Are Emergent Abilities of Large Language Models a Mirage?

Abilities not present in smaller models but present in larger ones, so they can’t be predicted by extrapolating smaller models’ performance. Their sharpness (a seemingly instant jump from absent to present) and unpredictability (appearing at unforeseeable scales).

If abilities appear sharply and unpredictably, larger models might suddenly, without warning, acquire dangerous capabilities.

They are mostly a mirage created by the researcher’s analysis: chiefly a metric that nonlinearly or discontinuously transforms a smoothly improving per-token error rate, and secondarily too little test data to measure small models’ performance precisely, so they look wholly incapable.

Multiple Choice Grade (1 if the most probability mass is on the correct option, a discontinuous step) and Exact String Match (1 only if the whole output matches, nonlinear in the per-token error); together they cover over 92% of hand-annotated emergent abilities.

If per-token cross-entropy falls as a power law in model size \(N\), the probability of a correct token \(p_N\) rises smoothly. Requiring all \(L\) tokens correct gives accuracy \(\approx p_N^{\,L}\), which stays near zero and then rises steeply on a log-scale plot, more so for longer targets.

Token edit distance is about \(L(1 - p_N)\), linear in the per-token error, so the same models show smooth, continuous, predictable improvement with scale: the “emergence” disappears.

(1) A metric that nonlinearly or discontinuously scales the per-token error rate; (2) insufficient resolution to estimate small models’ performance (resolution is set by 1/test-set size); (3) too few model sizes sampled in the large-parameter regime.

On 2-digit multiplication and 4-digit addition, accuracy looked emergent for 4–5-digit targets, but scoring the same outputs with token edit distance showed smooth, predictable improvement with scale, degrading quasi-linearly as target length grew.

With more test data (higher resolution), every InstructGPT/GPT-3 model had above-chance accuracy, improving smoothly with scale and falling roughly geometrically with target length, as the toy model predicts; zero accuracy had been an artifact of small test sets.

Emergence is tied to metrics, not to task–model pairs: of 39 BIG-Bench metrics at most 5 showed it; human-annotated emergent abilities appeared under only 4 metrics. Since BIG-Bench often scores a task with several metrics, the absence under the others is telling.

Tasks where LaMDA looked emergent under the discontinuous Multiple Choice Grade showed smooth improvement under the continuous Brier score (squared error of the probability given to the right answer).

By choosing sharp metrics: shallow autoencoders on CIFAR-100 improved reconstruction error smoothly, but a metric counting images reconstructed below an error cutoff jumped abruptly; Transformers classifying Omniglot characters looked emergent when scored as getting all images in a sequence right.

No: they argue only that some previously claimed emergent abilities appear to be mirages created by researcher choices. Some private models couldn’t be analyzed, and whether human judgments of quality are themselves thresholded remains open.

Broken power laws treat emergence as real, a break in the scaling curve; Schaeffer et al. show it can arise under a single smooth power law. The quanta hypothesis says tasks need sets of discrete skills that larger models are likelier to have all of; both could be true for some abilities.

Task and metric are separate, meaningful choices; continuous or linear metrics are better for forecasting scaling, while nonlinear ones need much more test data; claims need controls for multiple comparisons (BIG-Bench has about \(10^6\) task–metric–model triples); and models and outputs should be open to independent study.