Learn · Intermediate

Emergent abilities: do models suddenly gain skills, or are we measuring badly?

An emergent ability is a skill that a small model does not have, a large model does, and which seems to switch on abruptly somewhere in between rather than improving gradually. The idea was formalised by Jason Wei and colleagues in 2022 and became one of the most influential — and most contested — claims in modern machine learning. A 2023 rebuttal from Rylan Schaeffer and colleagues at Stanford argued that much of the apparent suddenness is an artifact of how the abilities were scored. Both papers are largely correct, and the disagreement is more useful than either conclusion.

The original observation

Scaling laws say that a model's loss — its average error at predicting the next token — falls smoothly and predictably as you add parameters, data and compute. No jumps, no surprises. That was among the field's most robust findings.

But Wei and colleagues noticed something that did not fit. On certain tasks, model performance sat at chance level across several orders of magnitude of scale, then rose sharply once models passed some threshold. Three-digit arithmetic. Answering questions in Persian. Identifying an intended meaning from context. Their definition was precise: an ability "not present in smaller models but present in larger models," which "cannot be predicted simply by extrapolating" from the smaller ones.

The implication was unsettling. If overall loss improves smoothly but specific capabilities appear without warning, then nobody can know what a model will be able to do until it has been built and tested. Many of these were catalogued in BIG-bench, a collaborative benchmark assembled to probe exactly this. Chain-of-thought prompting showed a similar pattern — it actively hurt small models and helped large ones.

The rebuttal

Then Schaeffer, Miranda and Koyejo asked a question about the ruler rather than the thing being measured, and won a best-paper award for it.

Consider four-digit addition scored by exact match: right or wrong, no partial credit. Suppose a model's per-digit accuracy improves smoothly from 70% to 80% to 90%. The chance of getting all four digits right goes from about 24% to 41% to 66%. Smooth underlying improvement, but because the metric demands perfection, measured performance hugs zero for a long time and then climbs steeply. The cliff is manufactured by the scoring rule.

Their argument was that this pattern explains a large share of reported emergent abilities. Swap exact-match for a metric with partial credit — token edit distance, say — and the same runs often produce a smooth curve. Emergence, on this account, is a property of the measurement, not the model.

The analogy is a pass mark. Track the proportion of students passing an exam as study hours increase and you see a sharp transition at the threshold. Track their actual scores and you see a gentle slope. Nothing discontinuous happened to anyone's learning; the discontinuity is in where the line was drawn.

Where this actually lands

The mirage paper did not demonstrate that nothing surprising happens at scale, and it is careful not to claim that. It demonstrated that a specific and common measurement choice manufactures sharp curves from smooth ones — which is a serious methodological finding, and it disposes of a good number of the original examples.

But several things survive it. Some capability transitions do not obviously reduce to metric artifacts, and the grokking phenomenon — where a network trains to memorisation, plateaus for a long time, and then abruptly generalises with no change in the training procedure — is a genuine discontinuity in learning dynamics that has nothing to do with scoring rules.

More importantly, the practical problem is untouched. Even if every underlying skill improves smoothly, the thing users and regulators care about is whether the model can do the task, and tasks usually have thresholds. A model that gets 90% of the steps right in a chain of ten still fails the job. Smooth improvement in components produces genuinely sudden changes in usefulness, and "the underlying curve was smooth" is little comfort if you could not predict when it would cross the line that matters.

Why it matters for safety

This is where the debate stops being academic. Frontier labs now publish capability thresholds and responsible scaling policies that commit them to safeguards when a model reaches some level of dangerous capability. Those frameworks assume capabilities can be anticipated.

If Wei is right, they cannot be, and evaluation must be exhaustive and post-hoc — you build the model, then find out. If Schaeffer is right, the underlying skills can in principle be extrapolated with better-designed metrics, and forecasting becomes tractable. The practical upshot from the rebuttal is constructive rather than dismissive: measure with metrics that give partial credit, because they carry information about where a model is on a curve rather than only whether it has crossed a line.

The lesson that generalises beyond this argument is one worth carrying into any benchmark claim. When a graph shows something dramatic, ask what the y-axis is actually counting before concluding something dramatic happened. A great deal of what looks like a discovery about models turns out to be a fact about how they were benchmarked.

Ground Truth. (2026, September 9). Emergent abilities: do models suddenly gain skills, or are we measuring badly?. Ground Truth. https://groundtruth.day/learn/emergent-abilities-and-whether-they-are-real.html

@misc{groundtruth:emergent-abilities-and-whether-they-are-real,
  title  = {Emergent abilities: do models suddenly gain skills, or are we measuring badly?},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/learn/emergent-abilities-and-whether-they-are-real.html}
}

Topics: fundamentals · scaling · evaluation · emergence · interpretability