Scaling up language models has been empirically shown to improve performance and unlock emergent abilities. Conversely, observing worse performance as a function of scale ("inverse scaling") would indicate that scaling encourages behaviors that are misaligned with human preferences. The Inverse Scaling Prize identified eleven such inverse scaling tasks, evaluated on models of up to 280B parameters and up to 500 zettaFLOPs of training compute. This paper takes a closer look at these inverse scaling tasks. We evaluate models of up to 540B parameters, trained on five times more compute than those evaluated in the Inverse Scaling Prize. With this increased range of model sizes and training compute, ten out of the eleven tasks exhibit what we call "U-shaped scaling" -- performance decreases up to a certain model size, and then increases again up to the largest model evaluated. U-shaped scaling can be seen as emergent ability unlocked by scaling and implies that inverse scaling may not hold for larger models.

翻译：缩放语言模型已被经验证明能够提升性能并解锁涌现能力。相反，观测到随模型规模增大而性能下降（"逆缩放"）则表明，缩放会鼓励与人类偏好不一致的行为。逆缩放奖识别出11项此类逆缩放任务，并在参数高达2800亿、训练计算量达500泽塔FLOPs的模型上进行了评估。本文对这些逆缩放任务进行了深入探究。我们评估了参数高达5400亿的模型，其训练计算量是逆缩放奖评估模型的五倍。在模型规模与训练计算量的扩展范围内，11项任务中有10项呈现出我们称之为"U形缩放"的现象——性能先随模型规模增大而下降，随后在最大规模评估模型中再度回升。U形缩放可视为缩放所解锁的涌现能力，并暗示逆缩放可能不适用于更大规模模型。