Review Emergent Capabilities

Emergent capabilities are abilities or measured behaviors that appear unexpectedly, or appear to arise sharply, as a model, training process, or system changes in scale or design.

A capability may not be obvious in smaller systems and then become visible in a larger one. Some apparent jumps can also result from the evaluation method, scoring threshold, prompting, or added system components.

Unexpected capability complicates forecasting and testing. Claims of emergence need careful measurement and separation of model behavior from the surrounding application.

A benchmark score appears near zero for smaller models and then rises sharply, prompting researchers to test whether the pattern reflects true discontinuity or the scoring method.

Emergent does not necessarily mean mysterious, conscious, untrained, or impossible to explain after analysis.

The capability became visible unexpectedly; the measurement still needs examination.

* [[scaling-laws|Scaling Laws]] * [[capability-provenance|Capability Provenance]] * [[foundation-model|Foundation Model]] * [[evidence-selection|Evidence Selection]] * [[calibration|Calibration]]

* [Wei et al., *Emergent Abilities of Large Language Models*](https://arxiv.org/abs/2206.07682) — primary paper describing abilities that appear in larger models but not smaller ones under evaluated settings. * [Schaeffer, Miranda, and Koyejo, *Are Emergent Abilities of Large Language Models a Mirage?*](https://arxiv.org/abs/2304.15004) — counteranalysis showing how metric choice can create apparently sharp emergence.

* Examine the evaluation design with [[evidence-selection|Evidence Selection]] before declaring a discontinuity.