KAIST’s Graduate School of AI Semiconductor, led by Professor Hoi-Jun Yoo, announced the development of a personalized large language model (LLM) accelerator named “Soulmate” that evolves based on user characteristics. This advancement is seen as key technology that moves beyond “AI for everyone” to enable a “hyper-personalized AI” that adapts to a person’s conversational style and preferences.
Soulmate employs on-device AI technology, processing data on the device itself rather than through external servers (clouds). The research team integrated Retrieval-Augmented Generation (RAG) for generating personalized responses from stored conversations and Low-Rank Adaptation (LoRA) for instant learning from user feedback directly into the semiconductor. This enables a real-time personalized AI system with a response time of 0.2 seconds.
The device also employs a mixed-rank architecture to optimize processing based on the importance of information, dramatically reducing power consumption. The semiconductor operates at an ultra-low power of 9.8 milliwatts (mW), approximately 1/500th of a smartphone processor’s power consumption, enabling complex learning and reasoning functions without depleting battery life.
Moreover, all personal data is processed internally, implementing a “secure AI” structure that fundamentally prevents data breaches. The team anticipates that this technology will herald the era of true personalized AI services when combined with next-generation platforms such as smartphones, wearables, and personal AI devices.
Professor Yoo stated that this research provides the technological foundation for AI to become a true companion by mimicking the process of friendship-building, moving beyond being a mere tool to becoming a “best friend” that understands individuals perfectly while protecting privacy.
This research, with Dr. Hong Seong-yeon as the first author, was recognized as a “highlight paper” at the International Solid-State Circuits Conference (ISSCC) in San Francisco, drawing global academic attention. The team successfully demonstrated a prototype semiconductor chip that dynamically adapts its response style in real-time to user reactions. The Soulmate AI semiconductor is expected to be commercialized next year through the startup company Onerio AI, founded by faculty members.