KAIST CluSTER Triples AI Fine-Tuning Speed

KAIST researchers developed CluSTER, a method that groups data by learning effect to distribute it across GPUs. This approach reduced training time by up to 69.6% while maintaining nearly the same performance. Experiments showed models trained on 31% of data completed in about three hours instead of ten. The findings were presented in October 2026 at EMNLP 2026.