涌现能力Emergent Abilities
Emergent Abilities
小模型没有、大模型突然会了的能力——以及它可能是评测假象的争论。Abilities small models lack that large models suddenly have — and the dispute that they may be a metric artifact.
Wei 等人把涌现定义为：小模型不具备、大模型具备的能力。这种能力无法由小模型的缩放曲线外推预测。
思维链提示就是文中举的涌现例子。
Schaeffer 等人提出相反解释：许多「涌现」源于评测指标的非连续选择。
改用连续指标后，性能随规模平滑变化；涌现未必是缩放的根本属性。
Wei et al. define an emergent ability as one absent in smaller models but present in larger ones, unpredictable by extrapolating small-model scaling curves.
Chain-of-thought prompting is their flagship example.
Schaeffer et al. offered the opposite account: many apparent emergences come from discontinuous metric choices.
With continuous metrics, performance changes smoothly with scale; emergence may not be a fundamental property of scaling.
「大到某个规模突然会变聪明」是营销常用话术。正反两篇都读过，你才有判断依据。'Suddenly gets smart past some size' is a marketing staple. Read both sides before you judge the claim.