AI Tech Flash ── 2026-11-02
KAIST Cuts AI Fine-Tuning Time by Up to 69.6%
KAIST develops CluSTER, cutting fine-tuning time by up to 69.6%
What changes
According to Digital Today, a KAIST research team has developed a technology that cuts additional AI model training time by up to 69.6 percent. Fine-tuning means training an existing AI model further with domain-specific data, such as coding or medicine, to improve its performance.
KAIST said the method used only about 31 percent of the total data while largely maintaining model performance. That is enough to finish training that previously took 10 hours in about 3 hours.
Why it matters
According to KAIST, multiple GPUs usually split the data during training. A GPU is a processor used for AI computation. Repeatedly processing data with similar training effects can reduce training efficiency relative to the amount of computation.
Professor Lee Jae-gil said that to improve training efficiency, it is important not only to use more GPUs but also to choose intelligently which data to train on. He said the study can reduce the training burden of large-scale AI models by using existing GPU resources more efficiently.
What it is
CluSTER is a fine-tuning technique for large language models. A large language model is a text-generating AI trained on large amounts of text.
According to Digital Today, it selects data with similar training effects and allocates it efficiently across multiple GPUs.
Who and where
The technology was developed by a team led by computer science professor Lee Jae-gil at KAIST, the Korea Advanced Institute of Science and Technology. Doctoral student Kim Hyun-jin of the same department is the first author.
The findings were presented in October at EMNLP 2026, an international conference in natural language processing. Digital Today does not say whether the technology will be made publicly available.
Why it was built
In fine-tuning, multiple GPUs typically split the data for processing. KAIST said that repeatedly processing data with similar training effects can lower efficiency relative to the amount of computation.
The problem the team aimed to solve is this redundancy: keeping performance while using less data and less time.
How it works
According to Digital Today, the team first analysed how data affects AI training. It then grouped data that produces similar effects.
The team selected the necessary data from each group and distributed it evenly across multiple GPUs. This reduced redundant training while reflecting diverse training information.
Key points
- KAIST said that using about 31 percent of the total data cut training time by up to 69.6 percent.
- The team fine-tuned models with about 7 billion and 13 billion parameters on three coding and medical datasets, verified through eight performance evaluations.
- The tests focused on a four-GPU environment, and the team also confirmed better performance than existing methods in an eight-GPU environment.
Conclusion
The KAIST team said CluSTER cuts fine-tuning time by up to 69.6 percent while largely maintaining performance.
The verification covered two model sizes and three datasets. That is the extent of what Digital Today's report establishes.
Quotes and sources
“The experiments showed that using only about 31 percent of the total data reduced training time by up to 69.6 percent while largely maintaining model performance.”── 디지털투데이 (Digital Today), 2026-11-02
“Fine-tuning is a process that improves performance by training an existing AI model further with domain-specific data such as coding or medicine.”── 디지털투데이 (Digital Today), 2026-11-02
“To improve the efficiency of AI training, it is important not only to use more GPUs but also to choose intelligently which data to train on,”── 디지털투데이 (Digital Today), 2026-11-02
Source articles
- 디지털투데이 11/02 08:58
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.