Menu
← FIELD NOTESOPINION 2026.11.10 · 13 min

Your model's emergent ability is a property of your ruler, not your model.

The capability that 'switched on' at some scale is often an artifact of a discontinuous metric and a lucky seed — swap the ruler and the cliff smooths into a curve. 'Emergence at N parameters' is not a planning constant you can budget against.

A planning meeting, a capability the team needs that the current model does not have, and a slide with a number on it: at some parameter count the ability switches on. The roadmap is built around that number. Scale to N, the new capability appears, the product ships. Everyone in the room has seen the plots — flat along the bottom for the small models, then a near-vertical climb past some threshold — and they make the number feel like a law of nature, a thing you can put on a Gantt chart and budget compute against.

The number is not a law of nature. In a measurable share of the reported cases, the cliff on that plot is not a fact about the model at all. It is a fact about the ruler used to draw it — the scoring metric — and about how few times the experiment was run. Change the metric from a harsh all-or-nothing one to a continuous one, on the exact same model outputs, and the cliff flattens into a smooth, predictable curve. Re-run training with a different random seed and the same model at the same scale can look emergent or look smooth. The breakthrough you are planning around is, in those cases, a property of your measurement apparatus, not a capability your weights acquired at some magic size — and a property of your apparatus is not something you can put on a roadmap.

This matters because “emergence at N parameters” gets used as a planning constant — a budget line teams commit headcount and GPU-hours against, and the original framing invited exactly that. This post is about why that line is not a constant, what the measurement artifact actually is, and where the honest residue of genuine, unexplained discontinuity still lives.

The original definition built the trap

The word came from a specific paper, and the definition it gave is the thing every later result pushes back on. Emergent Abilities of Large Language Models (arXiv 2206.07682) — Wei et al., which coined the term — defines an ability as emergent if “it is not present in smaller models but is present in larger models,” and frames the phenomenon as fundamentally “unpredictable.” The two load-bearing properties are sharpness and unpredictability. From them the paper draws its planning-relevant conclusion: “emergent abilities cannot be predicted simply by extrapolating the performance of smaller models,” and “the existence of such emergence implies that additional scaling could further expand the range of capabilities of language models.”

Read that last clause the way a roadmap reads it. If new capabilities appear unpredictably at larger scale, and you cannot see them coming from the small models, the rational move is to scale and find out — to treat scale itself as the lever that mints capabilities. That turns a research observation into a budget line: pick the capability you want, estimate the scale at which it “turns on,” commit the compute. The trouble is that both load-bearing properties turn out to be, in a large fraction of cases, artifacts of how the measurement was set up.

The cliff lives in the scoring rule

The mechanism is almost embarrassingly simple. Are Emergent Abilities of Large Language Models a Mirage? (arXiv 2304.15004) — a NeurIPS 2023 outstanding-paper award — makes the case directly: “emergent abilities appear due to the researcher’s choice of metric rather than due to fundamental changes in model behavior with scale. Specifically, nonlinear or discontinuous metrics produce apparent emergent abilities, whereas linear or continuous metrics produce smooth, continuous predictable changes in model performance.”

Picture a task scored by exact-string-match: the model must produce a long answer token-for-token, one point for a flawless match and zero otherwise. As it scales, per-token accuracy rises smoothly — but the whole-string score stays pinned near zero until per-token accuracy is high enough that the model occasionally nails an entire sequence, and then it climbs fast. The smooth underlying improvement is real and continuous; the metric squashes it flat and then releases it, manufacturing a cliff out of a slope. Multiple-choice graded the same harsh way does the same thing. Swap in a metric that gives partial credit — token edit distance, or a proper scoring rule on the probabilities — and the same outputs trace a smooth curve with no cliff.

The Stanford HAI writeup of the work, AI’s Ostensible Emergent Abilities Are a Mirage, puts the practitioner version plainly: the discontinuity lives in the scoring rule, so rewarding partial progress turns the cliff into a slope. The model improves the same way the entire time; the metric decides whether you see that improvement as a smooth ramp or a sudden switch.

This is not a single cherry-picked task. The paper quantifies how concentrated the effect is: in a meta-analysis of BIG-Bench, the possible emergent abilities appear “with at most 5 out of 39” metrics, and ”> 92% of emergent abilities appear under one of two metrics: Multiple Choice Grade and Exact String Match” — both the harsh, all-or-nothing kind that squash a smooth curve flat. And the disappearance is demonstrated, not asserted: “the LaMDA model family’s emergent abilities disappear when measured under a continuous BIG-Bench metric: Brier Score.” Same models, same outputs, swap the ruler, the cliff is gone. The authors’ conclusion is the title of this post: emergent abilities “may be creations of the researcher’s choices, not a fundamental property of the model family on the specific task.”

