研究人员开发了一种名为 Token-Superposition Training (TST) 的新预训练方法，旨在提高大型语言模型训练的效率。TST 包括一个两阶段过程：初始的叠加阶段，其中 token 与多热交叉熵目标相结合进行训练，然后是标准训练的恢复阶段。对高达 100 亿参数模型的评估表明，在同等损失条件下，TST 可将预训练时间缩短高达 2.5 倍。 AI
影响 该方法可以显著降低训练大型语言模型所需的计算成本和时间，从而可能加速研究和开发。
排序理由 该集群包含一篇详细介绍大型语言模型新预训练方法的学术论文。
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2605.06546v1 Announce Type: new Abstract: Pre-training of Large Language Models is often prohibitively expensive and inefficient at scale, requiring complex and invasive modifications in order to achieve high data throughput. In this work, we present Token-Superposition Tra…
Pre-training of Large Language Models is often prohibitively expensive and inefficient at scale, requiring complex and invasive modifications in order to achieve high data throughput. In this work, we present Token-Superposition Training (TST), a simple drop-in method that signif…