Thinking model with Chain of thought prompting

Introduction

Most AI models today are non-thinking by default — they just give you the answer directly without showing how they got there. But sometimes, especially for complex problems like math, reasoning, or planning, we want the model to think step-by-step before answering.

That’s where Chain-of-Thought (CoT) comes in — a technique where we train or prompt a model to reason through intermediate steps before producing the final result.

In this blog, we’ll walk step-by-step on how to turn a non-thinking model into a thinking model using Chain-of-Thought reasoning.

1. Understanding Non-Thinking vs Thinking Models

Non-Thinking Model

• Directly jumps to the answer.

• No reasoning process is visible.

• Works fine for simple tasks.

• Example:
  Q: What is 23 × 17?
  Model: 391

Thinking Model (with CoT)

• Breaks down the problem into small logical steps.

• Shows the reasoning before answering.

• Works better for multi-step reasoning tasks.

• Example:
  Q: What is 23 × 17?
  Model: First, 23 × 10 = 230. Then, 23 × 7 = 161. Add them: 230 + 161 = 391. So, answer = 391.

2. Why Chain-of-Thought Helps

• Better Accuracy: Especially for reasoning-heavy questions.

• Transparency: You can see why the model gave that answer.

• Debugging: Easier to find errors in logic.

• Generalization: Handles new, complex problems better.

---

3. Steps to Build a Thinking Model from a Non-Thinking Model

Step 1: Define the Task/

Decide where you need reasoning.
Example:

• Math word problems

• Logical puzzles

• Code debugging

• Planning tasks

---

Step 2: Use Prompt Engineering

You don’t need to retrain the whole model. Sometimes, just changing the prompt works.
Example:
Instead of asking: What is the total cost of 5 pens at ₹12 each and 3 books at ₹50 each?

Think step-by-step.
First, find the cost of the pens.
Then, find the cost of the books.
Finally, add them together.

Step 3: Teach the Model with Examples

Show the model how to think using few-shot learning.

Before:

Q: 2 + 2 = ?
A: 4

After

Q: If you have 2 apples and then buy 2 more, how many apples in total? A: First, start with 2 apples. Then, add 2 apples to them. That makes 4 apples.

Step 4: Apply Chain-of-Thought Prompting

This is where you explicitly tell the model to think step-by-step.
You can say:

• "Let's think step-by-step."

• "Explain your reasoning before giving the final answer."

---

Step 5: (Optional) Fine-Tuning

If you want a permanent thinking model, fine-tune the model with reasoning data.

• Collect Q&A pairs with reasoning steps.

• Train so it always produces steps first, then the answer.

---

Step 6: Post-Processing the Output

flowchart TD
    A[Start with Non-Thinking Model] --> B[Identify Complex Tasks]
    B --> C[Prompt Engineering with Step-by-Step Instruction]
    C --> D[Add Examples of Reasoning in the Prompt]
    D --> E[Model Generates Reasoning + Answer]
    E --> F[Optionally Fine-Tune for Permanent Thinking]
    F --> G[Deploy Thinking Model]

Sometimes you don’t want to show the reasoning to users, but still want the model to think internally.
You can:

• Get the reasoning from the model.

• Show only the final answer to the user.

Conclusion

By applying Chain-of-Thought prompting, you can transform a simple “answer-only” model into a reasoning powerhouse. The best part? You don’t always need heavy retraining — just smarter prompts and a few examples can get you there.

Next time your model answers “391” without explaining, teach it to think first, answer second.