Hardware soft errors cause large language model hallucinations
Understanding LLM Hallucinations Caused by Soft Errors
Software EngineeringHardware Architecture
Summary
Large language models (LLMs) sometimes make things up, which is called hallucination. While many studies look at software causes, this paper shows that tiny errors in the computer hardware can also cause these mistakes. The authors created a way to simulate these hardware soft errors during the model’s thinking process and studied how they affect different LLMs and tasks like math and coding. They classified types of hallucinations caused by these errors and measured how often they happen at different error levels. This work helps connect hardware reliability to errors seen in AI model outputs.
What this means in practice
- •For data center engineers: Improve hardware monitoring by incorporating LLM hallucination patterns caused by soft errors to detect and mitigate AI inference faults.
- •For cloud service operators: Develop testing tools that simulate hardware soft errors to assess the reliability of LLM-based services under realistic fault conditions.
Authors
Jinghao Wen, Dongning Ma, Ruixuan Wang, Shaohuang Wang, Xun Jiao
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 tasks, including mathematical reasoning and code generation. We summarize the taxonomy of different failure modes based on LLM hallucination behaviors and analyze their distributions under different error rates. Our find- ings underscore the hallucination-specific diagnostics within the reliability of LLM implementation on hardware.