Emergent Abilities in LLMs

9 min read

Learn

Quick Reading

Estimated 9 mins

Prereq

Intermediate

Basic ML concepts helpful

Interactive

Static Playbook

Static guide & reference tables

Capabilities that seem to appear out of nowhere

An emergent ability, as the term is used in the LLM scaling literature (notably Wei et al., 2022), refers to a capability that is essentially absent — performing at or near random-chance level — in smaller models, but appears sharply and unpredictably once model scale crosses some threshold, rather than improving gradually and predictably the way scaling laws describe loss. Examples cited in the original survey include multi-step arithmetic, certain multi-task language understanding benchmarks, and specific reasoning-chain tasks, where performance on a benchmark stays flat near zero across several smaller model sizes and then jumps substantially at a larger size, with no model of intermediate scale showing a gradual ramp.

emergent_vs_smooth_metric.txt

text

Smooth metric (e.g. next-token prediction loss):
  Loss  |\
        | \___
        |     \______
        |            \__________
        +------------------------> model scale  (predictable, smooth improvement)

"Emergent" metric (e.g. exact-match accuracy on a multi-step task):
  Acc   |                               ______
        |                              /
        |_____________________________/
        |  (near-zero for many scales, then a sharp jump)
        +------------------------> model scale

The counterargument: emergence as a measurement artifact

A widely discussed 2023 paper by Schaeffer, Miller, and Koyejo ('Are Emergent Abilities of Large Language Models a Mirage?') argued that much of this apparent discontinuity is a byproduct of the chosen evaluation metric, not an intrinsic property of the underlying model capability. Many emergent-ability benchmarks use discontinuous, non-linear metrics like exact-match accuracy on multi-step tasks — if a task requires getting every one of $k$ sequential steps correct, and each step's per-step accuracy improves smoothly and gradually with scale, the overall exact-match probability (the product of per-step correctness probabilities) can still appear to jump sharply once per-step accuracy crosses a threshold, purely as a mathematical consequence of multiplying many smoothly-improving probabilities together — not because anything qualitatively new happened inside the model at that scale.

metric_choice_matters.py

python

import numpy as np

def per_step_accuracy(scale):
    # Smooth, gradual improvement with scale (illustrative sigmoid)
    return 1 / (1 + np.exp(-(scale - 5)))

scales = np.linspace(0, 10, 11)
steps_required = 5  # a multi-step task requiring all steps correct

per_step = per_step_accuracy(scales)
exact_match = per_step ** steps_required   # probability all 5 steps are correct

for s, p, e in zip(scales, per_step, exact_match):
    print(f"scale={s:.0f}  per-step acc={p:.3f}  exact-match acc={e:.3f}")
# per-step accuracy improves gradually; exact-match looks like it "emerges" sharply

Two readings of the same observations

View	Claim	Supporting evidence	Key limitation
Genuine emergence	Qualitatively new capabilities appear at scale thresholds, not predictable from smaller models	Several benchmarks show sharp, hard-to-predict jumps at specific scales	Hard to rule out metric artifacts as the real cause
Metric artifact / mirage	Underlying capability improves smoothly; metric choice creates the appearance of a jump	Switching to smoother, partial-credit metrics often reveals gradual improvement on the same tasks	Doesn't necessarily explain every reported emergent behavior, and some tasks may still show genuine discontinuities

This debate has real planning consequences

Whether emergent abilities are genuine or a metric artifact isn't just academic — if specific valuable capabilities genuinely can't be predicted from smaller-scale experiments, that materially limits how confidently a lab can plan a large, expensive training run around achieving a target capability, compared to the Chinchilla-style predictability of loss itself. Most careful practitioners now evaluate candidate benchmarks with both discontinuous and smooth/partial-credit metrics before drawing conclusions about whether a capability is emerging gradually or not.

Both views can be simultaneously true

It's possible — and increasingly the consensus reading — that some reported 'emergent' jumps really are metric artifacts of exactly the kind described above, while others reflect genuine, harder-to-predict phase-transition-like changes in what a model's internal representations support. Resolving which case applies to any specific capability generally requires digging into per-step or partial-credit performance rather than relying on the headline discontinuous metric alone.

What's next?

One specific capability frequently cited in the emergence debate — the ability to perform a new task from a handful of examples given purely in a prompt, with no weight updates at all — deserves its own closer look, since the mechanics of how in-context learning actually works inside a transformer are still an active area of interpretability research.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →