KAIST method cuts LLM fine tuning time by up to 69.6%

KAIST announced CluSTER in a press release, a method that reduced large language model fine tuning time by up to 69.6% while maintaining nearly the same performance. Training that previously took 10 hours could take about three hours under the highest measured reduction.

CluSTER groups training examples according to their effects on model learning, selects data from each group, and distributes the groups across multiple GPUs. This reduces repeated computation on examples that teach similar skills while preserving varied training information.

In the team's own experiments, researchers fine tuned models with about 7 billion and 13 billion parameters on three coding and medical datasets and eight benchmarks. The models retained nearly the same performance when trained on about 31% of the full data. Tests mainly used four GPUs, with additional results showing improvements over conventional methods in an eight GPU environment. The research was presented at EMNLP 2026 in October.