How Predictable Are Large Language Model Capabilities? A Case Study on BIG-bench
This work addresses the need for efficient and accurate prediction of LLM performance, which is incremental but practical for users, developers, and researchers in AI.
The study tackled the problem of predicting large language model (LLM) capabilities using past experiment records, achieving an R^2 score greater than 95% with an MLP-based predictor and identifying a subset of tasks 3 times smaller than BIG-bench Hard that is equally informative for evaluating new model families.
We investigate the predictability of large language model (LLM) capabilities: given records of past experiments using different model families, numbers of parameters, tasks, and numbers of in-context examples, can we accurately predict LLM performance on new experiment configurations? Answering this question has practical implications for LLM users (e.g., deciding which models to try), developers (e.g., prioritizing evaluation on representative tasks), and the research community (e.g., identifying hard-to-predict capabilities that warrant further investigation). We study the performance prediction problem on experiment records from BIG-bench. On a random train-test split, an MLP-based predictor achieves an $R^2$ score greater than 95%, indicating the presence of learnable patterns within the experiment records. We then formulate the problem of searching for "small-bench," an informative subset of BIG-bench tasks from which the performance on the full set can be maximally recovered. We find a subset as informative as BIG-bench Hard for evaluating new model families, while being $3\times$ smaller. Additionally, we find competitive subsets by clustering task representations learned by our MLP-based predictor and selecting tasks close to cluster centroids, highlighting the importance of task diversity in constructing "small-bench."