Special Topic · Open Weights, Distillation & Local Runs

Emergence: Why Capabilities Appear Suddenly

Capabilities jump in steps once a model crosses a scale threshold — plus the academic dispute this phenomenon is still under

THE QUESTION THIS PAGE ANSWERS

ANSWER FIRST

What is the key idea behind “Emergence: Why Capabilities Appear Suddenly”?

Capabilities jump in steps once a model crosses a scale threshold — plus the academic dispute this phenomenon is still under

DECISION RULE

Make the claim earn its place. Use this page as a decision aid, not a definition to memorize. Connect the idea to one real task, one observable result, and one failure that would change your mind.

TRY NEXT

Write one question you could answer with evidence after trying this idea.

WATCH FOR

A conclusion that sounds complete but leaves the key assumption untested.

The short answer
Capability often does not grow smoothly with parameter count. For one class of tasks, a model's score hugs zero across a wide range of scales, showing no apparent progress at all. Once the parameter count passes a certain point, the score jumps to a usable level within a short stretch. Before that jump, the data points you already have tell you nothing about what comes next.
Two curves with completely different shapes

Put parameter count on the horizontal axis and the score on some task on the vertical axis, and you see two kinds of trajectory. One climbs steadily with scale, gaining a little with every doubling of parameters. The other hugs the floor for a long stretch and then suddenly lifts off at some point.

Threshold range (illustrative) 100% 75% 50% 25% 0 1M 10M 100M 1B 10B 100B 1T Parameter count (log scale) Task score
Ordinary ability, rising steadily with scale Emergent ability, jumping once the threshold is crossed
The chart above is illustrative; the values on both axes do not correspond to any specific model or benchmark. The phenomenon and the shape of the curves follow Wei et al., Emergent Abilities of Large Language Models, TMLR 2022 (arXiv:2206.07682).

The trouble lies in the left half. Before the jump, this curve looks exactly like "this approach is a dead end." With only the early data in hand, there is no way to extrapolate whether it will lift off later, and no way to predict where. Nor is there currently a theory that can tell you in advance at what scale a given ability will appear.

That reads flatly enough. Dragging it once is what makes it land. Below, the model scale is yours to set: push it up from 0.5B and watch six abilities switch on one at a time.

Drag from the smallest to the largest and notice that the lamps do not come on together — they pop up one by one.
Model scale 0.5B
0 / 6
The thresholds above are order-of-magnitude illustrations drawn from observations repeatedly cited in published research and industry discussion; they are not precise measurements. Change the architecture, the training data, or the evaluation method and every one of these positions shifts noticeably.
Six examples that come up again and again

All of the abilities below get cited as examples of emergence in published research and industry discussion. The scales in parentheses are order-of-magnitude references only. Change the architecture, the training data, or the evaluation method, and the same ability shifts noticeably.

Multi-step mathematical reasoning
Small models get a share of single-step arithmetic right. Ask them to chain several steps of reasoning before answering and their accuracy stays pinned near zero for a long time.
around 10B
Cross-lingual transfer
Reasoning patterns learned from English text carry over to languages that barely appeared in training. Below the required scale, switching languages makes the score collapse.
around 7B
Structured output and tool calling
Reliably emitting valid JSON and filling in the right argument types for a function signature. Nobody designed this ability in; it was found after training finished.
around 7B
Chain-of-thought reasoning (CoT)
Having the model write out its reasoning before answering. On small models this buys nothing, and it sometimes derails questions they would otherwise have answered correctly.
around 100B
Theory of mind (ToM)
Working out what the other person in a conversation knows and how they will think about it. This one is the most disputed: some studies found the ability vanished after rewording the test items by a few words.
Disputed
Persona consistency over long conversations
Holding on to an assigned identity, tone, and boundaries dozens of turns in, without drifting and without being talked out of it in a few messages.
around 13B
Chain-of-thought produces clear gains only at around the 100B scale; see Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, NeurIPS 2022 (arXiv:2201.11903). On theory of mind, see Kosinski, arXiv:2302.02083 (2023); around the same time Ullman, arXiv:2302.08399 showed that the behavior disappears after small alterations to the task. The parameter counts for the remaining items are order-of-magnitude references — the industry has no agreed authoritative threshold, so please do not cite them as precise figures. Verified 2026-08-07.
Academically, this is not settled
One line of research argues that a fair share of "emergence" is an artifact created by the evaluation metric. Change the scoring method and the same models on the same questions produce a smoothly rising curve.

The argument runs like this. Multi-step math problems are usually scored on whether the final answer is exactly right; one wrong step anywhere and the whole item scores zero. That scoring is discrete, and it compresses continuous progress into a binary outcome. A model that has improved from wildly wrong to almost right still scores zero. Only when it clears that last bit of the threshold does the score jump off the floor all at once.

Switch to a continuous metric — for instance, the log probability of the correct answer token by token — and the curve from the same set of experiments becomes a smooth climb. The jump disappears, which says the jump came from the ruler, and that the model itself was improving steadily all along.

Discrete metric

Only a clean pass counts. All the progress along the way is recorded as zero and cashed in at the end, which looks exactly like a sudden jump.

Continuous metric

Partial credit counts too. Progress across the same set of models is recorded in full, and the curve turns into a smooth climb.

Schaeffer, Miranda & Koyejo, Are Emergent Abilities of Large Language Models a Mirage?, NeurIPS 2023 (arXiv:2304.15004). The paper argues that emergence is largely induced by the researcher's choice of metric: nonlinear or discrete metrics produce jumps, while linear, continuous metrics smooth the curve out. The two mini charts above illustrate the principle and are not taken from the paper's data. Verified 2026-08-07.

This rebuttal does not overturn the phenomenon entirely. What it establishes is something more specific: when you see a step-shaped curve, first confirm whether the step comes from the model or from your ruler. For anyone building a product, that reminder is more useful than the debate itself. If your acceptance criterion is "it only passes if it runs end to end," you will most likely see a step in your own data too.

What it means if you build products
1

Run your own test before moving to a bigger model

Benchmarks measure general tasks, and your scenario may sit right near some threshold. Moving up one size class might solve it instantly, or change nothing at all. Spend half a day building a test set of twenty or thirty samples and run it on real data; that beats reading ten benchmark writeups.

2

Don't force a prompt to cover what a small model cannot do

When a task stays wrong no matter how you rewrite the prompt, try one size class up first. If the bigger model gets it right immediately, those elaborate prompts were only papering over a gap in scale. Prompts like that are usually long and brittle, and they break the moment you change model or version.

3

When something "suddenly works," don't conclude too fast either

The dispute above applies to your own tests as well. Cross-check with a more lenient scoring scheme — credit per step, or credit per field. The conclusion only holds if both rulers show the jump.

One line to remember: scale is an independent variable and deserves to be tested on its own. It may hand you capabilities you never expected, or it may give you nothing. Only your own task can answer that.
What comes next

Emergence shows that scale buys capability. It also leaves a question open: if bigger models are stronger, why is the whole industry working so hard to make models smaller? Squeezing hundreds of billions of parameters down to a few billion costs exactly the kind of capability this section describes. What compression buys, and whether the price is acceptable, is where the next section starts.

Put “The short answer” back into its constraints

“Put parameter count on the horizontal axis and the score on some task on the vertical axis, and you see two kinds of trajectory.” shows that a model, license, access route, or leaderboard is information—not an answer outside context. The real choice depends on task, data boundary, latency, quality floor, and operating cost.

Write elimination criteria before chasing the top score

The comparison in “The trouble lies in the left half.” should use the same real inputs while observing correctness, failure behavior, response time, and cost. A model leading a public leaderboard may still fail your license, privacy, or peak-latency constraints.

Without a test set, there is no reliable winner

Start with “Emergence shows that scale buys capability.”: choose inputs that could genuinely change the decision and write down one counterexample that would reverse your choice. That is more useful than memorizing a single ranking.

From “The short answer” to “Two curves with completely different shapes”

“The short answer” grounds the problem in “Capability often does not grow smoothly with parameter count. For one class of tasks, a model's score hugs zero across a wide range of scales, showing no apparent progress at all. Once the parameter count passe…”. “Two curves with completely different shapes” then moves it toward “Put parameter count on the horizontal axis and the score on some task on the vertical axis, and you see two kinds of trajectory. One climbs steadily with scale, gaining a little with every doubling of parameter…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.

Carry the judgment into the next situation

For model selection, write non-negotiable constraints from the real task first. Compare quality, failure behavior, latency, licensing, and cost on the same inputs; use a leaderboard only as a starting point.

  • “The short answer”: Capability often does not grow smoothly with parameter count. For one class of tasks, a model's score hugs zero across a wide range of scales, showing no apparent progress at all. Once the parameter count passe…
  • “Two curves with completely different shapes”: Put parameter count on the horizontal axis and the score on some task on the vertical axis, and you see two kinds of trajectory. One climbs steadily with scale, gaining a little with every doubling of parameter…
  • “The closing point”: Partial credit counts too. Progress across the same set of models is recorded in full, and the curve turns into a smooth climb

The final “The closing point” brings the discussion to “Partial credit counts too. Progress across the same set of models is recorded in full, and the curve turns into a smooth climb”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.

Mark as learned Your reading progress updates automatically
← PreviousNext →

Keep reading

The next useful article in the thread.

ARTICLE DISCUSSION

Leave one useful thought here.

Keep the idea that clicked, the question that stayed open, or a small note for the next learner.

Discussing Emergence: Why Capabilities Appear Suddenly Open Weights, Distillation & Local Runs
3discussionsArticle discussion · synced with the Circle
View in the learning circle
AM
Asha MorganContent editor
INSIGHTField note

I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.

ARTICLE DISCUSSION7 helpful
LH
Lin HarperIndie developer
INSIGHTInsight

After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.

ARTICLE DISCUSSION5 helpful
KM
Kiki MooreProduct operations
QUESTIONQuestion

When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.

ARTICLE DISCUSSION4 helpful