South Korean telecom operator LG U+ announced that two of its research papers on AI Token optimization technology have been accepted for the NeurIPS 2026 On-Device Intelligence workshop. This technology targets devices with limited computational resources, such as mobile phones, by reducing the computational overhead of AI models to improve processing speed and decrease power consumption. The optimization scheme for a small language model focused on call services can increase processing speed by approximately 4.5 times compared to traditional methods, while reducing battery consumption by about 77%.
AI models process information by breaking down content such as text into tokens, which are the basic units for computation. As on-device AI capabilities increase, mobile phones must run models within the constraints of limited memory, processing power, and battery capacity, making computational efficiency directly impact response speed and battery life. LG U+'s research focuses on token processing efficiency, aiming to reduce latency and energy consumption during the model inference process to enhance the efficiency of AI functions running locally on mobile devices.
LG U+'s AI Token Optimization Technology Significantly Enhances Mobile Performance
In the context of call services, LG U+ applied model lightweighting and inference acceleration techniques to small language models (sLM). Compared to relying directly on large cloud models, small language models have relatively lower computational resource demands, making them more suitable for operation on devices such as mobile phones. LG U+'s test results show that after optimization, the processing speed of the relevant models reached about 4.5 times that of the original method, while battery consumption decreased by approximately 77%.
Additionally, LG U+ has developed technology that allows AI models to continue training within smartphones. Traditional AI training typically requires data to be transferred to a server for processing, but this technology reduces computational load and memory usage, enabling models to learn from user data locally on the device. LG U+ stated that in actual smartphone tests, locally trained models achieved accuracy comparable to server-trained models, and user data did not need to leave the device, allowing for improved model adaptation to personal usage environments.
Item Specification Processing Speed Increase Approximately 4.5 times Battery Consumption Reduction Approximately 77%
Disclosure: This article contains links to products from partner merchants. If you purchase through these links, TechRitual may earn a commission, but this does not affect product evaluations and recommendations. For more details, please refer to the privacy policy.