LG Uplus Papers Accepted at NeurIPS 2026

LG Uplus said on the 29th that two research papers on token optimization, aimed at speeding up AI responses and reducing battery use, were accepted to the On-Device Intelligence workshop at NeurIPS 2026. NeurIPS is regarded as one of the world’s three leading AI conferences, alongside ICML and ICLR. This year’s event will take place in Sydney, Australia, from December 6 to 12; workshops are sessions devoted to specific research topics, separate from the main conference.

The researchers studied ways to run and train small language models (sLMs) efficiently within smartphones’ limited memory and power. They adapted an sLM for phone-call services using model compression and inference acceleration. LG Uplus said the system maintained performance at the level of the latest cloud AI while processing requests about 4.5 times faster than the existing method and reducing battery use by about 77%. The approach is designed to use limited computing resources and battery power efficiently while minimizing performance loss.

A separate study examined additional AI training on a smartphone without sending personal data outside the device. The researchers optimized the computing and memory needed for training, and LG Uplus said tests in real smartphone conditions achieved accuracy at the same level as server training. The company is considering ways to use the research in various AI services. It plans to extend the optimization work proven on smartphones to server graphics processing unit (GPU) environments so customers can use AI faster and more conveniently.

LG Uplus said on the 29th that two research papers on token optimization, aimed at speeding up AI responses and reducing battery use, were accepted to the On-Device Intelligence workshop at NeurIPS 2026. NeurIPS is regarded as one of the world’s three leading AI conferences, alongside ICML and ICLR. This year’s event will take place in Sydney, Australia, from December 6 to 12; workshops are sessions devoted to specific research topics, separate from the main conference.

The researchers studied ways to run and train small language models (sLMs) efficiently within smartphones’ limited memory and power. They adapted an sLM for phone-call services using model compression and inference acceleration. LG Uplus said the system maintained performance at the level of the latest cloud AI while processing requests about 4.5 times faster than the existing method and reducing battery use by about 77%. The approach is designed to use limited computing resources and battery power efficiently while minimizing performance loss.

A separate study examined additional AI training on a smartphone without sending personal data outside the device. The researchers optimized the computing and memory needed for training, and LG Uplus said tests in real smartphone conditions achieved accuracy at the same level as server training. The company is considering ways to use the research in various AI services. It plans to extend the optimization work proven on smartphones to server graphics processing unit (GPU) environments so customers can use AI faster and more conveniently.