Predicting Emergent Capabilities by Finetuning
Charlie Snell, Eric Wallace, Dan Klein, Sergey Levine
2024-11-26
Summary
This paper explores how to predict the capabilities of future language models by fine-tuning existing ones, focusing on understanding when new abilities will emerge as models are developed.
What's the problem?
As language models (LLMs) grow in size and complexity, they sometimes develop unexpected new abilities, known as emergent capabilities. However, it's difficult to predict when these abilities will appear or how effective future models will be, making it challenging for researchers to plan for advancements in AI technology.
What's the solution?
The authors propose a method called emergence prediction, which involves fine-tuning current LLMs on specific tasks to determine when new capabilities might arise. They found that by adjusting the training data and analyzing the results, they could predict whether future models would perform well on certain tasks. They validated their approach using established benchmarks and demonstrated that they could accurately forecast the performance of larger models based on smaller ones.
Why it matters?
This research is important because it helps improve our understanding of how language models develop new skills as they are scaled up. By being able to predict these emergent capabilities, researchers can better design future models and optimize their training processes, ultimately leading to more powerful and effective AI systems.
Abstract
A fundamental open challenge in modern LLM scaling is the lack of understanding around emergent capabilities. In particular, language model pretraining loss is known to be highly predictable as a function of compute. However, downstream capabilities are far less predictable -- sometimes even exhibiting emergent jumps -- which makes it challenging to anticipate the capabilities of future models. In this work, we first pose the task of emergence prediction: given access to current LLMs that have random few-shot accuracy on a task, can we predict whether future models (GPT-N+1) will have non-trivial accuracy on that task? We then discover a simple insight for this problem: finetuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models. To operationalize this insight, we can finetune LLMs with varying amounts of data and fit a parametric function that predicts when emergence will occur (i.e., "emergence laws"). We validate this approach using four standard NLP benchmarks where large-scale open-source LLMs already demonstrate emergence (MMLU, GSM8K, CommonsenseQA, and CoLA). Using only small-scale LLMs, we find that, in some cases, we can accurately predict whether models trained with up to 4x more compute have emerged. Finally, we present a case study of two realistic uses for emergence prediction.