Emergent abilities are capabilities that appear in large neural network models but are effectively absent in smaller models trained on the same data with the same objective. They are described as emergent because the capability seems to switch on abruptly rather than improving gradually as scale increases — the model crosses some threshold of parameters or compute and suddenly performs a task it could not perform at all before. Examples documented in the research literature include multi-step arithmetic, chain-of-thought reasoning, analogy completion, and code synthesis.

Think of it like water freezing: as you lower the temperature by small increments nothing dramatic happens, but at exactly zero degrees the liquid abruptly becomes a solid with entirely new structural properties. The phase transition is a collective phenomenon that does not exist at smaller scales.

One frequently cited example is multi-digit multiplication. Smaller LLMs produce random-looking answers to questions like "What is 37 times 84?" but models beyond a certain parameter count begin to produce correct results reliably. The capability appears to emerge between roughly 10B and 100B parameters for this specific task depending on training data. Researchers have proposed that some apparent emergence is an artefact of using discontinuous metrics — switching from "fraction correct" to a metric that rewards partial credit can make the curve look smooth again.

Emergent abilities are closely tied to research on Foundation Models, LLMs, and scaling laws. They relate to Chain of Thought prompting because multi-step reasoning is one of the clearest documented emergence cases, and they connect to Pretraining because the abilities develop from the base training run, not from later Fine-tuning or RLHF.

You will encounter this term in discussions of model scaling, AI safety (unpredictable capabilities can appear without warning), and capability evaluations that attempt to benchmark models at different parameter counts to anticipate what abilities the next generation might exhibit.