Chain of Thought and Self-Consistency: Reasoning Better with AI
Two techniques to guide an LLM's reasoning and catch hallucinations: ask it to show its work step by step, and have it cross-check against alternative paths.
Two techniques to guide an LLM's reasoning and get more reliable answers: Chain of Thought (CoT) and Self-Consistency. We'll use one running example: how much would you pay for three shirts at $25 each with a 20% discount? The answer is $60.
Chain of Thought (CoT)
It means asking the LLM to show its reasoning step by step before giving the final answer. Each step becomes visible and auditable, which makes errors easier to detect and improves accuracy. It's like solving a problem out loud:
20% of 25 = $5 → 25 − 5 = $20 → 20 × 3 = $60- Originally, you had to write "think step by step" in the prompt.
- Today, most models do this almost by default.
- "Reasoning models" are versions that spend more time on an internal chain of thought before answering.
- If a model jumps straight to the answer ($60), add "think step by step" to force it to verbalize the reasoning.
- Different valid chains can reach the same result. In the Copilot test: $75 full price → $15 discount → $60 final price.
How to apply it (e.g., in Microsoft Copilot):
- Open a new chat and ask the question.
- If the answer is direct, add "think step by step".
- Review the reasoning chain for errors.
Self-Consistency
It means asking the LLM to solve the same problem through several different paths, then comparing the results to pick the most common answer or confirm that they all agree.
How much would you pay for three shirts at $25 each with a 20% discount? Give me three methods to calculate this.| Method | Calculation | Result |
|---|---|---|
| 1. Discount on the total | 3 × 25 = $75 → 20% of 75 = $15 → 75 − 15 | $60 |
| 2. Discounted unit price | 25 − 20% of 25 = $20 → 20 × 3 | $60 |
| 3. Remaining percentage | 100% − 20% = 80% = 0.8 → 75 × 0.8 | $60 |
Why does it work? If all methods agree, your confidence increases. If one method disagrees, that's a clear signal of a hallucination or calculation error: you can inspect which chain failed or keep the majority answer.
When to use multiple methods
Not every problem needs three paths. It's worth it for:
- Mathematical or financial calculations, where a small error (e.g., a misplaced decimal) changes everything.
- Legal or regulatory interpretation with several valid angles.
- Decisions where you want to see the reasoning from different perspectives.
In math problems, all methods should give an identical result. In interpretive topics, the diversity of paths adds nuance instead of one single answer.
Combined: a double defense
| Chain of Thought | Self-Consistency | |
|---|---|---|
| What it does | Forces the model to show its work | Forces the model to cross-check itself with alternative paths |
| Benefit | Makes reasoning auditable | Makes inconsistencies visible |
Together they give you a double defense: show the work (CoT), then contrast it against alternative paths (self-consistency).
Exercise
Take the example about banking system laws (the one from grounding). Ask a question and request three methods to reach the answer. Check whether all three paths reach the same conclusion and note any discrepancies.
Key takeaways
- CoT = ask for step-by-step reasoning → transparency and fewer errors.
- Self-consistency = ask for multiple methods → validation and hallucination detection.
- Use both on high-stakes problems (math, finance, legal).
- Agreement between methods means more confidence; disagreement is a red flag to review.