On-Device Agentic AI: LG Uplus ixi-O Based on EXAONE

As the center of gravity in AI technology shifts from the cloud to personalized devices, LG Uplus's AI assistant ixi-O has been upgraded into an on-device agent by incorporating LG's foundation model, EXAONE.

This project focused on reducing infrastructure costs by lowering cloud dependency, and strengthening security by processing sensitive call data directly on the device. The key applied services are as follows.

• AI Call Analysis: Converts unstructured call scripts transcribed via STT into structured information, including summaries, keyword and entity extraction, and follow-up action recommendations (such as schedule registration).
• AI Call Answering: Moves beyond existing rule-based responses to a conversational structure that grasps the caller's intent and exchanges dialogue naturally.

For the base model, the lightweight EXAONE 4.0 1.2B model, optimized for on-device environments, was adopted. This model is sized appropriately for smartphone environments while showing excellent performance on key benchmarks such as knowledge reasoning, instruction following, and Korean language understanding, and supports long context at the 65K level, which is advantageous for processing long call scripts.

For data training, a pipeline was built in which a high-performance Teacher Model generates labels, which are then verified and supplemented before being used to train the Student Model, EXAONE. In addition, all tasks were trained using a multi-task approach to minimize memory usage and maximize operational efficiency.