Zero-Shot Chain-of-Thought
Five words that transformed AI reasoning: “Let’s think step by step.” This single phrase — appended to any prompt without examples — unlocks the model’s latent ability to show its work, dramatically improving accuracy on math, logic, and commonsense tasks.
Introduced: Zero-Shot Chain-of-Thought was published in 2022 by Kojima et al. in their landmark paper “Large Language Models are Zero-Shot Reasoners.” The researchers discovered that simply appending the phrase “Let’s think step by step” to a prompt — with no worked examples whatsoever — could boost accuracy on the MultiArith benchmark from 17.7% to 78.7%. This finding revealed that large language models already possessed latent reasoning capabilities; they simply needed a trigger phrase to activate step-by-step thinking rather than jumping directly to an answer.
Modern LLM Status: The core insight of Zero-Shot CoT — that models reason better when instructed to show their work — has been natively absorbed into modern LLM architectures. Claude, GPT-4, and Gemini now engage in internal chain-of-thought reasoning by default on complex queries. The explicit “Let’s think step by step” trigger is no longer strictly necessary with frontier models, though it can still improve performance on edge cases and smaller models. Zero-Shot CoT’s lasting contribution is proving that reasoning could be elicited without curated examples — a principle now embedded in how all modern LLMs are trained and deployed.
Permission to Think Out Loud
Before Zero-Shot CoT, unlocking step-by-step reasoning in language models required carefully crafted few-shot examples — hand-written demonstrations showing the model how to break problems into steps. This was effective but labor-intensive. Kojima et al. asked a radical question: what if the model already knows how to reason, and just needs to be told it’s allowed to?
The answer was a single phrase. By adding “Let’s think step by step” to any prompt, the model shifts from generating an immediate answer to producing an intermediate reasoning chain first. It works because large language models have absorbed countless examples of step-by-step problem-solving from their training data — the trigger phrase simply activates that latent capability.
Think of it like asking someone to show their work on a math test. They already know the process — they just need the instruction to write it down instead of jumping straight to the final answer.
Traditional Chain-of-Thought prompting requires providing worked examples for every new task type — a math example for math, a logic example for logic, a coding example for code. Zero-Shot CoT eliminates this dependency entirely. A single, task-agnostic phrase works across arithmetic, symbolic reasoning, commonsense questions, and logical deduction. This universality made it one of the most widely adopted prompting discoveries in AI history.
The Zero-Shot CoT Process
Two-stage prompting that unlocks latent reasoning
Compose Your Question Normally
Write the problem or question exactly as you would in any standard prompt. No special formatting, no worked examples, no demonstrations. The question can be about math, logic, commonsense reasoning, or any task that benefits from deliberate thinking.
“A farmer has 15 sheep. All but 8 die. How many sheep does the farmer have left?”
Append the Trigger Phrase
Add “Let’s think step by step” (or a similar reasoning trigger) at the end of your prompt. This shifts the model from answer-generation mode into reasoning mode. Instead of pattern-matching to a likely answer, the model begins producing an explicit chain of intermediate thoughts.
“A farmer has 15 sheep. All but 8 die. How many sheep does the farmer have left? Let’s think step by step.”
Model Generates Reasoning Chain
The model produces a step-by-step reasoning trace before arriving at its conclusion. Each step builds on the previous one, making the logic visible and verifiable. This intermediate chain acts as a form of “scratch work” that guides the model toward the correct answer.
“The farmer starts with 15 sheep. The phrase ‘all but 8 die’ means 8 sheep survive. So the farmer has 8 sheep left.” — Without CoT, models often answer “7” by subtracting 8 from 15.
Extract the Final Answer
The reasoning chain concludes with a final answer that is grounded in the visible logic. Because each step is explicit, you can audit the reasoning, identify where errors might occur, and trust the conclusion more than a bare assertion. In production systems, this final answer can be extracted programmatically from the reasoning output.
The answer is 8 sheep. The reasoning chain made the tricky phrasing explicit, preventing the common subtraction error.
See the Difference
How a single phrase transforms model output
Direct Answer
A store has 23 apples. They sell 8 in the morning and receive a delivery of 15 more. Then they sell 12 in the afternoon. How many apples are left?
There are 20 apples left.
Zero-Shot CoT
Step 1: Start with 23 apples.
Step 2: Sell 8 in the morning: 23 − 8 = 15 apples.
Step 3: Receive 15 more: 15 + 15 = 30 apples.
Step 4: Sell 12 in the afternoon: 30 − 12 = 18 apples.
There are 18 apples left. Each arithmetic step is visible, making the answer verifiable at every stage.
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
48 US states now require AI transparency in key areas. Critical thinking remains your strongest tool against misinformation.
