[Edaily Reporter Kim Hyun-ah ] LG Uplus(032640)has achieved success by having two papers accepted for the NeurIPS 2026 workshop—a leading global artificial intelligence (AI) conference—following its research into technology for efficiently running and training small language models (sLMs) on smartphones.

By applying model lightweighting and inference acceleration technologies, the company increased AI processing speed by approximately 4.5 times compared to existing methods while reducing battery consumption by about 77%.

LG Uplus announced on the 29th that two research papers on AI token optimization were accepted for the NeurIPS On-Device Intelligence Workshop. NeurIPS is considered one of the world’s top three AI conferences, alongside ICML and ICLR. This year’s conference will be held in Sydney, Australia, from December 6 to 12.

This research focused on efficiently running and training small language models in environments with memory and power constraints, such as smartphones. Tokens are the data units AI uses to understand questions and generate responses. Efficient token processing can speed up response times and reduce computational resource and power consumption.

LG Uplus developed a small language model for use in voice services and optimized it for the smartphone environment by applying model lightweighting and inference acceleration technologies. The company explained that, as a result, processing speed was increased by approximately 4.5 times compared to existing methods, while battery consumption was reduced by about 77 percent. The company focused on minimizing performance degradation while utilizing limited computational resources.

The company also researched technology that trains AI within the smartphone itself, without sending personal data to external servers. By reducing the computational and memory requirements for training, the company conducted additional training on the device and verified performance in a real-world smartphone environment. LG Uplus stated that it confirmed the same level of accuracy as models trained on servers, even without transmitting personal data outside the device.

LG Uplus is exploring ways to apply the technology developed through this research to its own AI services. The company plans to extend the optimization techniques validated on smartphones to server-based graphics processing unit (GPU) environments to improve the response speed and resource efficiency of its AI services.

Lee Sang-yeop, Chief Technology Officer (CTO) of LG Uplus, said, “The acceptance of this paper recognizes our research achievements in running and training AI quickly and efficiently even in constrained environments,” adding, “We will apply the relevant technology to actual services and expand our research scope to include server environments.”