Chain Of Verification: Definition and Examples

Chain of Verification (CoVe) is a structured method designed to reduce hallucinations in language models. The principle is simple yet powerful: rather than blindly trusting an initial response, the model is guided through a systematic self-verification process in multiple steps.

The process unfolds in four phases. First, the model generates an initial response (baseline response). Next, it formulates verification questions targeting the factual claims contained in its response. Then, it answers each of these questions independently, without being influenced by its initial response. Finally, it compares the results of this verification with its original answer and produces a revised version, correcting any detected inconsistencies.

This approach leverages an important empirical finding: language models are often more reliable when answering short, targeted questions than when generating long, complex responses. By breaking down verification into micro-questions, the risk of factual errors and hallucinations is significantly reduced.

Introduced by researchers at Meta AI in 2023, Chain of Verification belongs to the family of chain-of-thought reasoning techniques but distinguishes itself by its specific focus on factual reliability rather than logical reasoning. It is particularly useful in contexts where information accuracy is critical.