Large language models can appear to gain new abilities abruptly as they get bigger. A Stanford University team argues that, at least in some cases, the surprise may come less from the model and more from the ruler used to measure it.
The puzzle behind sudden AI skills
Two years ago, 450 researchers created the Beyond the Imitation Game benchmark, or BIG-bench, to test large language models across 204 tasks. These models power chatbots like ChatGPT, and the benchmark was meant to probe what they could do as they scaled.
On many tasks, performance improved in a steady pattern: larger models did better. But on some tasks, the results looked different. Scores stayed near zero, then rose sharply, creating the impression that a model had crossed a hidden threshold.
The authors described this as “breakthrough” behavior. Other researchers compared it to a phase transition in physics, like water becoming ice. In a paper published in August 2022, researchers called such abilities “emergent,” meaning they seemed to appear only after a system became sufficiently complex.
That framing mattered because it suggested that large language models might develop new capabilities in ways that are hard to forecast. The question now is whether some of those jumps were truly sudden or whether the tests made gradual progress look abrupt.
What the Stanford paper argues
A new paper from three researchers at Stanford University challenges the strongest version of the emergence claim. Sanmi Koyejo, Rylan Schaeffer, and Brando Miranda argue that some apparent leaps are created by measurement choices.
Koyejo, a computer scientist at Stanford and the paper’s senior author, put the issue directly: “The transition is much more predictable than people give it credit for.” He added, “Strong claims of emergence have as much to do with the way we choose to measure as they do with what the models are doing.”
The basic point is simple. If a test gives credit only for a perfect answer, a model can look useless until it suddenly appears competent. But if the same test gives partial credit for getting closer to the answer, progress can look gradual.
The Stanford researchers are not arguing that bigger models fail to improve. They acknowledge that larger large language models can solve tasks smaller models cannot, including tasks they were not explicitly trained to perform. Their argument is narrower: whether improvement looks smooth or sudden can depend on the metric, and sometimes on how many test examples are available.
Why three-digit addition matters
The source article uses three-digit addition to show the stakes of the measurement problem. In the 2022 BIG-bench study, GPT-3 and another large language model named LAMDA appeared unable to solve addition problems at smaller scales. Then GPT-3 seemed to gain the ability at 13 billion parameters, and LAMDA seemed to do so at 68 billion parameters.
That result made addition look like an ability that emerged at a threshold. But the Stanford researchers noted that the models were judged by exact accuracy. A nearly correct answer still counted as a failure.
For example, if the problem is 100 plus 278, an answer of 376 is plainly closer than −9.34, even though neither is exactly correct. Under an all-or-nothing score, both can be treated the same.
Koyejo and his collaborators used a metric that gave partial credit. Instead of asking only whether the full answer was right, they looked at whether the model predicted each digit correctly. As model parameters increased, the models produced increasingly correct digit sequences.
That change in scoring made the supposed leap look more like a slope. In this case, the ability to add appeared gradual and predictable rather than sudden and mysterious.
Scale is still changing what models can do
The debate is happening because large language models have grown quickly. GPT-2 had 1.5 billion parameters. GPT-3.5, the model powering ChatGPT, uses 350 billion. GPT-4, which debuted in March 2023 and now underlies Microsoft Copilot, reportedly uses 1.75 trillion.
Parameters are a way of describing the many connections a model can form among words. Large language models train on enormous collections of text from online sources, including books, web searches, and Wikipedia. As models gain more parameters, they can represent more relationships in that text.
That scale has produced major improvements in performance. The dispute is not whether larger models become more capable. The dispute is how researchers should describe and predict those gains.
For AI safety, model evaluation, and product decisions, that distinction matters. If an ability truly appears without warning, prediction is harder. If the ability was developing all along but hidden by a blunt metric, then better tests may make future behavior easier to understand.
Why the debate is not over
Other researchers do not think the Stanford paper settles the question. Tianshi Li, a computer scientist at Northeastern University, said the paper does not fully explain which metrics will show abrupt improvement, or when they will do so. “So in that sense, these abilities are still unpredictable,” she said.
Jason Wei, a computer scientist now at OpenAI and an author on the BIG-bench paper, has argued that earlier reports of emergence were sound. For tasks such as arithmetic, his view is that the final correct answer is what matters.
Alex Tamkin, a research scientist at the AI startup Anthropic, saw value in the Stanford work but also limits. “But this is not the full story. We can’t say that all of these jumps are a mirage,” he said. He added that the literature still shows cases where models improve in a jump-like fashion even with one-step predictions or continuous metrics.
Xia “Ben” Hu, a computer scientist at Rice University, also pointed toward the next generation of models. As large language models grow, he said, they may draw knowledge from other tasks and other models.
The practical conclusion is cautious. Some dramatic AI breakthroughs may be less mysterious than they first appear. But researchers are still working toward a stronger science of prediction for large language models, especially as the systems become larger and more broadly useful.