Are Emergent Abilities of Large Language Models a Mirage?

1 Introduction

1. How were “emergent abilities” of large language models defined, and what two properties make them intriguing?
   Abilities not present in smaller models but present in larger ones, so they can’t be predicted by extrapolating smaller models’ performance. Their sharpness (a seemingly instant jump from absent to present) and unpredictability (appearing at unforeseeable scales).
   • Schaeffer et al., 2023, Are Emergent Abilities of Large Language Models a Mirage? (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §1
2. Why do emergent abilities matter for AI safety, according to Schaeffer et al.?
   If abilities appear sharply and unpredictably, larger models might suddenly, without warning, acquire dangerous capabilities.
   • Schaeffer et al., 2023, Are Emergent Abilities of Large Language Models a Mirage? (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §1
3. What alternative explanation do Schaeffer et al. propose for emergent abilities?
   They are mostly a mirage created by the researcher’s analysis: chiefly a metric that nonlinearly or discontinuously transforms a smoothly improving per-token error rate, and secondarily too little test data to measure small models’ performance precisely, so they look wholly incapable.
   • Schaeffer et al., 2023, Are Emergent Abilities of Large Language Models a Mirage? (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §1
4. Which two metrics account for most claimed emergent abilities on BIG-Bench?
   Multiple Choice Grade (1 if the most probability mass is on the correct option, a discontinuous step) and Exact String Match (1 only if the whole output matches, nonlinear in the per-token error); together they cover over 92% of hand-annotated emergent abilities.
   • Schaeffer et al., 2023, Are Emergent Abilities of Large Language Models a Mirage? (Advances in Neural Information Processing Systems 36 (NeurIPS 2023)), §1