arXiv：cs.LG· Trevor McCourt, Ila R. Fiete, Isaac L. Chuang·· 4 小时前AI 评分52
arXiv 论文提出容错基础模型：LLM 可在不可靠硬件上训练且错误韧性随规模增强
Fault-tolerant foundation models
AI 导读
arXiv 论文（arXiv:2610.10311）显示，大语言模型可被训练以容忍硬件不可靠性，且错误韧性随模型规模增大而增强而非退化。作者基于 4 万 GPU 小时在模拟故障数字硬件上的训练运行推导出修正的神经缩放律，并推测经适当训练的 LLM 可能是形式上容错的，这意味着在低能耗故障硬件上运行 AI 推理或可带来可观的能耗节省。
正文
Abstract:Emerging computer hardware often trades reliability for energy efficiency; here we show that large-language models (LLMs) can be trained to tolerate this unreliability, and that rather than degrading, their error resilience actually increases as they grow. Modified neural scaling laws inferred from 40,000 GPU-hours of training runs on simulated faulty digital hardware quantify this trend and suggest that models learn to compute within "good" error-correcting codes, whose relative overhead remains finite no matter how large the model gets. This finding leads us to conjecture that appropriately trained LLMs may be formally fault-tolerant; if true, running AI inference on low energy, faulty hardware may be a path to substantial energy savings over the status quo.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Hardware Architecture (cs.AR) |
| Cite as: | arXiv:2610.10311 [cs.LG] |
| (or arXiv:2610.10311v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.10311 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Trevor McCourt [view email]
[v1]
Wed, 7 Oct 2026 16:05:48 UTC (6,366 KB)
来源：arXiv：cs.LG · arxiv.org