The Pulse
KAIST’s CluSTER Reports 69.6% Time Cut in Preprint, 64.6% in EMNLP Paper
KAIST says its CluSTER method cut fine-tuning time by up to 69.6% in tests using a reduced dataset, while the published EMNLP 2026 paper reports a 64.6% reduction. The approach groups training examples by their learning effects and distribu
AI.info Team · Written with AI · fact-checked ·
KAIST researchers say fine-tuning can take less time not by adding GPUs, but by choosing and distributing training data differently. In its November 2, 2026 announcement, the university reported that its CluSTER method cut training time by up to 69.6% in tests, while models trained on about 31% of the full dataset retained nearly the same performance. The 69.6% figure also appears in the arXiv preprint; the published EMNLP 2026 paper reports that CluSTER reduces training time by 64.6%.
The work addresses a trade-off in adapting large language models: developers need examples to teach a model a specialized task, but processing examples with similar learning effects can consume computing time while adding little new learning. CluSTER aims to reduce that repetition and distribute varied examples across GPUs.
CluSTER groups examples by learning effect
Rather than grouping training examples only by topic or wording, CluSTER groups them according to how they affect model learning. KAIST’s example is two arithmetic problems with different objects—apples and pencils—that teach the same addition skill. The method selects examples from these groups and allocates them across GPUs so the devices process a varied range of learning information.
The published paper describes the method as gradient-space clustering and data-parallelism-aware balanced allocation. It says the allocation seeks coverage across both clusters and GPU workers, while cluster-weighted updates preserve the original data distribution. That weighting description comes from the paper, not the KAIST news release.
The paper’s authors—Hyunjin Kim, Youngeun Nam, Jaemin Han, Wonhyeok Choi and Jae-Gil Lee—write:
“We propose CluSTER, a Cluster-aware balanced Sampling framework for Training Efficient data Reduction in DP instruction tuning.”
Tests covered coding and medical tasks
The team fine-tuned models with approximately 7 billion and 13 billion parameters using three training datasets from coding and medical fields. KAIST says the researchers evaluated the method with eight performance benchmarks. In the university’s account, models trained on about 31% of the full dataset retained nearly the same performance, while training time fell by as much as 69.6%.
The figure depends on which version is being described. KAIST’s announcement and the arXiv preprint report a reduction of up to 69.6%; the published paper in the EMNLP 2026 proceedings reports 64.6% across multiple instruction-tuning datasets. The figures should not be treated as interchangeable. KAIST translates its reported maximum reduction into a comparison: a training run taking 10 hours could finish in about 3. The researchers primarily tested with four GPUs and also reported better performance than conventional methods in an eight-GPU environment.
Those figures describe the researchers’ tests, not a guarantee for other models, datasets or hardware. KAIST says the approach could help organizations adapt models for tasks such as coding assistance, customer service and medical information retrieval, reducing the time and GPU costs of additional training.
“To improve the efficiency of AI training, it is important not only to use more GPUs, but also to intelligently select which data should be used for learning,”
— Jae-Gil Lee, professor in KAIST’s School of Computing
The research findings were presented at EMNLP 2026 in October. The paper, “Clustering-Based Balanced Sampling and Allocation with Data Parallelism for High-Performance Fine-Tuning,” was published in the conference proceedings.