Chain-of-Verification (CoVe)
The method comes from the 2023 paper Chain-of-Verification Reduces Hallucination in Large Language Models. It has four steps. The model drafts a baseline response, plans verification questions that test the draft's claims, answers each question independently so the answers aren't biased by the draft, and then writes a final, verified response. The authors report fewer hallucinations on list-based questions, closed-book question answering and long-form text.
The independence step matters most. If the model answers the checks while looking at its own draft, it tends to repeat its mistakes, so run each verification question as a separate call, or at least without the draft in view. It works even better when the checks are answered from source documents (grounding) instead of from memory. It's related to reflection and self-consistency, and Qurtoo's article on the verification prompt pattern shows how to build it into everyday calls.
Example
Step 1 (call 1): Answer the question: "Which of our suppliers are based in Ohio?"
Step 2 (call 2): List one yes/no verification question for each supplier in the draft, such as "Is Acme Corp's registered address in Ohio?"
Step 3 (one call per question, without the draft): Answer each verification question using only the supplier records below.
Step 4 (call 4): Rewrite the original answer, keeping only suppliers whose check came back yes.When to use it
- Answers that list facts, names or figures
- Long-form drafts where one wrong claim undermines the rest
- When you can afford several calls per answer
When not to use it
- Creative writing or opinions with nothing to verify
- Latency-sensitive paths
- When the model has no reliable way to check a fact; give it sources first

