Tool

Chain of Thought (CoT)

Why Chain of Thought Is Needed

When a model solves a math problem, a correct answer doesn't always mean it truly understood the problem — it might have skipped steps and guessed right, or memorized the answer.

For many tasks, what matters isn't "is the final answer correct?" but "are the intermediate steps correct?":

  • Extracting conditions from the problem
  • Judging relationships between conditions
  • Choosing the right method
  • Checking for self-contradiction

If you let the model output an answer directly, errors in the middle are hard to spot. CoT solves this: making the model show its reasoning process first, then give the conclusion.


What Is Chain of Thought

One-line definition: CoT (Chain of Thought) is a method that guides models to展开 reasoning step by step, making intermediate judgments explicit to reduce skipping and guesswork.

Intuitive example:

User question: A store has 5 apples. They sold 3, then restocked 2. How many now?

Plain answer: 4

CoT style answer:
1. Started with 5
2. After selling 3, 2 remain
3. After restocking 2, it becomes 4
4. The final answer is 4

Why it works: Breaking down the thinking process makes errors easier to spot and easier to correct.


How to Do It: Common CoT Forms

Zero-shot CoT: No examples, just add "please think step by step." Best for quick reasoning enhancement with minimal implementation cost.

Few-shot CoT: Give several standard reasoning examples first, then let the model mimic them. Best for fixed question types or format requirements. This aligns with Zero-shot / Few-shot thinking.

Self-Consistency: Generate multiple reasoning paths, then pick the most consistent result. Best for high-value reasoning tasks, but consumes more tokens.

Hidden Reasoning: Use CoT internally for analysis, but only output conclusions and key reasoning externally. Better suited for production environments.


When to Use / Not Use

Scenarios suited for CoT:

  • Math and symbolic reasoning: calculations, comparisons, formula transformations
  • Logical judgment: conditional derivation, elimination, causal analysis
  • Code debugging: locating errors, analyzing call chains, finding root causes
  • Agent planning: deciding what to check first, what to do next, when to stop

Scenarios where CoT is less needed:

  • Pure factual recall: e.g., "What's the capital of China?" — no reasoning needed
  • Simple single-turn Q&A: no need to break down steps

CoT makes the model "think first," but after "thinking" it still needs to "do" — this leads to ReAct: the closed loop of think + act + observe.


Remember this: CoT makes models "think before answering" — making intermediate steps explicit so errors are easier to spot and correct.

Related terms: ReAct · Zero-shot / Few-shot