[Paper Review] Are Emergent Abilities of Large Language Models a Mirage?

Rylan Schaeffer, Brando Miranda+1 more|arXiv (Cornell University)|Apr 28, 2023

Topic Modeling132 citations

TL;DR

The paper argues that so-called emergent abilities in large language models arise from metric choice and data limitations rather than fundamental scaling. It presents a simple mathematical model and three complementary tests showing emergent abilities can disappear with linear/continuous metrics, better statistics, or across vision tasks.

ABSTRACT

Recent work claims that large language models display emergent abilities, abilities not present in smaller-scale models that are present in larger-scale models. What makes emergent abilities intriguing is two-fold: their sharpness, transitioning seemingly instantaneously from not present to present, and their unpredictability, appearing at seemingly unforeseeable model scales. Here, we present an alternative explanation for emergent abilities: that for a particular task and model family, when analyzing fixed model outputs, emergent abilities appear due to the researcher's choice of metric rather than due to fundamental changes in model behavior with scale. Specifically, nonlinear or discontinuous metrics produce apparent emergent abilities, whereas linear or continuous metrics produce smooth, continuous predictable changes in model performance. We present our alternative explanation in a simple mathematical model, then test it in three complementary ways: we (1) make, test and confirm three predictions on the effect of metric choice using the InstructGPT/GPT-3 family on tasks with claimed emergent abilities; (2) make, test and confirm two predictions about metric choices in a meta-analysis of emergent abilities on BIG-Bench; and (3) show to choose metrics to produce never-before-seen seemingly emergent abilities in multiple vision tasks across diverse deep networks. Via all three analyses, we provide evidence that alleged emergent abilities evaporate with different metrics or with better statistics, and may not be a fundamental property of scaling AI models.

Motivation & Objective

• Question whether emergent abilities are intrinsic to model scaling rather than artifacts of measurement.
• Propose a mathematical model linking per-token error, metric choice, and observed emergence.
• Test predictions with InstructGPT/GPT-3 to assess metric effects on emergent arithmetic abilities.
• Conduct a meta-analysis of BIG-Bench results to evaluate metric dependence of emergent abilities.
• Demonstrate induced emergent abilities in vision tasks by altering evaluation metrics.

Proposed method

• Provide a simple mathematical model where per-token cross-entropy loss L_CE(N) follows a power-law with model size N, leading to a nonlinear mapping to task scores under nonlinear or discontinuous metrics.
• Show how nonlinear metrics (e.g., accuracy on long sequences) and discontinuous metrics (e.g., multiple-choice grade) can create sharp, emergent-looking transitions from small to large models.
• Demonstrate that continuous/linear metrics (e.g., token edit distance, Brier score) yield smooth, predictable improvements with scale, abating emergent behavior.
• Test three predictions on InstructGPT/GPT-3: metric changes reveal smooth improvements; higher-resolution statistics reveal nonzero small-model performance on nonlinear metrics; target length affects performance predictably.
• Meta-analyze BIG-Bench results to see emergent abilities appear mainly under a few nonlinear/discontinuous metrics and vanish under continuous metrics.
• Show how to induce emergent abilities in vision models by designing discontinuous metrics for reconstruction and sequential classification tasks.

Experimental results

Research questions

• RQ1 Do emergent abilities depend on the metric used to evaluate model performance?
• RQ2 Can nonlinear or discontinuous metrics create apparent sharp transitions with model scale that disappear under linear/continuous metrics?
• RQ3 Do higher-resolution statistics (more test data) reveal nonzero small-model performance on tasks with claimed emergent abilities?
• RQ4 Can emergent abilities be induced in non-language domains (vision) by altering evaluation metrics?
• RQ5 How robust are emergent claims in BIG-Bench to metric choice and model family?

Key findings

• Emergent abilities appear predominantly under nonlinear or discontinuous metrics such as Exact String Match and Multiple Choice Grade.
• Increasing test data resolution reveals above-chance performance for smaller models, indicating nonzero capabilities when measured with nonlinear metrics.
• Switching to linear or continuous metrics (e.g., Token Edit Distance or Brier Score) yields smooth, predictable improvements, eroding the emergent effect.
• Meta-analysis shows emergent abilities occur for at most a subset of metrics; two metrics (Multiple Choice Grade and Exact String Match) account for most claimed emergent abilities.
• Metrics can be used to induce new emergent-like abilities in vision models across architectures by designing appropriate evaluation criteria.
• Empirical results on InstructGPT/GPT-3 support that emergent abilities disappear under better statistics and with metric changes.