The Illusion of Emergent Abilities in Large Models

34 concepts20 papers1 proposal

When scaling up large language models, researchers frequently observe a striking phenomenon: a model completely fails a task across several orders of scale, scoring a flat 0% on an evaluation metric, only to suddenly shoot up to high accuracy once it crosses a specific size or compute threshold. This behavior, popularized as an "emergent ability" by Jason Wei et al. in Emergent Abilities of Large Language Models, has been interpreted as a qualitative phase transition—a moment where a model spontaneously unlocks complex reasoning capabilities.

However, a growing body of research suggests that these dramatic leaps are often an illusion. Instead of a fundamental shift in the model's internal logic, "emergence" is frequently an artifact of how we measure success. When evaluated through the lens of continuous, high-resolution metrics, the jagged cliffs of emergence dissolve into smooth, predictable curves of incremental improvement.

The Mechanism of Thresholding and Discretization

To understand why continuous learning looks like a sudden breakthrough, consider how we define metrics like "Accuracy" or "Exact Match." These are discontinuous step functions. They binarize a model's output into a hard 0 or 1.

During training, a model optimizes a smooth, continuous loss function (such as cross-entropy), gradually shifting its internal probability distributions to favor correct tokens over incorrect ones. If a task requires the model's prediction error to fall below a strict threshold to be counted as "correct," the accuracy metric will remain at zero even as the underlying loss steadily decreases.

Underlying Loss (Smooth)       Thresholded Accuracy (Discontinuous)
          \                                     |
           \                                    |         /--- 100%
            \____                               |________/
                                                ^
                                          Threshold Crossed

This mathematical artifact is not unique to deep neural networks. Noam Itzhak Levi et al. demonstrated this analytically in Grokking in Linear Estimators -- A Solvable Model that Groks without Understanding. By analyzing simple linear estimators, the authors proved that a gradual, exponential reduction in generalization loss is transformed into a sudden, discrete jump in performance when evaluated via a hard error threshold. The model appears to suddenly "grok" the task, but the transition is purely an observational byproduct of the metric's discretization.

When applied to frontier AI models, this thresholding effect destroys our ability to predict downstream performance. As Rylan Schaeffer demonstrates in Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?, all-or-nothing metrics mask the incremental progress occurring within a model's internal representations, making performance jumps appear spontaneous rather than evolutionary.

The Multiplicative Effect in Compositional Tasks

The illusion of emergence compounds heavily when a task requires multi-step reasoning or compositionality.

Imagine a task that requires a model to complete $N$ independent steps correctly to receive a point. If the model has a probability $p$ of completing any single step correctly, and the evaluation metric only rewards total success (an all-or-nothing exact match), the final score scales as $p^N$.

If the model's per-step capability $p$ improves linearly with scale, the final metric $p^N$ will scale non-linearly. For a 5-step task, if $p$ improves from 20% to 80%, the final score undergoes a dramatic explosion from 0.03% to 32.7%.

Rylan Schaeffer et al. systematically exposed this dynamic in Are Emergent Abilities of Large Language Models a Mirage?. By changing the evaluation of prominent benchmarks from non-linear, all-or-nothing metrics to smooth, continuous ones that award partial credit, the authors showed that the sharp transitions disappear, resolving into predictable power-law scaling curves.

This multiplicative phenomenon is also highly visible outside of language modeling. In Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task, Maya Okawa et al. showed that diffusion models learning to compose multiple visual attributes exhibit a binary-like emergence. The model appears to suddenly master the composition, but only because it has finally mastered the last of the constituent "atomic" skills required to satisfy the multiplicative evaluation criteria.

Format Alignment and Evaluation Insensitivity

A third factor driving spurious emergence is the alignment between a model's scale and its ability to follow strict output formats. A smaller model might possess the underlying knowledge to solve a problem, but fail to receive credit because it outputs the answer in a non-standard format that trips up strict evaluation regexes.

As Sheng Lu et al. point out in Are Emergent Abilities in Large Language Models just In-Context Learning?, traditional exact-match scoring creates artificial performance gaps by heavily penalizing semantically correct answers that violate formatting constraints. When prompts are optimized or metrics are refined to accommodate smaller models, the apparent sudden jumps flatten into predictable linear improvements.

This format-alignment bottleneck means that "emergence" often just marks the point where a model's scale finally allows it to match the test task's expected schema. In Training on the Test Task Confounds Evaluation and Emergence, Ricardo Dominguez-Olmedo et al. demonstrated that if smaller models are explicitly fine-tuned or prepared for the target task's format, their performance follows a predictable log-linear scaling law at much smaller parameter counts. The sudden jump vanishes because the formatting barrier has been removed.

Restoring Resolution to Evaluation

To move past the metric mirage, researchers are developing high-resolution evaluation frameworks that expose underlying progress.

One powerful approach is massive sampling. In Predicting Emergent Abilities with Infinite Resolution Evaluation, Shengding Hu et al. introduced a metric called PASSUNTIL. By sampling thousands of outputs to capture extremely low-probability successes in smaller models, they uncovered a smooth, continuous task scaling law that allowed them to accurately predict a model's downstream performance long before scaling it up.

Low-Resolution Metric (e.g., Exact Match) -> Captures only high-probability success -> Jagged Jump
High-Resolution Metric (e.g., PASSUNTIL)   -> Captures low-probability success   -> Smooth Curve

This shift toward continuous tracking is standardizing across other AI domains:

  • Dense Retrieval: Yan Fang et al. utilized continuous contrastive entropy in Scaling Laws For Dense Retrieval to map smooth retrieval improvements, avoiding the abrupt steps seen in discrete ranking metrics.
  • Computer Vision: To combat the "cliff-edge effect" where minor pixel shifts cause massive score drops in instance segmentation, Long Chen et al. proposed sortedAP in SortedAP: Rethinking evaluation metrics for instance segmentation, providing a continuous score that degrades gracefully with quality.

Is Emergence Entirely a Mirage?

While the metric mirage hypothesis explains many apparent breakthroughs, the community continues to investigate where genuine functional shifts might occur. Zhengxiao Du et al. note in Understanding Emergent Abilities of Language Models from the Loss Perspective that pre-training loss remains a universal predictor of performance, but identifying whether specific transitions represent true internal logic shifts or measurement artifacts remains a core challenge.

Offering a nuanced view, Rosie Zhao et al. argue in Distributional Scaling Laws for Emergent Capabilities that while binary metrics exaggerate breakthroughs, scaling does fundamentally alter the probability of a model run successfully clustering into a high-performance mode across different random seeds.

Ultimately, as emphasized by Anna Rogers et al. in Position: Key Claims in LLM Research Have a Long Tail of Footnotes, the field requires strict empirical rigor. Many extraordinary claims of spontaneous capability acquisition can be traced back to non-smooth metrics or data contamination. By designing metrics that match the continuous nature of gradient descent, we gain a clearer, more predictable view of how AI capabilities truly evolve.

Key Foundational Work

Emergent Abilities of Large Language Models (2022) — contributed 4 concept elements to this direction.
How we BLESSed distributional semantic evaluation (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_metric_induced_emergence_80a95120,
  title = {metric-induced emergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metric-induced-emergence/rdr_80a95120909143a896db58a8092461b0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_metric_induced_emergence_80a95120_v1,
  title = {metric-induced emergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metric-induced-emergence/rdr_80a95120909143a896db58a8092461b0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake