Are Emergent Abilities of Large Language Models a Mirage?

The paper on emergent abilities showed tasks where models sit at chance until some scale and then jump. This paper asks whether the jump is in the models or in the measurement.

The argument starts from something everyone agrees on: per-token loss falls smoothly with model size. Suppose the chance of producing each correct token rises smoothly toward 1. If a task is scored by exact match on a five-token answer, the score is roughly that chance to the fifth power, which sits near zero for a long time and then climbs steeply. If the same outputs are scored by token edit distance, the score tracks per-token error and moves smoothly. Multiple-choice grade, which gives 1 only when the right option has the most probability, is a step function of the same kind; Brier score, which measures the probabilities themselves, is smooth. Small test sets add a second effect: a small but real success rate rounds to zero.

The authors test this three ways. On GPT-3 and InstructGPT models doing 2-digit multiplication and 4-digit addition, the jump under exact match on long targets disappears under edit distance, and with more test data even exact match is above chance at every size. On BIG-Bench, emergent abilities concentrate in a few metrics: over 92% of hand-annotated cases use Multiple Choice Grade or Exact String Match, and LaMDA's jumps under Multiple Choice Grade vanish under Brier score. In vision models, choosing a thresholded metric produced emergence-like curves where none had been reported.

This is a paper about evaluation design, and every deployment has an evaluation.

All-or-nothing metrics hide progress. A document extraction scored only on "every field correct" can sit at 10% while field accuracy climbs from 70% to 95%. Report the strict metric the customer cares about and a partial-credit one beside it, so the team can see whether changes are moving things.

Small test sets make rare successes invisible. With 100 test items, a model that succeeds 0.1% of the time scores zero 90% of the time. Before declaring a task impossible for a model, check whether the test set is large enough to see the rate you care about.

Forecasts need smooth metrics. If a customer asks whether a bigger model or more fine-tuning data will fix a task, plot a continuous metric across sizes or data amounts. The strict metric will look flat until it does not.

Watch for jumps in your own dashboards. A sudden change in a pass-rate chart after a model upgrade may be a threshold effect on a metric, not a new capability.