Reasoning & Logic: Chain of Thought and Decomposition

In Module 3, we move into the realm of reasoning. LLMs are not inherently logical; they are probabilistic. However, with the right techniques, you can guide them to simulate complex logical reasoning.

Chain of Thought (CoT) is the most significant breakthrough in prompt engineering since few-shot prompting. The core idea is simple: Ask the model to "think step by step" before answering.

If you ask a complex math question directly, the model might guess the answer immediately, which often leads to errors.

Standard Prompt:

If I have 5 apples, eat 2, buy 3 more, and give half to my friend, how many do I have? Model Output (Guessing): 3

CoT Prompt:

If I have 5 apples, eat 2, buy 3 more, and give half to my friend, how many do I have? Let's think step by step.

Model Output:

1. Start with 5 apples.
2. Eat 2: 5 - 2 = 3 apples.
3. Buy 3 more: 3 + 3 = 6 apples.
4. Give half to friend: 6 / 2 = 3 apples. Answer: 3

By generating the intermediate steps, the model gives itself more "computational time" (more tokens) to reason correctly.

• Zero-Shot CoT: Just adding "Let's think step by step." (Simple, effective).
• Few-Shot CoT: Providing examples of step-by-step reasoning in the prompt. (Much more powerful for specific domains).

Tree of Thoughts (ToT) extends CoT by asking the model to explore multiple reasoning paths simultaneously.

Prompt Strategy:

"Imagine three different experts are answering this question. Each expert will write down 1 step of their thinking, then share it with the group. Then, they will critique each other's steps and decide which is the most promising path to follow."

This is great for creative writing, planning, or complex problem-solving where linear thinking might miss the best solution.

For very large tasks, CoT might still fail because the context window gets cluttered or the reasoning chain breaks. The solution is Decomposition.

Technique: Break the problem down into sub-problems explicitly.

Prompt:

To solve the user's request, first identify the key components needed. Then, solve each component individually. Finally, combine the solutions.

Example: "Write a Python script to scrape a website and save it to a database."

1. Sub-task 1: Write the scraping code.
2. Sub-task 2: Write the database schema.
3. Sub-task 3: Write the database insertion code.
4. Sub-task 4: Combine them.

In the next module, we will explore Persona & Context, learning how to make the model adopt specific roles and handle large amounts of information.