오성준 연구실

AI is no longer a research curiosity. It is reshaping how we live and work. To fully exploit its benefits, we must address critical gaps in trustworthiness. Current foundational models like LLMs have critical trustworthiness problems: they hallucinate false information, fail at continual learning, resist knowledge editing (making GDPR compliance impractical), leak private information embedded in parameters, and require prohibitive compute for training and personalisation. These issues are blocking the widespread adoption of AI and the productivity revolution it promises. Our approach: Knowledge-Intelligence Separation. Just as the code-data separation in the 1960s enabled the modern software industry, we believe this separation is the key to unlocking AI’s full potential. When knowledge is stored in interpretable, editable external modules while intelligence (reasoning, generalisation) remains in the model, we enable faster customisation, training data attribution by design, and knowledge editing and unlearning . Compositional generalization, the ability to recognize familiar parts in novel contexts, is a defining property of intelligent systems. Although modern models are trained on massive datasets, they still cover only a tiny fraction of the combinatorial space of possible inputs, raising the question of what structure representations must have to support generalization to unseen combinations. We formalize three desiderata for compositional generalization under standard training (divisibility, transferability, stability) and show they impose necessary geometric constraints: representations must decompose linearly into per-concept components, and these components must be orthogonal across concepts. This provides theoretical grounding for the Linear Representation Hyp