In the LLM literature, "emergent" usually means what Wei, Tay, et al. (2022) defined: a capability absent in smaller models and present in larger ones, which cannot be predicted by extrapolating the smaller models' curve (see emergent abilities). Krakauer et al. (2026) note that this is a thin notion compared with that of complexity science, where emergence is neither surprise nor discontinuity. Minimally, it is the reorganization of a system that supports a new, often smaller description, one that screens off microscopic details irrelevant to predicting its behavior (2026, p. 2): fluid dynamics with respect to molecules, or Anderson's (1972) "more is different."
Hence, to speak of emergence in an LLM, performance is not enough; one must show the new coarse-grained representations that sustain it inside the network (2026, p. 2). They also distinguish two kinds. In physics and chemistry, emergence is knowledge-out: complex structure from simple elements. In systems that learn—organisms, economies, LLMs—it is knowledge-in: the complexity comes from a complex environment (2026, p. 5), and "the more information about the world that resides in language, the weaker emergence claims become" (2026, p. 10). And they separate emergent capability from emergent intelligence: a calculator has capabilities, but it does not use its abstractions to draw analogies or to solve new problems economically ("less is more") (2026, p. 2).
The technical discussion runs in parallel. Schaeffer et al. (2023) argued that many jumps are a mirage produced by discontinuous metrics, such as exact match, which smooth out under continuous metrics. Baherwani et al. (2026) find genuine jumps during training, tied to the abrupt learning of sparse attention patterns, which appear at different points depending on the initialization.
The word also carries feelings. For Santelli (2026), the fact that models acquire skills their designers never sought—something like learning some physics when taught Latin—fuels both enthusiasm and fear. It helps to separate epistemic emergence (we did not predict it) from ontological emergence (it does not reduce to the parts): the first is a fact about us; the second, a strong thesis that every successful mechanistic explanation puts under pressure.