zero-shot chain-of-thought

**Zero-shot chain-of-thought (Zero-shot CoT)** is the remarkably simple technique of appending the phrase **"Let's think step by step"** (or a similar instruction) to a prompt — without providing any reasoning examples — to trigger the language model to generate its own step-by-step reasoning before producing a final answer. **The Discovery** - Standard **few-shot CoT** requires carefully crafted reasoning examples in the prompt — effective but labor-intensive to create for each task. - Researchers discovered that simply adding **"Let's think step by step"** to the end of a zero-shot prompt (no examples at all) dramatically improves reasoning performance. - This single phrase can improve accuracy on math and logic tasks by **40–70%** compared to standard zero-shot prompting. **How Zero-Shot CoT Works** - **Without CoT**: "What is 23 + 47 × 2?" → Model often gives wrong answer by misapplying order of operations. - **With Zero-Shot CoT**: "What is 23 + 47 × 2? Let's think step by step." → Model responds: ``` Step 1: First, compute 47 × 2 = 94 Step 2: Then, add 23 + 94 = 117 Answer: 117 ``` **Two-Stage Process** 1. **Reasoning Extraction**: Append "Let's think step by step" → model generates a reasoning chain. 2. **Answer Extraction**: After the reasoning, prompt "Therefore, the answer is" → model produces the final answer. - Some implementations use both stages explicitly; others let the model naturally conclude with an answer. **Why It Works** - The phrase **activates reasoning patterns** learned during pretraining — the model has seen many examples of step-by-step reasoning in its training data. - Without the prompt, the model defaults to **pattern matching** or **direct recall** — which often fails for problems requiring multi-step logic. - The instruction makes the model **allocate more computation** (more tokens) to the problem before committing to an answer. **Effective Trigger Phrases** - "Let's think step by step" — the original and most studied. - "Let's work this out step by step to be sure we have the right answer." - "Let's solve this carefully." - "Think about this step by step before answering." - Research shows the exact phrasing matters — some variations work better than others for specific models. **Limitations** - **Less Effective Than Few-Shot CoT**: On many benchmarks, few-shot CoT with well-crafted examples still outperforms zero-shot CoT. - **Model Size Dependent**: Zero-shot CoT primarily works with large models (>100B parameters). Smaller models may produce incoherent reasoning. - **Task Dependent**: Works well for math, logic, and commonsense reasoning. Less effective for creative tasks or tasks requiring domain-specific procedures. - **Unfaithful Reasoning**: The model may generate plausible-looking but logically flawed reasoning — the presence of steps doesn't guarantee correctness. **Practical Impact** - Zero-shot CoT is the **most cost-effective reasoning improvement** available — it requires no example crafting, no fine-tuning, and works across many tasks. - It's become a **standard baseline** in prompt engineering — virtually every complex prompt now includes some form of "think step by step" instruction. Zero-shot chain-of-thought is one of the **most influential discoveries** in prompt engineering — a single phrase that unlocks latent reasoning capabilities, demonstrating that how you ask is as important as what you ask.