Emergent abilities
Capabilities that appear suddenly above a certain model scale — chain-of-thought reasoning, in-context learning, instruction following — and are absent or near-zero in smaller models.
The term comes from the 2022 Wei et al. paper "Emergent Abilities of Large Language Models." The claim: some tasks remain at near-random performance until model scale crosses a threshold, then suddenly become solvable. Multi-step arithmetic, instruction following, and chain-of-thought all show this pattern.
A 2023 follow-up paper argued some of these emergence curves are an artifact of the metric (e.g., exact-match scoring) rather than a true phase transition. The debate is unresolved but matters less in practice — modern small models hit emergent abilities at much smaller scales via better training.
For agent builders, the practical lesson: capability tiers are real. If your agent needs reliable multi-step planning, you cannot substitute a sub-7B model for a frontier one even with good prompting. The capability simply is not there.
Frequently asked
Are emergent abilities a real phenomenon or a measurement artifact?+
Both arguments have evidence. Some curves smooth out under different metrics; others remain sharp. Treat "emergence" as a useful working concept, not a hard scientific law.
What scale unlocks chain-of-thought reasoning?+
In 2022 models, ~62B parameters. In 2026, well-trained 3B models often do CoT reasonably. Training quality has pulled the threshold down faster than parameter count alone.