KAIST's CluSTER cuts LLM fine-tuning time by up to 69.6%
The technique trains on about 31% of the data while nearly preserving performance, turning a 10-hour job into roughly 3 hours, according to digitaltoday.co.kr.
KAIST researchers have developed a large language model fine-tuning technique called CluSTER that reduces additional training time by up to 69.6% while nearly maintaining model performance, the university announced on November 2, 2026, according to digitaltoday.co.kr.
CluSTER groups data that produces similar training effects and selects the needed data from each group, distributing it evenly across multiple GPUs to cut redundant training. The team fine-tuned models of about 7 billion and 13 billion parameters on three coding and medical datasets and validated the method with 8 performance evaluations.
Using only about 31% of the full data nearly preserved model performance while reducing training time by up to 69.6%, turning a 10-hour training job into about 3 hours. Experiments centered on a 4-GPU environment, and 8-GPU setups also outperformed existing methods.
The research team is led by Lee Jae-gil, a KAIST computer science professor. A KAIST computer science PhD student is the study's first author.
Presented at EMNLP 2026
The research results were presented in October 2026 at EMNLP 2026, an international natural language processing conference.
Quick answers
What is CluSTER?
It is a large language model fine-tuning technique developed by a KAIST research team that groups data with similar training effects and distributes selected data evenly across multiple GPUs.
How much training time does CluSTER save?
It reduces additional training time by up to 69.6%, turning a 10-hour training job into about 3 hours, while using about 31% of the full data.
Where were the results presented?
They were presented in October 2026 at EMNLP 2026, an international natural language processing conference.