What is Emergent Ability?
Also called emergence.
An emergent ability is a capability that appears in larger models but is absent or near random in smaller ones of the same family, seemingly arriving abruptly rather than improving gradually with scale. Reported examples include multi-step arithmetic and certain instruction-following behaviors. Whether such jumps are genuine phase transitions or artifacts of how performance is measured is actively disputed.
The original observation was that on some tasks, performance stays at chance across several model sizes and then rises sharply past a threshold. This was striking because it implied capabilities could not be forecast from smaller runs, complicating both research planning and safety assessment, since a new ability might appear without warning in a larger model.
A prominent critique argues that many reported emergences are measurement artifacts. Metrics such as exact-match accuracy are discontinuous: a model producing four of five correct digits scores zero until it produces all five. Measuring the same runs with a continuous metric such as token-level probability often reveals smooth, predictable improvement with no jump at all.
The critique does not settle the matter entirely. Some capability changes remain hard to explain as pure metric effects, and the discontinuity of a metric can itself be the practically relevant thing, since a system that gets four of five digits right is useless for the task regardless of how smooth the underlying curve is.
For practitioners the operational lesson is narrower than the debate. Do not assume a smaller model in a family will handle a task just because a larger one does, and do not assume the reverse either. Test the specific model against the specific task with the metric that matters, because family-level claims about capability transfer poorly.
Key points
- Capability appearing sharply above a certain model scale
- Complicates forecasting capability from smaller runs
- Often attributed to discontinuous metrics like exact match
- Continuous metrics frequently show smooth improvement instead
- Test each model on your task rather than assuming transfer
In practice
A team finds that a smaller model in a family scores near zero on a five-step unit conversion task while a larger one scores well, and concludes the ability emerged. Rescoring with partial credit shows the smaller model getting most intermediate steps right and failing on the final rounding. The underlying competence was improving steadily; the exact-match metric simply reported nothing until the whole chain was correct.