Foundations
Emergent abilities
Capabilities absent in smaller models that appear abruptly at scale; a contested claim — critics attribute the sharp jumps to the choice of nonlinear metrics.
Emergent abilities are capabilities that, per the definition of Wei et al. (2022), are not present in smaller models but appear in larger ones — and cannot be predicted by extrapolating the performance curve of smaller models. Frequently cited examples include in-context learning, instruction following, and step-by-step reasoning (see reasoning model).
As a large language model grows in parameter count and training data, accuracy on some tasks stays near chance for a long time, then jumps sharply past a certain threshold. This abrupt “phase transition” led researchers to speak of qualitatively new, unplanned capabilities unlocked by scale.
The claim is contested. Schaeffer et al. (2023, a NeurIPS Outstanding Paper) showed that the sharpness often stems from the choice of nonlinear metrics (such as exact-match accuracy): continuous metrics yield smooth, predictable improvement, so the “emergence” disappears. The debate remains active in 2025-2026 because it directly affects how we forecast risks of frontier models and how we interpret benchmark results.