How Large Language Models Develop Unexpected Skills

2024

A recent study challenges the notion that large language models (LLMs) acquire emergent abilities suddenly and unpredictably. The study, conducted by researchers at Stanford University, suggests that these abilities actually develop gradually and predictably, depending on how they are measured.

LLMs, like the ones powering chatbots such as ChatGPT, learn by analyzing vast amounts of text data. As the size of these models increases, so does their ability to complete tasks, including ones for which they were not explicitly trained. This growth in performance has led to the perception of emergent abilities in LLMs, which are collective behaviors that appear once a system reaches a high level of complexity.

However, the Stanford researchers argue that the perception of emergence is influenced by how LLMs are measured. They conducted experiments with addition tasks, showing that the ability to add did not emerge suddenly at a certain threshold, as previously thought. Instead, they found that as the size of the LLM increased, its ability to predict the correct sequence of digits in addition problems improved gradually and predictably when measured using a different metric that awarded partial credit.

While this study challenges the idea of emergence in LLMs, other researchers point out that it does not fully dispel the notion. Some argue that the unpredictability of emergent abilities lies in the difficulty of predicting which metrics will show abrupt improvement in an LLM. Nevertheless, this research highlights the importance of considering how we measure the abilities of LLMs and raises questions about how these models will continue to evolve in the future.

As LLMs grow larger and more complex, they are likely to exhibit new and unexpected behaviors. Understanding how these behaviors emerge and how they can be predicted is crucial for the development of AI technologies.

Source