Google Research scientists Morteza Zadimoghaddam and Matthew Fahrbach introduced GIST, a new algorithm for selecting optimal data subsets. It balances data diversity and data utility with a provable guarantee, addressing the challenge of training models on massive datasets without processing all the data.

Modern machine learning, especially with large language models and computer vision, requires handling enormous datasets. Subset selection is crucial for efficient training, but balancing diverse, non-redundant data with high-utility information is an NP-hard problem. GIST tackles this by approximating a series of optimization problems.

GIST uses a bicriteria greedy algorithm that iterates through distance thresholds to find the best trade-off. It provides a strong mathematical guarantee, finding a subset worth at least half the optimal value. This offers a practical safety net for practitioners managing the diversity-utility conflict in real-world ML tasks.