理解软错误引发的大语言模型幻觉

AI 摘要

该论文首次系统研究硬件软错误（瞬态比特翻转）如何触发大语言模型幻觉，提出首个软错误诱导幻觉的分类体系，识别出四类主要幻觉模式。作者开发并开源了高效灵活的错误注入框架 LLMEI，可在不同规模 LLM 推理过程中注入软错误。研究在多种模型架构和任务（数学推理、代码生成）上评估了不同错误率下幻觉类型的分布，揭示了硬件可靠性对 LLM 实现的影响。

正文节选

Understanding LLM Hallucinations Caused by Soft Errors Abstract. While hallucinations in Large Language Models (LLMs) have been extensively studied from a software perspective, limited attention has been paid to hardware-level factors. In this paper, we performed a case study to understand LLM hallucination caused by soft errors on hardware. We introduce LLMEI to inject soft errors into LLMs during inference, based on what we evaluate the impact of soft error across diverse LLM architectures and

发布时间：2026-10-19 12:00

抓取时间：2026-10-19 13:48

来源机构：arXiv