Chain-of-Thought

Chain-of-Thought (CoT) is a prompt design technique that, when asking a large language model (LLM) to solve a complex problem, instructs the model to explicitly generate not just the final answer but also the intermediate "steps of thinking" that lead to it. It is a prompting approach that has attracted significant attention for improving accuracy on multi-step tasks such as reasoning and calculation. With a standard prompt, the model aims directly for an output. Chain-of-Thought, by contrast, guides the model to "first organize its thoughts, then reason step by step toward a conclusion." This draws out the model's internally stored knowledge and reasoning capabilities in a sequential manner, enabling more accurate and logically sound responses. Example (arithmetic problem): Standard prompt: "If you buy 3 apples at 120 yen each, how much do you pay?" → Model output: "360 yen" Chain-of-Thought prompt: "If you buy 3 apples at 120 yen each, how much do you pay? Please include your thinking steps in the answer." → Model output: "3 apples at 120 yen each comes to 120 × 3 = 360 yen. So the answer is 360 yen." By having the model output its intermediate reasoning, numerous studies have reported performance improvements on tasks such as complex question answering, mathematical reasoning, logic problems, and legal judgments. Chain-of-Thought can be combined with other high-performing prompting techniques such as Few-shot Learning, Tool-using Agents, and ReAct (Reasoning + Acting), and is actively being studied as a key technology for enhancing the reasoning capabilities of AI agents.