Jeong Jin-woo, a master’s student in the Department of Artificial Intelligence at Kookmin University (President Jeong Seung Ryul) and a member of Professor Kim Jangho’s research team, has developed a new model ensemble technique that effectively enhances the performance of quantized AI models, with support from an industry-academia collaboration project funded by LG Electronics’ CTO Division. The research findings were presented at the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2026, a world-renowned international academic conference in the fields of data mining and artificial intelligence.

Titled “Quantized Model Soup Shake-Up: Weight Perturbation for Enhanced Ensemble Diversity,” the study—in which Master’s student Jeong Jin-woo participated—analyzed the issue of reduced inter-model diversity in quantized AI models and proposed a new weight adjustment technique to address it.

As AI models have recently become increasingly large, the importance of model lightweighting technologies—which enable the efficient operation of high-performance AI even in environments with limited computational resources, such as smartphones, home appliances, and robots—has grown significantly. Quantization, a leading lightweighting technique, reduces memory usage and computational load by converting model weights to lower-precision values; however, it has the drawback that performance may degrade in low-bit environments due to limited model expressiveness.

The research team identified a problem: when applying the existing Model Soup approach to quantized models, the weights of different models converge to the same quantization values, reducing diversity among models and thereby limiting the ensemble effect. To address this, they proposed “Quantized Model Soup Shake-Up (QMSS),” which selectively modifies low-importance weights near quantization boundaries to increase diversity among models. This approach enhances the effectiveness of model ensemble without incurring additional inference costs, while maintaining the performance of individual models.

This research is significant in that it presents a new approach: rather than simply increasing the size of AI models to improve accuracy, it leverages the characteristics of the weight space that arise during the quantization process to enhance the performance of lightweight models themselves. In particular, it is expected to be applicable in on-device AI and edge AI environments, where high performance and limited computational resources must be balanced simultaneously.

Furthermore, this achievement is significant as it resulted from an industry-academia collaboration project between LG Electronics’ CTO Division and a research team at Kookmin University. By linking the challenges of AI model lightweighting and efficient inference—which are demanded in real-world industrial settings—with the university’s research on AI model optimization, the project demonstrated a successful example of fundamental technology research through industry-academia collaboration.

Professor Kim Jangho of the Department of Artificial Intelligence at Kookmin University stated, “As on-device AI expands, the importance of model optimization technology—which ensures high performance even in limited hardware environments—is growing even more.” He added, “This research is significant in that it did not view the characteristics of quantized models as mere constraints, but rather utilized them from a new perspective to ensure diversity among models.”

This research aligns with the future specialization strategy that Kookmin University is pursuing through “KMU Vision 2035: EDGE.” In particular, as part of the “AI+X” field—one of the university’s eight specialized areas—this research achievement demonstrates the university’s specialization direction by linking AI model optimization technology to actual industrial needs through industry-academia collaboration. Previous press releases from Kookmin University have also highlighted AI model optimization research conducted with industry partners as a key example of its “AI+X” specialization.

Meanwhile, the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) is recognized as the world’s most prestigious international academic conference in the fields of data mining and machine learning. Researchers from leading universities and global research institutions around the world participate in the conference to present cutting-edge research findings in artificial intelligence and data mining.