Chain-of-Thought (CoT)

Chain-of-Thought (CoT) prompting asks large language models to “think step by step” rather than jumping directly to an answer. By articulating intermediate reasoning, models avoid errors that arise from attempting complex reasoning in a single forward pass.

Key variants:

• Zero-Shot CoT: Simply adding “Let’s think step by step” to a prompt, which often improves accuracy significantly
• Few-Shot CoT: Providing examples that include reasoning steps, not just input-output pairs
• Self-Consistency: Running CoT multiple times and taking the majority answer
• Tree-of-Thought (ToT): Exploring multiple reasoning paths and backtracking when needed

CoT is a key driver of improved reasoning in modern LLMs. OpenAI’s o1 and o3 models internalize CoT-like reasoning during inference (test-time compute), representing a sophisticated evolution of this technique.