← Blog/blog/emergent-abilities-metric-mirage

The capability cliff that vanished when the ruler changed

A model improves from getting 80% of tokens right to 90%. That sounds gradual. Now require every token in a 20-token answer to be correct: its expected exact match jumps from about 1% to 12%. The model changed smoothly; the ruler bent the curve.

Schaeffer, Miranda, and Koyejo argue that some reported emergent abilities—capabilities said to appear sharply and unpredictably only at large scale—are this kind of measurement effect. Their claim is about fixed model outputs scored differently, not about denying that bigger models can acquire new abilities.

01

One exponent makes the cliff

Let p be the chance of producing one correct token. Under the paper's simplifying independence assumption, all L tokens are correct with probability pᴸ. Expected token errors instead equalL(1−p). One changes geometrically with answer length; the other changes linearly.

def score_pair(p_token, length):
    exact_match = p_token ** length
    expected_errors = length * (1 - p_token)
    return exact_match, expected_errors

The TypeScript implementation above drives every illustrative chart below; Python and C++ are faithful translations.

1 token5 tokens20 tokens
Exact-match probability for answers of 1, 5, and 20 independent tokens. The horizontal axis is smooth per-token skill from 45% to 95%; these are analytical illustrations, not measured model outputs.
02

A smooth scaling law can look like a phase change

The paper's toy model assumes per-token cross-entropy falls as a power law with parameter count. Cross-entropy is the penalty for assigning too little probability to the correct token. Turning that smooth loss into per-token correctness remains smooth; raising it to the twentieth power creates the apparent takeoff.

per-token correctness20-token exact match
The paper's illustrative power-law mechanism, rebuilt with relative model scale and a 20-token target. Both series come from the same smooth cross-entropy curve; values are illustrative, not paper-reported benchmark scores.
03

The benchmark can be too small to see progress

A test set of N answers measures accuracy only in steps of1/N. If a model's true exact-match probability is below that step, observing zero successes is ordinary sampling noise, not evidence of zero ability. Long answers make this worse because pᴸ becomes tiny.

Base-10 log of the number of 20-token test examples needed for a 95% chance of seeing at least one exact match. For 60% per-token correctness, the exact event is so rare that roughly 82,000 examples are needed.
04

What the paper actually found

AnalysisPaper-reported findingWhy it matters
Two dominant metrics>92% of annotated BIG-Bench emergent abilitiesExact string match or multiple-choice grade
Preferred BIG-Bench metrics39 examined; at most 5 showed emergenceThe effect clustered by metric, not only task
GPT arithmetic re-evaluationAll four model sizes above chance with more dataThe apparent zero plateau was resolution-limited
Reported observations from Schaeffer et al.; the analytical charts above are separate illustrations.

Multiple-choice grading creates a second discontinuity: whichever option has the largest probability gets 1, all others get 0. A continuous Brier score instead measures the squared error of the full probability forecast. The paper found that changing to continuous or approximately linear metrics often removed the apparent emergence while retaining graded improvement.

05

The caveat cuts both ways

So exact match is not a bad metric when exactness is the real requirement. It is a bad microscope for forecasting smooth scaling unless it is paired with token-level scores, confidence-sensitive metrics, enough test examples, and dense coverage of model sizes.

06

What I would probe next

  • Report exact match and token error on the same fixed outputs.
  • Attach confidence intervals that expose the zero-success uncertainty.
  • Pre-register metrics before searching many tasks for dramatic cliffs.

The measurement story sits above the model itself. To inspect the attention computation whose outputs these benchmarks score, poke at the interactive transformer page.

References

  1. Rylan Schaeffer, Brando Miranda, Sanmi Koyejo (2023). Are Emergent Abilities of Large Language Models a Mirage?. NeurIPS 2023
  2. Jason Wei et al. (2022). Emergent Abilities of Large Language Models. Transactions on Machine Learning Research