Creating Models in CoT Prompting

1. Introduction : What is Chain of Thought?

Imagine asking a student, “What’s 47 × 23?” and they just blurt out “1081.” That’s a correct answer , but you have no idea how they got it.
Now imagine the same student says:

“47 × 23 means 47 × (20 + 3). First, 47 × 20 = 940. Then, 47 × 3 = 141. Adding them gives 1081.”

This second method - explaining the reasoning is Chain of Thought prompting for AI. You guide the model to think step-by-step, which often leads to better, more reliable answers.

2. Why Chain of Thought Works

From decades of studying both human cognition and AI behavior, I can tell you:

• Humans reason sequentially - breaking problems into smaller steps reduces error.

• AI models mimic patterns in data - when you ask for step-by-step thinking, you align the model’s output with structured reasoning patterns it has seen in training.

• Reasoning reduces hallucination - forcing intermediate steps makes it harder for the model to “jump” to wrong conclusions.

3. The Beginner’s Approach : Zero-shot CoT

Example:

Q: If it’s 3 PM now, what time will it be in 9 hours? Explain step-by-step.

Model:

Step 1: 3 PM + 9 hours = 12 hours later minus 3 hours.
Step 2: 3 PM + 12 hours = 3 AM.
Step 3: 3 AM - 3 hours = 12 AM (midnight).
Answer: 12 AM.

Key tip: Use phrases like

• “Let’s think step by step.”

• “Show your reasoning before the answer.”

4. Moving to Intermediate : Few-shot CoT

Here, you train the model in-context by giving examples of reasoning before the real question.

Example:

Q: 12 × 4
A: Step 1: 12 × 4 = (10 × 4) + (2 × 4) = 40 + 8 = 48.

Q: 15 × 3
A: Step 1: 15 × 3 = (10 × 3) + (5 × 3) = 30 + 15 = 45.

The model learns the “show steps, then answer” pattern.

5. Advanced Techniques

5.1. Self-Consistency + CoT

• Instead of one chain of thought, generate multiple reasoning paths and pick the most common answer.

• This drastically improves accuracy in math, logic, and reasoning-heavy tasks.

How to do it:

Generate 5 possible reasoning chains and pick the most consistent final answer.

The model will “vote” on the most common solution.

5.2. Tree of Thought (ToT)

• A logical extension of CoT.

• Instead of one linear chain, the AI explores multiple branches of reasoning , like a decision tree - before picking the best path.

• Useful in planning, creative writing, game strategies.

5.3. Program-aided CoT (PaCoT)

• Combine CoT reasoning with code execution for calculations.

• The model writes the reasoning, uses code to verify, then finalizes the answer.

• Example: math problems, data analysis.

6. Real-world Analogy

Think of CoT like assembling IKEA furniture:

• Without instructions (no CoT), you might get it right, but it’s a gamble.

• With instructions (CoT), each piece fits, and the chance of success skyrockets.