驗證鏈機制降低大型語言模型中的幻覺

摘要

在大型語言模型中，產生看似合理但事實不正確的資訊，稱為幻覺，是一個尚未解決的問題。我們研究語言模型在給出回應時審慎考慮並更正錯誤的能力。我們開發了「驗證鏈」（CoVe）方法，模型首先（i）起草初始回應；然後（ii）計劃驗證問題以核實起草的內容；（iii）獨立回答這些問題，以避免受其他回應的影響；最後（iv）生成最終經過驗證的回應。在實驗中，我們展示了CoVe在各種任務中減少幻覺的效果，包括來自Wikidata的基於列表的問題、閉書式MultiSpanQA和長文本生成。

English

Generation of plausible yet incorrect factual information, termed hallucination, is an unsolved issue in large language models. We study the ability of language models to deliberate on the responses they give in order to correct their mistakes. We develop the Chain-of-Verification (CoVe) method whereby the model first (i) drafts an initial response; then (ii) plans verification questions to fact-check its draft; (iii) answers those questions independently so the answers are not biased by other responses; and (iv) generates its final verified response. In experiments, we show CoVe decreases hallucinations across a variety of tasks, from list-based questions from Wikidata, closed book MultiSpanQA and longform text generation.

驗證鏈機制降低大型語言模型中的幻覺

Chain-of-Verification Reduces Hallucination in Large Language Models

摘要

Support