"Conversing and Connecting Like a Friend": KAIST Develops Personalized AI 'SoulMate'

by Jeong Ilwoong

Published 17 Mar.2026 08:22(KST)

Just like a scene from a movie, an artificial intelligence (AI) semiconductor capable of understanding, resembling, and emotionally interacting with human thoughts and feelings has been developed in South Korea. This is a notable departure from the current large language models (LLMs) like ChatGPT, which can provide fluent answers in daily use but still fall short of genuine interaction with users.

On March 17, KAIST announced that the research team led by Professor Hyoejun Yoo at the Graduate School of AI Semiconductors has developed a personalized large language model (LLM) accelerator, called 'SoulMate,' which evolves on its own to fit each user's unique characteristics.

SoulMate is recognized as a core semiconductor technology that will usher in an era of 'hyper-personalized AI just for me,' going beyond 'AI for everyone' by learning and responding to each user's individual conversational style and preferences.

Professor Hweejun Yoo (right) and Hong Sungyun, doctoral researcher (left), are presenting their research results. KAIST

The key lies in its on-device AI technology, which processes data directly on the device without relying on external servers (cloud).

The research team has directly implemented retrieval-augmented generation (RAG) technology-which generates personalized answers based on conversations remembered by the AI-and low-rank adaptation (LoRA) technology-which immediately incorporates user feedback for learning-inside the semiconductor itself.

With this, SoulMate can respond to users within 0.2 seconds (216.4 ms) while simultaneously learning, making it usable as a real-time personalized AI system.

Another advantage of SoulMate is its mixed-rank architecture, which optimizes processing methods based on the importance of information, resulting in a dramatic reduction in power consumption.

The semiconductor that runs SoulMate can perform complex learning and inference tasks simultaneously with ultra-low power consumption of just 9.8 mW-about 1/500th of a typical smartphone processor-allowing it to operate on mobile devices like smartphones without battery concerns.

In particular, SoulMate features a 'security-complete AI' structure, where all personal data is processed only inside the device and never transmitted to external servers, fundamentally preventing any risk of personal information leakage.

The research team expects this technology to contribute to ushering in a true era of personalized AI services by combining with next-generation platforms such as smartphones, wearable devices, and personal AI devices.

Professor Yoo stated, "This research is meaningful in that it provides the technological foundation for AI to become a true companion by mimicking the process through which people build mutual friendships," adding, "In the future, AI will go beyond being a simple tool and become more like a 'best friend'-someone who perfectly protects your privacy and understands you better than anyone, anytime and anywhere."

Meanwhile, PhD researcher Seongyeon Hong participated as the first author of this study. The research results (paper) were selected as a 'Highlight Paper' at the International Solid-State Circuits Conference (ISSCC), held last month in San Francisco, USA, drawing significant attention from the academic community.

At the conference, the research team successfully demonstrated how the SoulMate AI semiconductor chip could instantly adjust its response style in real time based on user reactions. The SoulMate AI semiconductor is scheduled for commercialization next year.