One of the most exciting and most worrying ideas about large language models is that abilities can emerge. A small model cannot do arithmetic at all. A bigger one cannot either. Then a much bigger one suddenly can. If skills can appear without warning, we cannot predict what the next model will do.
A Stanford team looked at the charts behind this idea and noticed something odd. The sudden jumps appear mostly when a test is scored in a particular way. Their paper asks whether those jumps are a mirage created by the ruler. It won an Outstanding Paper award at NeurIPS 2023.
The authors argue that many claimed emergent abilities come from the metric, not from the model. Metrics that give credit only for a perfectly correct answer make smooth progress look like a sudden jump. Score the same answers with credit for partial success, and the jump disappears.
The claim: skills that switch on
The earlier work defined an emergent ability as one that is absent in smaller models and present in larger ones, so it cannot be predicted by extrapolating from small models. It has two striking features: sharpness, going from absent to present almost at once, and unpredictability, appearing at a scale nobody could foresee.
This matters for safety. If a dangerous ability could appear suddenly at some scale, warning signs might never show up. That makes it important to check whether the jumps are real.
The suspect: an all-or-nothing ruler
Think about a multiplication question whose answer has five digits, and a model that gets each digit right 80 percent of the time. It is a good model. But to be marked correct the whole answer must match, and the chance of that is 0.8 multiplied five times, about one in three. A weaker model that gets 50 percent of digits right scores about 3 percent.
That is the heart of the paper. If a model's per-token accuracy improves smoothly with size, then a metric that requires every token to be right bends the curve. It stays near zero for a long time and then shoots up. A metric that gives partial credit, such as counting how many tokens differ from the target, shows a steady climb.
