Chain of Thought is a prompting technique that consists of guiding a language model to break down a problem into intermediate steps before providing the final answer. Instead of jumping directly to the result, the model lays out its reasoning sequentially, which significantly improves its performance in tasks requiring logic, mathematics, or multiple deduction steps.
Its importance lies in the fact that it leverages an emergent capability of large models: reasoning better when they "think out loud." Simply adding instructions like "think step by step" is enough to activate this behavior. Common variants exist:
- Zero-shot CoT: reasoning step-by-step is requested without previous examples.
- Few-shot CoT: solved examples that demonstrate the reasoning are included.
A practical nuance: CoT works better in large models and does not always guarantee correct answers, as the model can generate plausible but erroneous reasoning. Even so, it makes the process more transparent and verifiable, making it easier to detect where the logic fails.