[论文解读] Chain-of-Verification Reduces Hallucination in Large Language Models
CoVe 让一个基础大语言模型起草回答、规划验证问题、独立回答这些问题,并生成最终经过验证的回答,从而在多任务中减少幻觉(错误信息)。
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.
研究动机与目标
- Motivate and quantify the hallucination problem in large language models across tasks.
- Introduce Chain-of-Verification (CoVe) to mitigate hallucinations by deliberate self-verification.
- Demonstrate that verification-question planning and independent answering improve factual accuracy.
- Compare joint, two-step, and factored verification strategies across tasks.
提出的方法
- Baseline LLM (Llama 65B) generates initial response without verification.
- CoVe generates a verification plan conditioned on baseline response.
- Execute verifications by answering each verification question (joint, 2-step, factored, or factor+revise variants).
- Generate a final verified response that incorporates verification results.
- Investigate variations where verification questions do not attend to the original baseline to avoid repetition.
- Evaluate with tasks: Wikidata list questions, Wiki-Category lists, MultiSpanQA closed-book, and longform biographies; compare to baselines and instruction-tuned models.
实验结果
研究问题
- RQ1Does CoVe reduce hallucinations across different task types?
- RQ2Can CoVe improve factual accuracy without sacrificing correct content?
- RQ3Which verification execution variant yields the best accuracy gains?
- RQ4What is the impact of verification planning quality on overall performance?
主要发现
- CoVe increases precision on Wikidata list-based tasks (0.17 baseline to 0.36 with CoVe).
- Hallucination counts (Negatives) drop from 2.95 to 0.68 with CoVe on Wikidata tasks.
- CoVe improves closed-book MultiSpanQA F1 from 0.39 (few-shot baseline) to 0.46–0.48 depending on variant.
- FactScore for longform biographies increases from 55.9 (few-shot baseline) to 71.4 with factor+revise reasoning.
- Factored and two-step CoVe variants outperform joint CoVe across tasks.
- CoVe can outperform several instruction-tuned or retrieval-augmented models on longform generation (e.g., surpassing InstructGPT, ChatGPT, and PerplexityAI on FactScore in some settings).
阅读本论文
在 Nubint 阅读器中打开原文,边读边就任意段落提问。
在 Nubint 中阅读本论文本解读由 AI 根据论文原文生成,引用前请核对原文。