← Back to blog

Why many emergent abilities are measurement artifacts

July 12, 2026 · 20 min read

Abstract

One of the sharpest disputes in LLM scaling research is whether "emergent abilities" are genuinely abrupt capability births or mostly a measurement illusion. The narrow question in this memo is not whether larger models become better. They do. The question is why some benchmark curves look flat for a long time and then appear to jump from failure to competence at a particular scale. My read of the literature is that many famous emergence plots are indeed artifacts of coarse metrics such as exact match, accuracy, or pass/fail thresholds applied to smoothly improving underlying behavior. But the mirage story is not the whole story either. Later work suggests that once performance is indexed by pretraining loss rather than parameter count, some tasks still exhibit threshold-like transitions from random guessing to useful behavior. The strongest conclusion is therefore narrower than both sides often claim: apparent emergence by model size is often a property of the measurement pipeline, while real threshold behavior may still exist in training or loss space.

Related Work

The primary source is Schaeffer, Miranda, and Koyejo, Are Emergent Abilities of Language Models a Mirage? (2023). Their central claim is methodological. If the underlying probability of success rises smoothly with scale, then nonlinear or discontinuous evaluation metrics can manufacture abrupt-looking jumps. They back that claim with a simple mathematical argument, a meta-analysis of BIG-Bench-style tasks, and empirical checks on GPT-3 and InstructGPT families.

The backdrop is Wei et al., Emergent Abilities of Large Language Models (2022), which made the phenomenon famous by highlighting task curves that appeared unpredictable from smaller models. That paper mattered because it shifted attention from scaling laws for loss toward the harder question of downstream capability forecasting.

Two useful follow-ups sharpen the debate rather than simply resolving it. Hu et al. (2023) argue that some small models already show measurable progress on "emergent" tasks once evaluation resolution is increased, for example with more sensitive code-generation metrics. Xia et al. (2024) go further and ask whether pretraining loss, rather than parameter count, is the right horizontal axis. On their account, different model-size and data-size combinations with the same loss often land on similar task performance, and some tasks still show threshold-like behavior when loss crosses a critical region.

Method/Mechanism

The mirage mechanism is simple. Suppose a model's latent competence on a task improves smoothly from 1% to 5% to 15% to 35% as scale increases. If we evaluate with a binary metric like exact-answer accuracy on a hard benchmark, the early regime can look like pure failure. A move from 1% to 5% still rounds to "almost never right." Only later, once the success probability passes the benchmark's coarse detection threshold, the curve looks like it suddenly turns on. Nothing discontinuous happened inside the model; the discontinuity came from mapping a smooth latent variable onto a hard observable.

This gets stronger when the metric is both nonlinear and low-resolution. Exact match ignores partial correctness. Pass/fail metrics collapse near misses and random guesses into the same bucket. Small test sets add statistical noise, so a tiny underlying gain may not even be detectable. Schaeffer et al. show that once one switches to more continuous metrics, better statistics, or repeated sampling, many canonical emergence curves become smoother and more predictable.

But there is a second mechanism that the mirage critique does not eliminate. If tasks depend on a prerequisite skill that remains below random-guess level until training loss falls under some value, then the observed transition can still be genuinely threshold-like. Xia et al. argue that this is easier to see when models are compared by pretraining loss rather than by raw parameter count. In that view, the key question is not "At how many parameters does ability X appear?" but "What learning state must the model reach before signal rises above the random baseline?"

Key Findings

Two case studies make the distinction concrete:

Four crisp insights follow:

A fifth implication matters for forecasting. If many dramatic-looking jumps are evaluation artifacts, then capability prediction should be less mystified and more measurement-driven. But if some tasks really do sit behind shared loss thresholds, then smooth scaling laws for training loss still will not automatically translate into smooth progress on every downstream benchmark.

Limitations

The mirage argument is strongest against a specific target: claims that sharpness in benchmark curves directly implies abrupt internal capability change. It is weaker as a universal debunking story. A benchmark can be badly measured and still be tracking a real latent threshold. Likewise, replacing binary metrics with continuous ones does not guarantee that all discontinuities disappear; it only removes one major source of artifact.

The loss-threshold view also has limits. Pretraining loss depends on tokenizer choice, corpus distribution, and architecture, so thresholds are not obviously portable across model families. The result is more satisfying scientifically than raw parameter thresholds, but it is not yet a universal coordinate system for capability forecasting. There is also a practical issue: many deployed models are evaluated after instruction tuning or RLHF, where downstream behavior can shift for reasons not well summarized by base-model loss alone.

Future Directions

The most important next step is benchmark refactoring. For every task that allegedly exhibits emergence, researchers should report at least one continuous or higher-resolution metric alongside the headline binary one. A second direction is mechanistic: if some skills truly turn on around a shared loss threshold, what internal circuits or representations become available there? That would connect the emergence debate to the more concrete circuit-birth stories seen elsewhere in interpretability.

A third direction matters for alignment. Safety evaluations often rely on thresholded outcomes such as "refused" versus "did not refuse" or "passed" versus "failed." If those outcomes hide smooth latent changes in dangerous capability, then governance decisions based on coarse dashboards may systematically underestimate gradual risk accumulation.

Open question: which frontier-relevant benchmarks still show robust threshold behavior after controlling for metric discontinuity, sample size, prompting, and pretraining-loss alignment?

Summary

Many celebrated emergent abilities in LLMs are better understood as measurement artifacts than as magical capability births. Schaeffer et al. make the core case: coarse nonlinear metrics can turn smooth performance gains into sharp-looking jumps. Hu et al. show that higher-resolution evaluation can recover predictable scaling that the coarse metrics conceal. Xia et al. then recover a narrower, more defensible notion of emergence by tying it to pretraining-loss thresholds rather than parameter count. The net result is a cleaner research program: distrust dramatic plots by default, upgrade the measurement stack, and only then ask whether a real capability threshold remains.

References