The seed draws a second cliff the metric cannot explain

If the story stopped at the metric, you could rescue the planning use of emergence: always use a continuous metric, and the curve will be smooth and extrapolable. But a second, independent reason the cliff is not a model property survives that fix.

Random Scaling of Emergent Capabilities (arXiv 2502.17356) proposes that “breakthroughs are instead driven by continuous changes in the probability distribution of training outcomes when performance is bimodally distributed across random seeds.” At a given scale, some training runs land in a basin where the capability works and others do not. As scale increases, the fraction of seeds in the good basin rises continuously — but any single run is one seed, so its scaling curve shows the capability flicking from off to on the moment your seed crosses over. The paper shows it directly: “different random seeds can produce either smooth or emergent scaling trends in synthetic length generalization tasks, multiple choice question answering, and grammatical generalization.” Same architecture, same task, same scale — emergent or smooth, depending only on which seed you drew.

This matters even if you have already taken the metric fix to heart: “our observations hold true even under continuous loss metrics, confirming that random variation must be considered when predicting a model’s performance from its scale.” The seed effect is not the metric artifact wearing a different hat — it is a second mechanism that bites the continuous-metric curve too. The two fixes do not collapse into one: a continuous metric removes the scoring-rule cliff, and you still have a seed-distribution cliff underneath it on a single run.

There is a sharper planning consequence buried here. The scale at which the seed distribution turns bimodal — the real transition — “appears at scales well before most seeds achieve breakthrough.” So the headline “emergence point,” the scale where your single run finally lights up, lags the actual capacity transition. The number on your slide is a population statistic reported as if it were a per-model event: budget against it and you are budgeting against the median of a distribution you sampled once.

The location of the cliff is movable, which means it is not a constant

The cleanest disproof of “emergence at N parameters is a planning constant” is that you can move the N, and a constant you can slide is not one.

Predicting Emergent Capabilities by Finetuning (arXiv 2411.16035) starts from exactly the planning pain the opening describes: “downstream capabilities are far less predictable — sometimes even exhibiting emergent jumps — which makes it challenging to anticipate the capabilities of future models.” Pretraining loss extrapolates beautifully; the capability you care about does not. Then it finds the lever: “finetuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models.” If the location of the cliff depends on what you did to the model after pretraining, the cliff is not an intrinsic property of model size.

The unpredictability — the second pillar of the original definition — is also softer than advertised. The same paper fits a parametric “emergence law” from small models and reports that “using only small-scale LLMs, we find that, in some cases, we can accurately predict whether models trained with up to 4x more compute have emerged,” validated on MMLU, GSM8K, CommonsenseQA, and CoLA. The jump the original framing called unforeseeable is, modelled correctly, foreseeable from a quarter of the compute — a failure of method, not an intrinsic wall.

That this is a live, named debate and not a contrarian hot take is worth establishing. Emergent Abilities in Large Language Models: A Survey (arXiv 2503.05788) frames the open question in the field’s own words — “are they truly emergent, or do they simply depend on external factors, such as training dynamics, the type of problems, or the chosen metric?” — and poses the definitional question the whole argument leans on: “if an ability appears emergent under a binary metric, but follows a continuous trajectory under a more fine-grained metric, should it still be considered emergent?”

This connects to two patterns documented elsewhere. The “lucky seed” problem is the training-time cousin of the evaluation-time one in the eval noise floor: too few samples — few seeds, few benchmark items — produce a number that looks like a stable measurement but is a draw from a wide band. And the deeper move is the one in RL sharpens, it doesn’t expand: a headline-metric jump is read as a new capability, when the metric, the seed, or the re-weighting of existing behavior produced it — and whether any of it is genuinely new is the disputed residue.

Where the cliff is real: the honest residue

The argument so far would be dishonest if it stopped here, because the metric explanation is not the whole story, and the strongest counter-position is real and has to be steelmanned, not waved off.

That counter-position is the loss-threshold account. Understanding Emergent Abilities of Language Models from the Loss Perspective (arXiv 2403.15796), a NeurIPS 2024 paper, re-anchors the x-axis from parameter count to pre-training loss — and on that axis the discontinuity survives the metric fix. Its central finding: “a model exhibits emergent abilities on certain tasks — regardless of the continuity of metrics — when its pre-training loss falls below a specific threshold.” That phrase, regardless of the continuity of metrics, is the rebuttal in four words: the cliff this paper sees is not the scoring-rule artifact, because it persists even with continuous metrics. The evidence is striking — “models with the same pre-training loss, but different model and data sizes, generate the same performance on various downstream tasks” — so performance is pinned by loss, not size. And “before reaching this threshold, its performance remains at the level of random guessing,” then it rises. On the loss axis, by this account, the jump is genuine.

Notice this paper agrees with the post’s method — pick a better ruler than raw parameter count — while disagreeing with its conclusion: it swaps to a better x-axis (loss, not parameters) rather than a better metric, and lands on a real discontinuity. “Switch to a continuous metric and the cliff smooths” and “switch to the loss axis and the cliff stays” are both supported, on different parts of the problem.

The survey is candid about the limits of the mirage story too. After recording that smoother metrics erase the leaps, it documents the exceptions: “in modular arithmetic, periodic elements, French-English translation, and IPA transliteration tasks, partial credit metrics exhibited a sharp performance jump.” Partial-credit — continuous — metrics, and the jump is still there. Its verdict is the one to carry out of this post: “while Schaeffer et al. highlight metric influence, these inconsistencies and alternative findings call for a more nuanced exploration.” That is exactly why the thesis is “a measurable share,” not “all”: a residue of tasks jumps even under continuous scoring, and on the loss axis the discontinuity has independent support. A careful planner treats that residue as unresolved, neither fully debunked nor fully real.

Even the predictability concession is partial. The survey notes that “the GPT-4 technical report suggests that GPT-4’s performance can be anticipated using less than 1/10,000th of its full computational resources” — loss extrapolates that far — but also that “the report acknowledges that certain emergent abilities remain unpredictable.” Even the most resourced labs concede specific capabilities they could not forecast. The unpredictability pillar is cracked — not demolished.

What this changes for how you plan

The practical consequence is not “emergence is fake, ignore scaling.” Scaling improves models; that was never in dispute. It is narrower: stop treating “the capability turns on at N parameters” as a constant you can budget against, because in a large fraction of cases that N is a property of three things you control or sample — the metric, the seed, the post-training — not of model size.

Concretely. When you see an emergence plot, ask what metric drew it; if it is exact-match or multiple-choice-grade, suspect the cliff is the metric squashing a smooth curve, and ask for the same data under a continuous one before planning around the threshold. Treat any single-seed curve as one draw from a distribution: its “emergence point” lags the real transition and shifts run to run, a median at best. And if you must forecast, fit an emergence law from small models — the jump that looks unforeseeable is, where it has been tried, predictable from a quarter of the compute.

The one thing not to do is the thing the original framing quietly licensed: write a parameter count on a slide, call it the moment a capability appears, and commit a roadmap to it as if the number were measured off the model. In the cases that make up the measurable share, you measured your ruler.

Reading list

  • Are Emergent Abilities of Large Language Models a Mirage? (arXiv 2304.15004) — establishes the core mechanism: discontinuous metrics (Multiple Choice Grade, Exact String Match) manufacture the cliff, > 92% of BIG-Bench emergence concentrates on two such metrics, and the cliff disappears under a continuous metric (Brier Score) on the same outputs.
  • Random Scaling of Emergent Capabilities (arXiv 2502.17356) — shows emergence is also a property of the bimodal distribution of outcomes across random seeds, so the same model and scale can look emergent or smooth, an effect that survives even under continuous loss.
  • Emergent Abilities in Large Language Models: A Survey (arXiv 2503.05788) — frames the live debate and supplies the honest counterweight: smoother metrics erase most leaps, but exceptions (modular arithmetic, translation, transliteration) keep a sharp jump even with partial credit.
  • Understanding Emergent Abilities of Language Models from the Loss Perspective (arXiv 2403.15796) — the steelmanned counter-position: re-anchored on pre-training loss, emergence persists regardless of metric continuity, with performance pinned by loss rather than model size.
  • Predicting Emergent Capabilities by Finetuning (arXiv 2411.16035) — shows the emergence point is movable (finetuning shifts it to smaller models) and forecastable, predicting emergence up to 4x more compute ahead on MMLU, GSM8K, CommonsenseQA, and CoLA.
  • Emergent Abilities of Large Language Models (arXiv 2206.07682) — the original Wei et al. definition the title rebuts: emergence as a sharp, unpredictable capability that cannot be extrapolated from smaller models.
  • AI’s Ostensible Emergent Abilities Are a Mirage — Stanford HAI’s plain-language provenance for the NeurIPS 2023 outstanding-paper result: the discontinuity lives in the scoring rule, so rewarding partial progress turns the cliff into a slope.

The plot is real and the climb on it is real, but the steepness of that climb is a choice you made when you picked the metric, and the exact scale where it starts is a sample you drew when you picked the seed. Swap the ruler for a continuous one and the cliff becomes a curve; re-run the seed and it shifts. What is left — the residue that stays sharp under a continuous metric and on the loss axis — is small, real, and genuinely disputed, and it is the only part you would be honest budgeting around. The rest was never your model. It was the shape of your ruler.

NEW ENGAGEMENT · INTAKE

Tell us about it.

The more specific you are, the more useful our first reply.

SERVICE AREA
↩ ENCRYPTED IN TRANSIT
ASK THE FIELD NOTES BETA