arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2403.03558cs.CL
收藏

基于不可解数学应用题的大语言模型幻觉基准测试

Benchmarking Hallucination in Large Language Models based on Unanswerable Math Word Problem

East China Normal University(华东师范大学);Fudan University(复旦大学);Shanghai AI Laboratory(上海人工智能实验室)

Yuhong Sun, Zhangyue Yin, Qipeng Guo, Jiawen Wu, Xipeng Qiu, Hui Zhao

AI总结:

本研究构建含5200道题的UMWP不可解数学应用题数据集,提出结合文本相似度与数学表达式检测的评估方法,测试31个大语言模型,证实上下文学习与RLHF训练可显著提升模型避免幻觉的能力。

摘要:

大语言模型(LLM)在各类自然语言处理(NLP)任务中表现出色,但在模糊语境下容易生成不可靠的推测内容,即幻觉。本文提出一种基于不可解数学应用题(MWP)评估问答(QA)场景下LLM幻觉的新方法。为支撑该方法,我们创新性构建了名为Unanswerable Math Word Problem(UMWP,不可解数学应用题)的数据集,包含5大类共5200道题目。我们开发了结合文本相似度与数学表达式检测的评估方法,以判定LLM是否认为题目不可解。对GPT-3、InstructGPT、LLaMA、Claude等31个LLM开展的大量实验结果表明,上下文学习与基于人类反馈的强化学习(RLHF)训练能显著提升模型避免幻觉的能力。研究证明,利用数学应用题是评估幻觉的可靠有效方法。我们的代码与数据可在https://github.com/Yuki-Asuuna/UMWP获取。

英文摘要:

Large language models (LLMs) are highly effective in various natural language processing (NLP) tasks. However, they are susceptible to producing unreliable conjectures in ambiguous contexts called hallucination. This paper presents a new method for evaluating LLM hallucination in Question Answering (QA) based on the unanswerable math word problem (MWP). To support this approach, we innovatively develop a dataset called Unanswerable Math Word Problem (UMWP) which comprises 5200 questions across five categories. We developed an evaluation methodology combining text similarity and mathematical expression detection to determine whether LLM considers the question unanswerable. The results of extensive experiments conducted on 31 LLMs, including GPT-3, InstructGPT, LLaMA, and Claude, demonstrate that in-context learning and reinforcement learning with human feedback (RLHF) training significantly enhance the model's ability to avoid hallucination. We show that utilizing MWP is a reliable and effective approach to assess hallucination. Our code and data are available at https://github.com/Yuki-Asuuna/UMWP.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