“Chain-of-Verification Reduces Hallucination in Large Language Models,” submitted to arXiv on September 20, 2023 by Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston of Meta, attacked hallucination - the generation of plausible but false facts - with a structured self-checking procedure.
Chain-of-Verification, or CoVe, runs in four stages. First the model drafts an initial response. Second it plans a set of verification questions designed to fact-check the claims in that draft. Third it answers those verification questions independently - crucially, without seeing its own original answer, so the check is not biased toward confirming the draft. Fourth it generates a final, revised response that incorporates what the verification turned up. The independence of the verification step is the central design choice; answering the questions in isolation reduces the model’s tendency to simply repeat its earlier mistakes.
The authors reported reduced hallucination across several settings: list-based questions over Wikidata, closed-book MultiSpanQA, and long-form text generation.
Why business readers should care: Chain-of-Verification is a concrete recipe for making a model fact-check itself before delivering an answer. For any use where confidently wrong output is costly, building a verification pass into the prompt flow is a low-cost way to catch errors.