Zero-Shot CoT in Action
See how the trigger phrase improves reasoning across different domains
“A bookstore sold 45 novels on Monday and 32 on Tuesday. On Wednesday, they received a shipment of 80 new novels but also returned 15 damaged ones to the publisher. How many novels do they have compared to the start of Monday? Let’s think step by step.”
Step 1: Novels sold on Monday: 45 (reduces inventory by 45).
Step 2: Novels sold on Tuesday: 32 (reduces inventory by another 32).
Step 3: Total sold: 45 + 32 = 77 novels gone from inventory.
Step 4: Shipment received Wednesday: +80 novels.
Step 5: Damaged returns Wednesday: −15 novels.
Step 6: Net change: −77 + 80 − 15 = −12 novels.
Answer: They have 12 fewer novels than they started with on Monday.
“There are three boxes. One contains only apples, one contains only oranges, and one contains both. All boxes are labeled incorrectly. You pick one fruit from the box labeled ‘Both.’ You draw an apple. Which box contains which fruit? Let’s think step by step.”
Step 1: All labels are wrong, so the box labeled “Both” does NOT contain both fruits.
Step 2: I drew an apple from the “Both” box. Since this box cannot actually contain both (wrong label), and it has an apple, it must be the Apples-only box.
Step 3: The box labeled “Apples” is also mislabeled, so it doesn’t contain apples. It must contain either oranges or both.
Step 4: The box labeled “Oranges” is mislabeled too, so it doesn’t contain oranges. It must contain either apples or both. Since we already found the apples box, it must be the Both box.
Step 5: That leaves the box labeled “Apples” as the Oranges-only box.
Answer: “Both” label = Apples only. “Oranges” label = Both fruits. “Apples” label = Oranges only.
“A man pushes his car to a hotel and tells the owner he’s bankrupt. What is happening? Let’s think step by step.”
Step 1: The scenario describes pushing a car to a hotel and declaring bankruptcy. In the real world, this would be unusual — you don’t typically push cars to hotels.
Step 2: Consider whether this could be a game or metaphor. In Monopoly, players use car-shaped tokens, land on hotel properties, and can go bankrupt by paying rent.
Step 3: “Pushing a car” = moving a car token. “Hotel” = a property with a hotel. “Bankrupt” = can’t pay rent. “Owner” = the player who owns that property.
Answer: The man is playing Monopoly. He landed his car token on a property with a hotel and doesn’t have enough money to pay, so he declares bankruptcy.
When to Use Zero-Shot CoT
Maximum impact on reasoning tasks, minimal overhead
Perfect For
Arithmetic word problems, percentage calculations, unit conversions — any task where showing intermediate steps prevents errors.
Syllogisms, constraint satisfaction, if-then reasoning, and any task requiring the model to chain logical inferences rather than guess at conclusions.
When you need reasoning improvement immediately without time to craft few-shot examples — just append the trigger phrase and go.
Models that lack native chain-of-thought reasoning benefit the most — the trigger phrase can produce dramatic accuracy gains on these systems.
Skip It When
Questions like “What is the capital of France?” need no reasoning chain — the trigger phrase just adds unnecessary tokens to a single-hop answer.
Step-by-step reasoning generates significantly more tokens. When you need fast, concise answers — chatbots, autocomplete, real-time systems — the extra tokens add unwanted delay.
Writing poetry, brainstorming ideas, or generating fiction — these benefit from fluency and spontaneity, not rigid step-by-step decomposition.
Use Cases
Where Zero-Shot CoT delivers the most value
Financial Calculations
Compound interest, tax scenarios, budget allocations, and investment return calculations where each arithmetic step must be visible and verifiable.
Code Debugging
Tracing program execution line by line, identifying where logic fails, and explaining code flow by walking through each operation step by step.
Homework and Tutoring
Teaching students how to approach problems by making the reasoning process explicit — showing not just the answer but the thinking behind it.
Medical Triage
Walking through symptom assessment, differential diagnosis, and decision trees where each reasoning step must be transparent and auditable by a clinician.
Strategic Decision-Making
Weighing trade-offs, analyzing multi-factor decisions, and evaluating risks by forcing the model to consider each factor explicitly before reaching a conclusion.
Compliance Verification
Checking whether a process meets regulatory requirements by reasoning through each rule, condition, and exception in a visible, checkable sequence.
Where Zero-Shot CoT Fits
The simplest reasoning trigger in a family of increasingly structured techniques
Zero-Shot CoT occupies a unique position in the prompting landscape: it delivers a large fraction of few-shot Chain-of-Thought’s accuracy improvement at essentially zero cost. No examples to write, no format to maintain, no task-specific tuning. When you need a quick reasoning boost and don’t have time to craft demonstrations, Zero-Shot CoT is the highest-leverage single change you can make to a prompt.
Related Techniques
Build on Zero-Shot CoT with these complementary techniques
Unlock Step-by-Step Reasoning
Try adding reasoning triggers to your own prompts or build enhanced prompts with our tools.