temperature sampling is a conventional approach to diversify large language
model predictions. As temperature increases, the prediction becomes diverse but
also vulnerable to hallucinations -- generating tokens t
本文提出了一种新的温度缩放采样方法 Long Horizon Temperature Scaling (LHTS),其优化样本的长线时间似然度,将该方法应用于图像扩散模型和字符 / 语言自回归模型,并证明在可能性和样本质量方面相对于目光短浅的温度缩放有优势,在多项选择类比任务的准确性上也有 10% 的提升。