Chain of Verification (CoVe)

Expert

Audit thoughts before answering to eliminate hallucinations.

01 The Concept

Chain-of-Verification (CoVe) is a prompting pattern designed to systematically reduce factual hallucinations. Instead of asking the model to write a response directly, the model is guided through four distinct phases:

1. Baseline Generation: Draft the initial response based on the prompt. 2. Verification Planning: Identify facts in the draft and generate targeted questions to verify them. 3. Verification Answering: Answer the planned questions independently (often in isolated contexts to prevent bias). 4. Synthesis & Revision: Combine the verified facts to construct a revised, error-free final response.

02 Weak vs. Strong

EX 01Historical Fact Summarization

You are a meticulous historical researcher. We need to summarize the major accomplishments of Marie Curie. Follow this 4-step Chain of Verification process: 1. <baseline_draft>: Write a concise 2-sentence draft of her key accomplishments and dates. 2. <verification_questions>: List 3 specific question-assertions from your draft that need verifying (e.g. specific years of Nobel prizes, names of elements discovered). 3. <verification_answers>: Answer each question in step 2 independently, referring ONLY to verified historical databases, not your draft. 4. <final_synthesis>: Produce the final, corrected summary of her accomplishments. Do not mention the draft or verification process in this final section.

→ Why it works

Explicitly divides the generation and verification stages. The model is forced to question its own draft assertions before finalizing, reducing the chance of carrying over errors.

03 Key Points

01Mitigate Hallucination: Proactively audit and correct facts before presenting final output.

02Fact Decomposition: Break down a complex answer into discrete, testable assertions.

03Isolated Execution: Answer verification questions without seeing the initial draft to avoid self-reinforcing errors.

04Surgical Revision: Only modify parts of the draft where verification failed.

05Production Pipelines: Highly effective when combined with structured retrieval or tool-based verification.

04 Model-Specific Notes

Claude excels at multi-step reasoning in a single turn if structured with XML tags. Use tags like <baseline_draft> and <verification_questions> to separate output phases.

05 For Your Role

Think of it like drafting a text message, double-checking the facts you wrote before hitting send, and rewriting the text if you made a mistake.

Next: Dynamic Few-Shot Selection →