LG Uplus said on Oct. 29 that two papers on AI token optimization technology were accepted for the on-device intelligence workshop at NeurIPS 2026, one of the world’s three major AI conferences.

NeurIPS is considered one of the world’s three major AI conferences along with ICML and ICLR. This year’s event will be held in Sydney, Australia, from Dec. 6 to 12. The research focuses on technology to run and train small language models efficiently even in smartphones’ limited memory and power environments.

LG Uplus developed a small language model for use in its call service and applied model compression and inference acceleration technology. It optimized smartphone computing resources and power use while maintaining performance at the latest cloud AI level. As a result, it increased processing speed about 4.5 times compared with existing methods and reduced battery consumption about 77 percent.

It also developed technology to train AI inside smartphones without sending personal data to external servers. It reduced the computing and memory use needed for training to enable additional on-device learning. Verification in an actual smartphone environment confirmed accuracy at a level similar to a model trained on a server, it explained.

LG Uplus is reviewing ways to apply the research results to its AI services. It plans to expand optimization technology secured in the smartphone environment to server GPU environments to increase processing speed and efficiency for AI services.

LG Uplus Chief Technology Officer Sang-yeop Lee (이상엽) said, "This NeurIPS paper acceptance recognizes research achievements aimed at running and training AI quickly and efficiently even in constrained environments." He added, "We will apply related technology to actual services and later expand the scope of research to server environments."