LG Uplus announced on October 29 that two of its research papers on artificial intelligence (AI) token optimization have been accepted for the on-device intelligence workshop at NeurIPS 2026.
NeurIPS is recognized as one of the top three AI conferences globally, alongside ICML and ICLR. It features presentations of the latest AI research by major global tech companies such as Google, Meta, and OpenAI, as well as leading university research teams. This year's event will take place from December 6 to 12 in Sydney, Australia.
As competition in AI services expands beyond model capabilities to how quickly and efficiently tokens can be processed, LG Uplus has focused on optimizing token handling. Tokens are the data units that AI uses to understand questions and generate responses. Faster and more efficient processing can enhance response times while reducing power and resource consumption.
LG Uplus has researched applying this token optimization technology in smartphone environments, enabling efficient execution and learning of small language models (sLM) even with limited memory and power. The company developed an sLM for use in call services and optimized it for smartphones by applying model lightweighting and inference acceleration techniques. As a result, it achieved a processing speed approximately 4.5 times faster than previous methods while reducing battery consumption by about 77%, all while minimizing performance degradation and efficiently utilizing the smartphone's limited computational resources and battery.
The company also explored technology that allows AI to learn directly on smartphones without sending personal data externally. By optimizing the computational and memory usage required for AI learning, LG Uplus conducted additional learning processes internally on the device. Performance validation in real smartphone environments showed that it achieved accuracy levels comparable to server-based learning without transmitting personal data outside the device.
LG Uplus is considering how to leverage the technologies developed in this research for its various AI services in the future. Specifically, it plans to expand the optimized capabilities validated on smartphones to server GPU environments, enabling customers to access AI services more quickly and conveniently.
Lee Sang-yeop, CTO of LG Uplus, stated, "The acceptance of our papers at NeurIPS recognizes our research achievements in executing and learning AI quickly and efficiently in constrained environments. We will expand our research scope to include server environments in the future."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.