What is Chain-of-Verification?
Chain-of-Verification (CoVe) was introduced in the paper Chain-of-Verification Reduces Hallucination in Large Language Models by Shehzaad Dhuliawala and colleagues at Meta AI. It targets one of the most persistent problems with language models: fluent, plausible statements that are simply wrong.
The key insight is that models answer short, focused questions more accurately than they produce long answers full of facts. CoVe turns a long answer into a set of short questions, checks each one, and uses the results to repair the original answer:
- Draft: generate a baseline response.
- Plan: write verification questions that test each factual claim in the draft.
- Verify: answer each question independently, without the draft in view, so the draft's errors are not simply repeated.
- Revise: produce the final answer, correcting or dropping claims the verification did not support.
The paper reports fewer hallucinations across list-based questions from Wikidata, closed-book question answering and long-form text generation.
When to Use Chain-of-Verification
Fact Lists
"Name politicians born in New York" style questions, where a few wrong entries hide among correct ones.
Biographies and Profiles
Dates, roles and affiliations are exactly the details models tend to invent.
Research Summaries
Check each claim about a paper, product or regulation before it reaches a report.
Content With Numbers
Articles and briefings where statistics, prices or dates must be right.
Technical Answers
Verify API names, flags and version-specific behaviour before trusting generated instructions.
Customer-Facing Answers
Support and sales answers where a confident wrong fact creates real cost.
How to Use Chain-of-Verification
- 1
Generate a baseline answer
Ask the question normally. Treat the result as a draft, not a final answer.
- 2
Plan verification questions
Ask the model to list one short, specific question per factual claim in the draft (who, when, where, how many).
- 3
Answer each question independently
Ideally in separate calls without the draft in context. In a single prompt, instruct the model to answer from its own knowledge, not from the draft.
- 4
Write the verified answer
Compare the draft with the verification answers. Keep consistent claims, fix contradicted ones and remove anything that could not be confirmed.
Prompt Examples
Question: Which programming languages were created at Bell Labs, and in which year was each first released? Step 1 — Draft: Answer the question. Step 2 — Plan: For each language and year in your draft, write one verification question. Step 3 — Verify: Answer each verification question on its own, from your knowledge, without looking at the draft. Step 4 — Final: Write the final answer. Keep only claims the verification supports and mark anything uncertain.
Answer each question independently and briefly. If you are not sure, say "uncertain" instead of guessing. 1. In which year was the C programming language first released? 2. Was C++ created at Bell Labs? 3. In which year was AWK first released?
Pros and Cons
| 🟢 Pros | 🔴 Cons |
|---|---|
| Measurably reduces hallucinated facts in the paper's benchmarks | Several passes: higher token cost and latency |
| Makes each claim explicit and checkable | Relies on the model's own knowledge; cannot fix what it does not know |
| Works with any model and needs no external tools | Independent verification needs separate calls for best results |
| Pairs well with retrieval for high-stakes content | Less useful for opinion, creative or reasoning-only tasks |
Chain-of-Verification Prompt Template
Copy this template and replace each {{ variable }}. It uses Prompt Edit's variable syntax, so you can paste it into the app and fill the fields from a form every time you reuse it.
Question: {{ question }}
Step 1 — Draft: Answer the question.
Step 2 — Plan: For each factual claim in the draft, write one short verification question.
Step 3 — Verify: Answer each verification question independently, from your own knowledge, without relying on the draft. Say "uncertain" if you are not sure.
Step 4 — Final: Rewrite the answer. Keep claims the verification supports, fix contradicted ones, and remove or flag anything uncertain.Frequently Asked Questions
What is Chain-of-Verification (CoVe)?
Chain-of-Verification is a technique from Meta AI (Dhuliawala et al., 2023) for reducing hallucinations. The model drafts an answer, plans fact-checking questions about the claims in that draft, answers those questions independently, and then writes a final answer that keeps only what the verification supports.
What are the four steps of Chain-of-Verification?
1) Generate a baseline response. 2) Plan verification questions that test the facts in that response. 3) Execute the verifications, answering each question independently so the answers are not biased by the draft. 4) Generate the final verified response, correcting or removing claims that failed verification.
Why must verification questions be answered independently?
If the model answers the verification questions while looking at its own draft, it tends to repeat the same mistakes. The paper found that the 'factored' variant, which answers each question separately without the original draft in context, removed more hallucinations than answering everything in one pass.
Does Chain-of-Verification eliminate hallucinations?
No. It reduces them, especially for lists of facts and long-form factual text, but it relies on the model's own knowledge. If the model does not know the correct fact, verification can confirm a wrong claim. For critical work, combine CoVe with retrieval from trusted sources or human review.
How is CoVe different from Reflexion or the Prompt Refinement Loop?
Reflexion critiques a failed attempt to learn a lesson for the next try, and the Prompt Refinement Loop improves the prompt itself. Chain-of-Verification targets factual accuracy within a single answer by turning its claims into questions and checking them.
Can I use Chain-of-Verification in ChatGPT or Claude?
Yes. The simplest version is one prompt with four labelled sections. For better results, run it as separate messages or API calls: draft, list questions, answer each question in a fresh context, then revise.
Origin & Sources
- Dhuliawala et al. (Meta AI) (2023). Chain-of-Verification Reduces Hallucination in Large Language Models. arXiv.