Contrastive Chain of Thought (CCoT)

Expert

Teach correct logic using both right and wrong reasoning steps.

01 The Concept

Standard Chain-of-Thought (CoT) only demonstrates correct paths. However, models often get tripped up by common edge cases or anti-patterns (e.g. off-by-one errors in coding or sign errors in math).

Contrastive Chain of Thought (CCoT) guides the model by providing few-shot examples that contain both an incorrect reasoning path (highlighting the mistake) and a correct reasoning path.

Showing the model what not to do alongside what to do dramatically improves reasoning accuracy and prevents repetitive regressions.

02 Weak vs. Strong

EX 01Database Transaction Guard (Python)

You are a staff backend engineer. We need a function to transfer funds between accounts. Here is a contrastive demonstration of reasoning: --- INCORRECT REASONING EXAMPLE: Task: Transfer money securely. Bad Logic: - Deduct balance from account A - Add balance to account B - Write changes to DB Why it fails: If the system crashes after step 1, A's money is deleted and B never receives it. The operation is not atomic. --- CORRECT REASONING EXAMPLE: Task: Transfer money securely. Good Logic: - Open a DB transaction block - Try to deduct balance from A - Try to add balance to B - Commit the transaction - Roll back if any step throws an error Why it succeeds: Ensures atomicity; either both steps complete or neither does. --- Write a Python function to transfer $100 from account 12 to 34 using this contrastive logic.

→ Why it works

By demonstrating the atomic failure case directly, the model understands the exact structural vulnerability of non-atomic transfers, ensuring it constructs a transactional rollback block.

03 Key Points

01Explicit Pitfalls: Show typical logic bugs or misinterpretations in the few-shot demo.

02Explanation of Errors: The 'bad' reasoning path must explain why it fails, so the model learns the constraint.

03Corrective Contrast: Always pair the 'bad' example with the 'good' example for immediate correction.

04Logic Lock-In: Highly effective for code generation, mathematical reasoning, and policy compliance tasks.

05Error Prevention: Programmatically bounds the model's logic away from typical failure states.

04 Model-Specific Notes

Claude reacts strongly to contrastive examples. Highlight contrastive examples using XML tags like <incorrect_example> and <correct_example>.

05 For Your Role

It's like learning to drive: the instructor shows you what happens if you slam the brakes too late, and then teaches you how to brake smoothly.

Next: Prompt Drift & Calibration →