Random Scaling of Emergent Capabilities

Language models famously improve under a smooth scaling law, but some specific capabilities exhibit sudden breakthroughs in performance. Advocates of "emergence" view these capabilities as unlocked at a specific scale, but others attribute breakthroughs to superficial metric thresholding effects. We propose that breakthroughs are instead driven by continuous changes in the probability distribution of training outcomes when performance is bimodally distributed across random seeds. We show that different random seeds can produce either smooth or emergent scaling trends in synthetic length generalization tasks, multiple choice question answering, and grammatical generalization. We reveal that sharp breakthroughs in metrics are produced by underlying continuous changes in their distribution across seeds. These distributions may become abruptly bimodal at a capacity threshold—but this threshold appears at scales well before most seeds achieve breakthrough. Our observations hold true even under continuous loss metrics, confirming that random variation must be considered when predicting a model's performance from its scale.

When AI language models get bigger, they usually get better in a smooth, predictable way. But sometimes a skill seems to appear out of nowhere: a model is terrible at a task, then suddenly great, like flipping a switch. People have argued about why this happens, and whether it means big AI models have hidden "breakthrough" moments waiting to be unlocked. We found that the switch isn't really sudden at all. We trained many copies of the same model, changing only the random starting conditions, and watched what happened. Some copies learned the skill and some didn't, and the chance of learning it grew gradually as models got bigger, even though each individual copy looked like it flipped a switch. This matters because it shows luck plays a big role in what an AI can do, so we shouldn't trust a single model to tell us what a given size is capable of.