SoulMate integrates Retrieval-Augmented Generation (RAG) and Low-Rank Adaptation (LoRA) to generate customized responses and learn from user feedback instantly, enabling updates in about 0.2 seconds.

Potential applications include personalized AI assistants on smartphones, emotion-aware interactions, and companion robots capable of nuanced emotional exchanges.

SoulMate operates at ultra-low power—about 9.8 milliwatts—with simultaneous learning and inference, delivering responses in roughly 0.2 seconds and enabling 24/7 on-device operation.

SoulMate is presented as a mobile intelligence SoC that can both infer and train on-device, with a focus on real-time personalization and privacy.

The chip uses a mixed-rank architecture that prioritizes important tokens to boost efficiency, delivering faster responses and significant energy savings.

Commercialization is planned for 2027 through the faculty-founded startup OnNeuro AI, with interest from industry players and SoulMate highlighted as a top paper at ISSCC.

The ISSCC showcase featured the work as a Highlight Paper, led by PhD student Seong-yeon Hong, under the title “SoulMate: A 9.8mW Mobile Intelligence System-on-Chip with Mixed-Rank Architecture for On-Device LLM Personalization,” underscoring real-time adaptation to user reactions.

Professor Yoo envisions SoulMate as a true companion AI that deeply understands users while preserving privacy, laying the groundwork for personal, portable intelligence.

KAIST researchers, led by Professor Hoi-Jun Yoo, unveiled SoulMate, touted as the world's first on-device AI semiconductor that learns and adapts to a user’s speech, habits, and emotions in real time, without relying on cloud services.

The project has backing from Korea’s MIST and IITP programs, with public notes released by Mirage News.

Data stays on-device, ensuring privacy and reducing leakage risk compared to cloud-based AI, as SoulMate aims to become a truly hyper-personalized assistant.