What are zero-shot, few-shot, and chain-of-thought prompting?

Zero-shot: "Classify this review as positive or negative: ..." Works well for common tasks with strong instruct models.

Few-shot

Review: "Battery died in a day" → negative
Review: "Superb camera, fast delivery" → positive
Review: "Okay screen but laggy" → ?

• Useful for custom formats, labels and edge cases.
• Tips: diverse examples, balanced labels, consistent format. Examples strongly influence output, so models may copy surface features.

Chain-of-thought

• "Let's think step by step" (zero-shot CoT, Kojima et al. 2022), or few-shot examples that include reasoning (Wei et al. 2022).
• Why it helps: the model gets more computation (tokens) to work through intermediate steps, and each step conditions the next.
• Self-consistency: sample several CoT answers and take a majority vote. More accurate, more expensive.

With reasoning models (models trained to think before answering), explicit "think step by step" prompting matters less. Focus instead on a clear task definition and the context.

Common mistakes

• Using CoT for simple lookups (adds latency and cost for no gain).
• Showing raw CoT to end users when it may contain errors or irrelevant text. Often the final answer alone is better.

Related

• Previous: Q11. What is hallucination, and why do LLMs hallucinate?
• Next: Q13. What does "7B" or "70B parameters" mean, and how much memory does a model need?
• LLM Fundamentals
• LangChain