Abstract
Recent work claims that large language models (LLMs) display emergent abilities—capabilities that are absent in smaller models but appear suddenly and unpredictably at larger scales. This paper presents an alternative explanation: that emergent abilities are not fundamental changes in model behavior, but rather artifacts of the researcher's choice of evaluation metrics. Specifically, nonlinear or discontinuous metrics produce apparent jumps in performance, whereas linear or continuous metrics reveal smooth, predictable improvements. By analyzing the InstructGPT/GPT-3 family, performing a meta-analysis on BIG-Bench, and inducing "emergence" in simple vision models, we demonstrate that these alleged abilities evaporate with different metrics or better statistics. Our findings suggest that emergence may be a mirage rather than a fundamental property of scaling AI models.
Introduction
The AI community has recently been captivated by the phenomenon of "emergence." In the context of Large Language Models (LLMs), emergence refers to the observation that scaling up parameters or compute leads to sudden, sharp transitions in performance on specific tasks, such as multi-digit addition or logical reasoning. These transitions are characterized by two features: sharpness (appearing instantaneously at a certain scale) and unpredictability (occurring at scales not previously foreseen).
If true, emergence poses a significant challenge for AI safety and alignment. If models can spontaneously develop new, powerful capabilities without warning, predicting and mitigating risks becomes exponentially harder. However, this paper argues that these "jumps" are largely measurement artifacts. By changing how we score model outputs, we can make emergent abilities appear or disappear at will.
The Mathematical Intuition: Metric as a Lens
The core of the argument lies in how metrics transform underlying model performance. Consider a model where the probability of predicting a single token correctly improves smoothly and predictably as the model scales (following a standard power law).
If we evaluate this model using a linear metric, such as the per-token error rate, the performance curve remains smooth. However, if we use a nonlinear, discontinuous metric like "Accuracy" or "Exact String Match"—where a multi-token sequence is only considered correct if every token is right—the resulting curve will show a sharp, emergent-like jump. As the per-token success rate crosses a certain threshold, the probability of getting an entire long sequence correct shifts from near-zero to significant, creating the illusion of a sudden breakthrough.
Empirical Evidence
1. Arithmetic in the GPT-3 Family
To test this hypothesis, we examined the InstructGPT/GPT-3 family on multi-digit multiplication. When measured by "Accuracy" (the standard metric used to claim emergence), the models appear to suddenly "learn" multiplication at a specific scale. However, when we re-evaluated the same model outputs using "Token Edit Distance"—a more continuous metric that gives partial credit for partially correct answers—the sharp jump vanished. Instead, the models showed a steady, linear improvement in their ability to predict the correct digits across all scales.
2. BIG-Bench Meta-Analysis
We conducted a meta-analysis of the BIG-Bench dataset, a massive collaborative benchmark. Our analysis revealed that emergent abilities are almost exclusively reported under specific metrics. Specifically, "Multiple Choice Grade" and "Exact String Match" account for over 92% of all claimed emergent abilities in the benchmark. When we analyzed the LaMDA model family, we found that swapping a nonlinear metric (Multiple Choice Grade) for a continuous one (Brier Score) caused the reported emergent abilities to disappear, restoring a predictable scaling relationship.
3. Inducing Emergence in Vision Models
To prove that emergence is a choice of the researcher rather than a property of the architecture, we intentionally "created" emergent abilities in simple, well-understood vision models. Using a LeNet (a convolutional network from the 1990s) on the MNIST dataset, we defined a new metric: the model is only "successful" if it classifies $K$ images correctly in a row. Under this discontinuous metric, the LeNet appears to have an "emergent" ability to classify sequences, despite the underlying model improving linearly. This demonstrates that emergence can be induced in any architecture simply by manipulating the evaluation criteria.
Discussion and Implications
Our findings suggest that the "sharpness" of emergence is often a result of the metric's lack of resolution. Smaller models often appear to have "zero" ability because the test sets are too small to capture their low-but-improving success rates. When the resolution is increased (through more data or more sensitive metrics), the continuity of improvement becomes clear.
This has significant implications for how we study and govern AI. If performance improvements are predictable and continuous, we can better anticipate the capabilities of future models. It also suggests that the field needs more robust benchmarking practices, including the use of multiple metrics and the public release of model outputs to allow for independent audits of claimed breakthroughs.
Conclusion
While we do not claim that LLMs cannot have emergent properties, we provide strong evidence that the currently claimed instances of emergence are likely measurement artifacts. By choosing metrics that are continuous and have high resolution, the "mirage" of sudden jumps evaporates, revealing a more predictable path of progress in AI scaling. Emergence, it seems, is often in the eye of the beholder—or more accurately, in the choice of their metric.
Summary updated