KAIST's CluSTER data selection cuts LLM fine-tuning time by up to 69.6% using about 31% of the data
Lee Jae-gil's team groups training examples with similar effect and spreads them evenly across GPUs; tested on coding and medical datasets with models of about 700 million and 1.3 billion parameters and presented at EMNLP 2026, so gains at frontier scale are untested.