Self-criticism

In short: Generate the answer, write checking questions, and revise.

What is: The model generates the answer, then formulates independent verification questions, answers each WITHOUT looking at the original answer, and revises based on them. One of the best defenses against invented facts.

When to use: Answers with facts, dates, numbers, names, where hallucination is costly.

When to avoid: Tasks with no verifiable facts (opinion, free creation).

💡 Tip: Do the verification in a separate step (or separate call) so the model isn't contaminated by the initial answer.

In short: Verify candidates, estimate confidence, reconstruct the problem.

What is: Self-Verification generates several solutions and verifies each by reconstructing the problem. Self-Calibration asks the model to estimate confidence in its own answer. Reversing CoT (RCoT) reconstructs the prompt from the answer to detect inconsistency.

When to use: Self-Calibration is the most practical: use it to route (low confidence → review or escalate to a human).

When to avoid: When you won't act on the confidence/verification, it's just overhead.

💡 Tip: Use confidence as an automation trigger: <70% → second opinion or human review.