Google in 2022. It involves providing a language model with a few examples that include not only the input and final answer, but also the intermediate reasoning steps that lead to that answer. This structure encourages the model to generate its own step-by-step reasoning when responding to a new query, rather than producing an answer directly.
The technique is particularly effective for tasks that require multi-step reasoning, such as arithmetic, commonsense reasoning, and symbolic manipulation. By decomposing a problem into a sequence of logical steps, CoT prompting improves the model’s accuracy and interpretability. The reasoning steps can be generated in natural language, making the model’s thought process transparent and easier to verify.
Chain of thought prompting does not require additional training or fine-tuning; it is a purely inference-time technique. It works by leveraging the model’s existing capabilities to generate coherent sequences of text, guided by the structure of the examples. Variants include zero-shot CoT, where the model is simply instructed to “think step by step” without any examples, and self-consistency, which samples multiple reasoning paths and selects the most consistent answer.
Why it matters
Chain of thought matters because it significantly improves the performance of large language models on complex reasoning tasks without requiring additional training. It makes model outputs more interpretable by exposing the reasoning process, which aids debugging and trust. In practice, CoT is used in applications like automated tutoring, code generation, and question answering, where accuracy and transparency are critical. It also provides a foundation for more advanced techniques like self-consistency and tree-of-thought prompting.
First appeared
Wei et al., Google, 2022.
Related terms
FAQ
How does it work?
Chain of thought works by including examples in the prompt that show the model how to break down a problem into intermediate reasoning steps. When asked a new question, the model generates its own step-by-step reasoning before producing the final answer, guided by the pattern established in the examples.
Does chain of thought require fine-tuning?
No, chain of thought is a prompting technique applied at inference time. It does not require any changes to the model’s weights or additional training. The model uses its existing knowledge to generate reasoning steps based on the structure provided in the prompt.
When should chain of thought be used instead of standard prompting?
Chain of thought is most beneficial for tasks that require multi-step reasoning, such as math word problems, logical puzzles, and complex question answering. For simple factual queries or tasks that do not benefit from explicit reasoning, standard prompting is often sufficient and more efficient.